SSRN Id3905180

Noname manuscript No.
(will be inserted by the editor)
Literature Review on Multi-spectral imaging for fruits and

vegetable.
Shilpa Gaikwad
Received: date / Accepted: date
Abstract In the field of agriculture fruit grading and picture and they became very popular [5]. The infor-
vegetable classification is an important and challeng- mation that we cannot visualize with the human eye,
ing task. The current fruit and vegetable classifications we can visualize with multi-spectral imaging. Different
are done manually by humans, which results in incon- spectral bands can be used in various applications such
sistent performance. There is an influence of external as sugar content detection. In multispectral imaging,
surroundings on this manual classification. Sometimes there is at least two to five spectral band involved.
getting an expert fruit or vegetable grader, is challeng-
A multi-spectral system consists of the following
ing and the performance of that person may become
spectral bands: visible spectra are in the range of 0.4
stagnant over some time. With the recent development
to 0.7 um. The range of near-infrared spectra (NIR) is
in computer technology and multi-spectral camera sys-
0.7 to 1 um, the short wave infrared (SWIR) is vis-
tem, it is possible to achieve an efficient fruit grading or
ible in the range of 1 to 1.7 um, the mid-wave in-
vegetable classification system. In this paper, we sum-
frared (MWIR) is visible in the range of 3.5 to 5 um
marize different automated fruit grading as well as veg-
and long-wave infrared (LWIR) is visible in the range
etable classification system, which is based on multi-
of 8 to 12 um. These are combined into a single sys-
spectral imaging.
tem [6]. Multi-spectral imaging can be done in a mini-
Keywords Multi-spectral · Fruit grading · vegetable mum of 3 and a maximum of 15 spectral bands. Spec-
classification tral imaging which covers a wide spectral range around
hundreds of contiguous spectral bands is called Hyper-
spectral imaging[7].The application of Multi-spectral
imaging includes: detecting or tracking military targets,
1 Introduction
detecting landmines i.e. presence of underground mis-
A multi-spectral image is the set of images for different siles, detection of a ballistic missile, weather forecast-
wavelengths, which is captured throughout the electro- ing, space-based imaging, and investigation of paintings
magnetic spectrum [2]. Wavelength separation for im- and documents[8–10].
age data is typically achieved with the help of colour Fig.1 shows the Schematic of Multispectral imag-
filters [3]. Multispectral imaging not only includes vis- ing. The light source consists of a large range of wave-
ible wavelength but sometimes the wavelength outside lengths(including visible and non-visible) is passed to a
the visible spectrum (i.e. infrared and ultra-violet) [4]. multispectral filtered disk to generate a specific wave-
With the recent advancement in manufacturing tech- length of light. A fruit sample is kept into a specific
niques, Light Emitting Diodes (LED’s) with multiple wavelength and a camera captures the image of the
colours, such as red, green, blue (RGB) came into the sample. Now the multispectral filter disk is rotated to
Shilpa Gaikwad
obtain other slices of the multispectral image. Finally,
Research Scholar the computer is used to process the multispectral im-
Symbiosis International University, Pune, India age with the help of algorithms to obtain the grading of
E-mail: shilpa.gaikwad.phd2020@sitpune.edu.in the fruit/vegetable sample. The camera used is without
Electronic copy available at: https://ssrn.com/abstract=3905180

2 Shilpa Gaikwad
Fig. 1 Schematic diagram of Multi-spectral imaging. The light source wavelengths (including visible and non-visible) is passed
to a multispectral filtered disk to generate a specific wavelength of light. A fruit sample is kept at a specific wavelength and a
camera captures the image of the sample. Now the multispectral filter disk is rotated to obtain other slices of the multispectral
image. Finally, the computer is used to process the multispectral image with the help of algorithms to obtain the grading of
the fruit/vegetable sample. The camera used is without any IR filter so that all the wavelengths of light can be observed.
any IR filter so that all the wavelengths of light can be HLB disease in citrus, frass on tomatoes, etc. 5.Flower
observed. or Fruit detection (Long-Wavelength sensitive receptors
The agriculture sector of India contributes an im- for bees, foreign materials in samples of vegetables.)
portant part to economic growth. The fruit grading ac- 6.Internal fruit parameters like moisture content (MC),
tivity is crucial in the agricultural sector due to the sugar content, vitamin C, pH, total soluble solids (TSS)
higher demand to get a good fruit quality in the mar- and firmness (TSS and firmness in strawberry) 7.Fruit
ket. Usually, the quality examination of fruits is car- Ripeness/ Maturity (banana) 8.Fruit Yield (growth in-
ried out by skilled labourers. The visual examination dicators of sugar beet, rice plants, tomato)
to manually sort the fruits is work-intensive, slow, in-
consistent and inaccurate in assessment or evaluation
by a different human. This review focuses on the re- 2 Literature Review
cent development in Multi-spectral imaging in the grad-
ing of fruits and vegetables. It is non-destructive, sim- In our daily nutrition fruits and vegetables are very im-
ple, rapid, which makes this technology perfectly suit- portant sources. From a consumers point of view, the
able for real-time monitoring and quality control.Fig.2 quality of fruit and vegetable is a vital feature. Fruit
shows the Overview of Multispectral imaging for vari- and vegetable quality include both internal and exter-
ous applications in fruits and vegetables. nal properties. The internal quality mainly is deter-
mined by smell, taste, texture, firmness of flesh, dis-
The main objectives of this paper are: eases, organic residues and nutritional quality which
1.Varietal identification of fruits (rice and tomato includes sugar content, starch, organic acids, total solu-
seeds) 2.Fruit Quality includes quality of pomegranate ble solids content, etc.[11]. The external quality mainly
fruit, Jatropha curcas seed quality, optical reflection is determined by shape, size, colour, appearance and
changes in wok-fried vegetables(carrots and celeriac), bruises[12].
hail damage assessment in potato, differentiation of vanilla
cream samples in two quality classes, etc. 3.Classifica-
tion of fruits (medicinal plant leaves, tree species and 2.1 Varietal identification
coral reefs) 4.Disease detection in fruits which includes
disease detection in fruit plants (citrus and peach or- Quality and quantity of yield production are found out
chard), rice sheath blight disease, D. Avenae (pathogen) using the variety identification technique. A combina-
in black oat seeds, Fire Blight (FB) infected pear trees, tion of multi-spectral imaging and chemometric data

Literature Review on Multi-spectral imaging for fruits and vegetable. 3
Fig. 2 Overview of Multispectral Imaging for applications in fruits and vegetables. There are eight applications for which
multispectral imaging is used. Varietal identification is generally used with tomato and rice seeds. The fruit quality deter-
mination application involves pomegranate fruit quality, Jatropha curcas seed quality, optical reflection changes in wok-fried
vegetables (carrots and celeriac), hail damage assessment in potato and differentiation of vanilla cream samples into fresh and
spoiled classes. The classification application involves classifying medicinal plant leaves, tree species and coral reefs. There are
many disease detection application for plants, seeds and tree species. Some common disease detection applications are disease
detection in citrus and peach orchard, frass on tomatoes, fire blight infected pear tree, D.Avenea fungi in black oats, and rice
sheath blight disease. Some application focuses on flower and fruit detection. These applications involve long-wavelength sen-
sitive receptors response in honey bees to multispectral images of flower and foreign material detection in fresh-cut vegetables
using images. Internal fruit parameters are one of the important applications where total soluble solids and firmness in fruits
are measured. The fruit ripeness or maturity detection can be done for bananas. Another application of multispectral imaging
is fruit and vegetable yield prediction. Growth indicators of sugar beet, the yield of tomato and rice plants can be performed.
Fig. 3 (a-e) Rice seeds classification. Multi-spectral images allow the rice to be classified as FD2 (a), QXY512 (b), HXD3 (c),
QXY822 (d), and WKJ11 (e). These five types are varietal identification numbers of rice[1].
helped to find the varieties of rice seeds in a fast way support vector machines (LS-SVM) models and prin-
without damaging the rice seeds. In this study, multi- cipal component analysis-back propagation neural net-
spectral imaging methods were applied for the separa- work (PCA-BPNN) were used. In the next step, their
tion of rice seeds based on different varieties. In this discrimination performance was compared to classify
method, the morphological and spectral features were the rice seeds into 5 different varieties. The spectral
extracted from multispectral imaging. Then the differ- data consisted of various rice seed features. Lastly, the
ent chemometric methods along with the partial least spectral and morphological information were combined
squares discriminant analysis (PLS-DA), least squares- respectively and the discrimination performance was

4 Shilpa Gaikwad
compared. After comparing, there exist differences in 2.1.1 Varietal identification setup
the varieties of rice seeds and they could be classified
according to variety as shown in (Fig.3(a-e)). Liu et al. Images from each seed sample (tomato or rice) are cap-
[1] used 19 spectral bands for online variety identifi- tured using a VideometerLab instrument as shown in
cation of rice seeds with the help of chemometric and Fig.4. This setup captured multispectral images in 19
Multispectral imaging methods. The classification ac- different wavelengths such as 375,405, 435, 450, 470,
curacy was up to 94 % for the LS-SVM model, 62 % 505, 525, 570, 590, 630, 645, 660, 700, 780, 850, 870,
for the PLS-DA and 84 % for the PCA-BPNN models. 890, 940 and 970 nm [13]. It consists of a sphere that
has a white coating to take care of the light getting uni-
formly scattered around the object. The sphere consists
of nineteen LED along its rim along with a camera that
Recently, Shrestha et al. [13] has proposed a method
is mounted at the top. Initially, a standard target was
of varietal separation and recognition of tomato culti-
used to calibrate the system radiometrically and geo-
vars using principal component analysis (PCA) along
metrically. A light setup was employed depending on
with Normalized canonical discriminant analysis (nCDA).
the type of the sample to make sure that direct compa-
The study is to calculate and compare the parents and
rable images were captured [1]. The background infor-
next-generation results of the method. For this study
mation in the captured image is not relevant. Thus, a
2 sample sets were considered. The first set comprised
CDA method was used to remove the background and
2 cultivars and their 2 crosses for studying parent and
only retain the region of interest (ROI). The Videome-
offspring affinity. 11 cultivars were used in the second
terLab software was used to extract data and transform
sample set for varietal identification. With the help of a
the pixel data once the images were captured. Various
VideometerLab instrument, multispectral images were
algorithms such as n-CDA, PCA, PLS-DA, BPCNN,
captured. This instrument can capture multispectral
and LS-SVM were used for this study. The seed images
images in 19 wavelengths. From these images, data was
were segmented based on a certain threshold. Morpho-
extracted and conversion of pixel data was performed
logical features such as the area (mm2 ), width/length,
using the VideometerLab software. For the discrimi-
and roundness values of seeds were extracted for anal-
nation between the cultivars nCDA was used. Various
ysis of the image [1].
nCDA conversions were used for the study. In one trans-
formation eleven cultivars were used, while in others
various matched combination nCDA conversions were
used. A blob database was constructed with all seed 2.2 Fruit quality
images. One blob represented one seed. Many features
were extracted and computed from the blob toolbox. The quality of fruit is a vital factor when considered
And from these features, tomato cultivars were dis- from point of view of the consumers. Human experts
criminated against. The shape features considered were examine the quality of the fruit [12]. But manual sort-
area, length, width, roundness. Similarly, the intensity, ing by visual inspection is work-intensive, slow, incon-
hue, saturation, CIE sistent and incorrect as it is the decision by another
Lab L *, CIELab a *, CIELab b * and RegionMSI mean person. With the invention of multispectral imaging
related to a blob feature were taken as colour features. techniques, automation of the grading process can be
All these features which are taken from the blob tool- done which will reduce labour cost, and increase the ef-
box are variables for varietal discrimination. The shape ficiency and accuracy of the sorting process. Fruit qual-
features constitute the area of the pixel per seed size ity includes both internal and external properties. The
images. The cultivars do not get a good separation due internal quality includes taste, texture, smell, flavour,
to this. This shows that the shape of the tomato seeds flesh firmness, diseases, and chemical/organic residues
are mainly comparable with each other so they cannot and nutritional quality which includes sugar content,
be used for varietal recognition. Blob feature Region- starch, organic acids, total soluble solids content, etc.
MSI mean could achieve the best separation when it The quality of fruit depends on the shape, size, colour,
was compared with the intensity of the pixel (mean). and skin defects (bruises). The maturity and defect of
The calibration set values are compared with the re- skin are the most crucial features that are used to find
sults obtained from the unknown set. And the model the fruit quality [17]. An automatic fruit and vegetable
can accurately identify between the pairs. The accu- grading system will help both farmers and consumers
racy achieved was 82 %. For step-wise PLS-DA models, by providing high-quality fruits in the market [18].
the classification error was 7% for the cross-validation Khodabakhshian et al. [19] has developed a mul-
as well as the prediction. tispectral imaging system to assess the pomegranate

Fig. 4 Structural setup for capturing multispectral images using the video meter lab instrument. Images from each seed
sample (tomato or rice) are captured using a VideometerLab instrument. This setup captured multispectral images in 19
different wavelengths. It consists of a sphere that has a white coating to take care of the light getting uniformly scattered
around the object. The sphere consists of 19 light-emitting diodes (LED) along its rim along with a camera that is mounted
at the top. Initially, a standard target was used to calibrate the system radiometrically and geometrically. A light setup was
employed depending on the type of the sample to make sure that direct comparable images were captured [13].
fruit quality. This system was dependent on the op- ital pH meter was used to measure the pH. The aver-
timal wavelengths. A visible/NIR spectroscopy in the age of these measured values for both parameters was
400– recorded. An average percentage of the citric acid was
1100 nm range was tested for finding the TSS, pH and measured after the value of TA was acquired with the
titratable acidity (TA). The performance of a multi- help of a Metrohm 862 compact titro sampler.
spectral imaging system was estimated by multiple lin- Preprocessing: A huge amount of spectral data is
ear regression technique. The resultant soluble solids produced with the help of Vis/NIR instruments. How-
have the value of r as 0.97, the Root Mean Square Er- ever, there is excessive data such as the background and
ror of calibration (RMSEC) of 0.21°Brix and the ratio noise in the data acquired from the spectrometer apart
performance deviation (RPD) is 6.7 °Brix. Hence, it is from the required information on samples. The effect of
shown that the models have a great ability to predict this irrelevant information is reduced by pre-processing
pH and TA. this data. This is done to achieve accurate, stable and
reliable calibration models. The pre-processing tech-
2.2.1 Block diagram of pomegranate fruit quality niques utilized for this purpose can be categorized as
columns pretreatments and rows pretreatments.
Spectrometer: ParLeS software was employed to implement the
As shown in Fig.6(a)The spectrometer is provided pretreatments for this study. Initially, the average of
with an optical fibre cable externally. To capture the four spectra was taken to get a single spectrum. An
Vis/NIR diffused reflectance spectra from the pomegranate absorbance value was achieved by converting this spec-
fruit the measurement system is operated in reflectance trum using the equation Abs = log (1 /R) where R
mode. For each fruit, an average value of a total of 30 is the quantity of reflectance. A linear correlation was
scans was taken. Later the entire data is saved in a achieved between the molecular concentration of the
computer for future operation. sample and the spectra. Lastly, various methods for
Evaluation of TSS, TA and pH: preprocessing such as centering, normalization, smooth-
Once the spectra were acquired, the TSS, TA, and of ing and differentiation were performed. Centering was
each fruit sample was tested. A manual fruit squeezer done to ensure the best results in terms of change across
was used to find the above-mentioned features. This the mean. Smoothing was used to find the best Signal
was followed by the filtration and centrifugation of the to noise (SNR) ratio. Multiplicative scatter correction
fruit. The TSS was measured in °Brix three times with (MSC) was used to eliminate the results of scattering
the help of a hand-held refractometer. Similarly, a dig- related to the average spectrum. Similarly, the back-

6 Shilpa Gaikwad
Fig. 5 (a) RGB images of the front and back surfaces of Jatropha curcas seeds and their respective reflectance images acquired.
These reflectance images were taken at a wavelength of 940 nm for the gray scale. The images are transformed with the help of
the nCDA algorithm and X-ray images. 3 colour channels namely, red, green and blue are used to represent the RGB images.
A single-colour value for each pixel is generated in the image [14]. (b) Plot showing all spectral channels of a celeriac sample
[15]. (c) The experimental plots show the 3 growth stages of the potato crops: Tuber Initiation (TI), Early Bulk (EB) and
Late Bulk (LB) phases. The false colour images were captured 95 days after planting the crops. Two varieties were used NH
(Norkotah)and RA (Ranger). Skip represents an unplanted area in the field[16].
ground spectra were eliminated and the spectral reso- to evaluate the results of the validation and calibration
lution was increased by using the 1st and 2nd derivative model.
preprocessing methods. Bianchini et al. [14] investigated the Multi-spectral
with X-ray image technique for Jatropha curcas seed
Calibration and Validation: quality for 100 seeds per lot. For the study, the seed
surface of 3 lots of Jatropha curcas was considered each
A model was developed using the PLS regression
having 3 vigor levels: low, high and medium vigor as
method between quality parameters (TA, TSS and pH)
shown in Fig. 5(a). High-resolution multi-spectral im-
from samples and spectral responses of the samples.
ages with a size of 40 micrometre/pixel were formed
ParLeS software was used for comparing one quality pa-
after capturing reflectance images at 19 different wave-
rameter and the spectral data. The calibration set con-
lengths. Hence, by combining multi-spectral with X-ray
sisted of 70 samples. The purpose of the PLS method
imaging methods, high accuracy of 0.96–0.98 % was
is to find a mathematical relation between a set of in-
achieved to find seed quality.
dependent variables/predictors which consists of the X
matrix (70 fruits × 700 wavelengths) and the dependent
variable which consists of the Y matrix (70 fruits × 1). 2.2.2 Block diagram of Jatropha curcas seed quality:
The dependent variables (Y) are defined by the quality
parameter (TSS, TA, and pH) value from the calibra- Multispectral imaging:
tion set. For validation 30 fruits were selected. When With the help of VideometerLab4 setup, multispec-
the predicted residual error sum of squares was mini- tral images were captured at 19 different wavelengths
mum, the calibration model was built using the maxi- refer Fig. 6(b). For this work, 3 seed lots were con-
mum number of latent elements. The correlation coef- sidered with 10 seeds per lot. The whole process was
ficient (r), Root mean square error of calibration (RM- repeated ten times. Multispectral images were acquired
SEC), Root mean square error of prediction (RMSEP) for both the front and back seed surfaces of ten samples.
and Ratio performance deviation (RPD) were measured The nCDA algorithm was implemented to capture only

Fig. 6 Block diagram for Fruit Quality (a) Chemometric block diagram to extract the information of pomegranate fruit.
The spectrometer is provided with an optical fibre cable externally. To capture the Vis/NIR diffused reflectance spectra
from the fruit sample the measurement system is operated in reflectance mode. The entire data is stored in a computer
for further operation. The data(raw) from the spectrometer is given to the preprocessor to remove irrelevant data. Finally
various methods of preprocessing such as mean, centering, smoothing, normalization is used. Further standard normal variant
and differentiation operation is carried out. In calibration and validation, a model was developed using the PLS regression
method between quality parameters (TA, TSS and pH) from samples and spectral responses of the samples. (b) Jatropha
curcas seeds evaluation using X-ray and multispectral imaging. During multispectral imaging, the nCDA algorithm is used to
capture only important wavelengths and further, the PCA technique is used to process the multispectral data to achieve the
proper classification of the seed. For the X-ray method, 100 seeds were radiographed three times using a multi-focal digital
radiography system. The CDA method is used to get the best differentiation among seed lot groups and to classify them as
normal, abnormal, and dead seeds. LDA algorithm was used to develop three models. The first model was generated with
the help of multispectral data at a wavelength of 940 nm. The second model was generated with the help of X-ray classes
data. Whereas the third classification model was generated using multispectral and X-ray imaging. (c) The VideometerLab
setup was used to capture the multispectral images of carrots and celeriac. A t-test was conducted to check for important
differences on a 5 % level in many percentage points of the light reflectance. This was done to examine the reflection changes
as a function of days. (d)The vanilla cream sample of 25gm weight was kept in a sterile stomacher bag having 225 mL of a
sterilized quarter-strength solution of Ringer to calculate the total viable counts (TVC). The VideometerLab setup was used
to capture the multispectral images for each cream sample. The next step is Data labelling, which was used to calculate the
mean reflectance spectrum and the respective standard deviation values. The model was constructed, to correctly find fresh
or spoiled samples. A UOS algorithm based on the DFS method was used to choose the best model features corresponding to
a certain test data used. Finally, an SVM classifier was used to classify the vanilla cream samples.
meaningful wavelengths using the candisc package. A X-ray imaging was performed on 100 seeds per lot.
facto miner package was used in the PCA technique to After X-ray imaging 25 seeds were replicated four times
operate on the multispectral data. PC1 and PC2 which and sown in sand placed in plastic boxes at a temper-
are the two principal components were plotted to choose ature of 30 ºC for 20 hours. After sowing for ten days
important wavelengths to differentiate the seed lot us- each seed was examined for three quality features: nor-
ing Pearson’s correlation test. A multivariate analysis mal seeds, abnormal seeds, and dead seeds. Later they
of variance (MANOVA) was performed to test the ef- were differentiated based on three categories. Based on
fect of low, high and medium vigor. With the help of the germination tests of the seeds and the integrity
VideometerLab software and the R package, statistical of the tissue in the radiographic images. CDA method
analysis were performed. used the candisc R package to get the best differentia-
X-ray imaging: tion among seed lot groups using a database extracted

8 Shilpa Gaikwad
from X-ray classes, quality features, and reflectance in- Lianou et al. [20] proposed an online feature se-
formation at the wavelength 940 nm. lection grading of Vanilla Cream quality using Multi-
Validation: spectral Imaging. The study was implemented for the
Linear Discriminant Analysis (LDA) algorithm was inspection of 2 microbiological quality classes of cream
used to develop three models. The first model was gen- samples, with the value of total viable counts (TVC)
erated with the help of multispectral data at a wave- ≤ 2.0 log CFU/g for fresh samples and TVC ≥ 6.0
length of 940 nm.. The second model was generated log CFU/g for spoiled samples. Both the classes show
with the help of X-ray classes data. Whereas the third great performance with their results. The overall clas-
classification model was generated with the combina- sification accuracy is 91.7 % for validating the model.
tion of X-ray and multispectral data. To develop the This model can be further extended to find the micro-
model a total of 300 seeds were used. 210 seeds were biological quality of classes by using 1 log step. This
used for training the model whereas, for validation pur- step was taken to validate the capability of the model
poses, 90 seeds were used. For training the data, the k- to evaluate increasing microbial populations.
fold cross-validation was employed where k=5. With
the help of the confusion matrix, accuracy, sensitiv- 2.2.4 Block diagram to find vanilla cream quality
ity, specificity, and KC were calculated to estimate the
model. Samples of vanilla cream subjected to microbi-
Bjørn S. Dissing et al. [15] proposed an imaging ological study
technique for capturing high-quality multi-spectral im- Figure 6 (d) shows block diagram of the vanilla
ages in carrots and celeriac in 14 days to find out opti- cream sample. The vanilla cream sample of 25gm weight
cal reflection changes. For this purpose, the vegetables was kept in a sterile stomacher bag having 225 mL of
were fried before and frozen at 30 0 C for 4 months be- the sterilized 1/4th concentration solution of Ringer. It
fore recording a multi-spectral image. Fig. 5(b) shows was homogenized in a Stomacher device at room tem-
the plot of all spectral channels of a celeriac sample. perature for 1 min. To calculate the TVC precise serial
The vegetables were kept at +5 0 C throughout the im- decimal dilutions in Ringer’s solution were considered.
age recording of 14 days. During this period, surface Finally, colonies were measured after an incubation pe-
changes, as well as the reflectance properties, were very riod of 48 hours with the temperature maintained at 30
subtle. But, numerically important differences for some °C. The microbiological quality of data was indicated
wavelengths and combinations of wavelengths were ob- as log (colony forming units) per gram of cream (log
served. A t-test was conducted to check for important CFU/g).
differences on a 5 % level in many percentage points of Image capturing and its study:
the light reflectance. Major variations were observed in The VideometerLab setup was used to capture the
the reflectance spectrum of the carrots and celeriac from multispectral images for each cream sample. Capturing
days 2 to 4. While for the celeriac, significant changes the spectral data, a data cube of m x n size in pixels
in the reflectance spectra till day 14 were noted. is captured for each cream sample. This resulted in the
production of a huge amount of data describing samples
in time-based experiments considering different storage
2.2.3 Block diagram of wok-fried vegetables to find conditions. Before the image was captured a light setup
optical reflection changes: called auto light is used and the calibration is done ra-
diometrically and geometrically. The samples of vanilla
The VideometerLab setup was used to capture the mul- cream were placed in an Ulbricht sphere which has a
tispectral images refer Fig. 6(c). T-tests were done on camera installed on the top and the subsequent multi-
the multispectral images to find the importance of re- spectral image of the cream surface was captured. To
flection changes between days. Pixel intensity values of eliminate the redundant data, segmentation was done
the features were measured at the 5th, 10th, 25th, 50th, using the in-house method. After this normalization
75th, 90th, and 95th percentiles for each vegetable piece was done for pixel values of each wavelength in the
kept in the refrigerator. This was done to examine the range [0,1]. The next step was to calculate the mean
reflection changes as a function of days. To categorize reflectance spectrum and the respective standard devi-
the differences 358 carrot samples and 389 celeriac sam- ation values were calculated.
ples were tested in the interval of six different sample Data labelling :
days. The p-values for the T-tests were plotted with Out of the 245 spectra of the cream samples, 65
the null hypothesis and it shows no variation in mean spectra were used in model training, 48 spectra were
values for the six-day interval. used for validating the model and the remaining 132

spectra were used in model testing. The input matrix 2.3 Classification
had 18 average and 18 SD values of the reflective spec-
tra extracted from 245 samples of vanilla cream. The Naeem et al. [21] applied Multi-spectral and Texture
output matrix had the results of microbiological data Feature Using Machine to categorize the Medicinal Plant
for finding TVCs in the corresponding samples. At first, Leaves using five spectral bands with a range of 460 nm
we applied binary classification that will automatically to 1560 nm. In this study, six medicinal plant leaves
find samples having TVC ≥ 6 log CFU/g and TVC ≤ were considered.Fig.7 shows the gray scale image of
2 log CFU/g. The model was constructed to correctly the 6 medicinal plant leaves. The classification accu-
find fresh or spoiled samples. racy achieved is Tulsi 99.10 % , Peppermint 99.80 %,
Dynamic Feature Selection (DFS) method: Bael 98.40 %, Lemon balm 99.90 % , Catnip 98.40 %,
An Unsupervised Online Feature Selection (UOS) and Stevia 99.20 % using the multi-layer perceptron
algorithm based on the DFS method was used to choose classifier.The multi-layer perceptron method achieved
the best model features corresponding to a certain test a high accuracy of 99.01 % as compared to the other
data used in this work. A three-step DFS approach methods.
was designed. The first step was training-dependent fea-
ture elimination. The second step was the online test- 2.3.1 Block diagram of Medicinal plant leaves
dependent feature elimination step. The third step was classification
the decision taken by considering the training set and
test samples collected. This will remove the test sample A computer vision-based experimental setup was used
features that are enclosed by samples of various classes for collecting multi-spectral and digital image data-base
or it is away from the other classes. The classification as shown in Fig. 8(a). The images were transferred into
model is built when the training set features fit all three the grey level format with a resolution of 800 × 800
conditions for each test sample. pixels by cropping the leaf region exactly. A Sobel fil-
Support Vector Machine (SVM) classifier: ter was used for edge/line detection of seed intensity.
SVM with linear kernel was used to classify the If the seed intensity threshold value is greater than 6,
vanilla cream samples. Every new test sample selects then mark the region as a region of observation. In this
a new set of features. These features are reinserted in work, five regions of observation were drawn on each
the data set for training and from this features were image and fused features were extracted from the data
selected to create new test data in the online feature set. The chi-square feature selection technique reduces
selection method. the feature vector space and selects the fourteen best
Zhou et al. [16] introduced crop hail damage as- features for medicinal plant leaves classification. Five
sessment in potato using aerial multispectral imaging. machine learning classifiers named multi-layer percep-
In this study, the four damage level in potato is 0%, tron (MLP), LogitBoost (LB), Bagging (B), Random
33%, 66% and 99%. Also, three growth stages of potato Forest (RF), and Simple Logistics (SL) were employed
were considered which includes: Tuber Initiation (TI), on the medicinal plant leaves database.
Early Bulk (EB), and Late Bulk (LB), in 2 varieties Abdollahnejad et al. [22] proposed tree species clas-
of potatoes as shown in Fig. 5(c). For the Early Bulk sification and health condition assessment with the help
stage, a severe damage effect of 99% was observed. of Unmanned Aircraft system (UAS) Multi-spectral Imag-
Due to severe damage, the Green Normalized Differ- ing for Mixed Broad leaf-Conifer Forest. This study de-
ence range vegetation index (GNDVI) value achieved veloped a new method to classify tree species and to
was lower, which resulted in a reduction in yield as com- recognize healthy, unhealthy, and dead trees affected by
pared to other treatments. The resultant Vegetation in- bark beetle infection. The results show an overall accu-
dex (VI) of GNDVI, normalized difference vegetation racy (OA) = 81.18 % and values of Kappa= 0.70 which
index (NDVI), soil-adjusted vegetation index (SAVI) prove the capability of the current method to classify
showed a satisfactory correlation of (r>0.77) with crop tree species. The health status condition of tree species
yield at the Early Bulk (EB) stages. The hail damage shows the value of an OA = 84.71 % and Kappa= 0.66.
treatment is done in intervals for 60 days. The effects The Support Vector Machine method proved as a good
of hail damage should be checked within 10 days after classifier of tree species. The fusion of vegetation in-
the damage. After an observation interval of 10 days, dices (VI) and texture analysis (TA) layers also results
plant regrowth will not reveal the hail damage effects. in increased OA by 4.24 %. This means that the system
Thus, the imaging method based on Unmanned Aerial provides the latest and new method of data extracted
vehicle (UAV) is a strong technique, for the assessment from UAS platforms. This method can be used to check
of hail damage in agriculture. huge areas affected by biological intrusion factors for

10 Shilpa Gaikwad
mapping and detection very fast. Other applications total six images considered in the study, only the first
are to estimate habitat conditions and tree inventory three bands were selected to be more specific in wave-
at low costs. length interval. A thematic map is a digital map that
consists of the bottom types. This was useful for captur-
2.3.2 Block diagram of tree species classification and ing training and test data and helped in the visual as-
health status pect of spatial distribution. The fourth kind of data was
a bathymetric map taken from multispectral images of
In this study, UAS bi-temporal aerial imagery was used the study area. In the preprocessing stage, three images
which incorporates a high-resolution 5-band spectrum were taken to classify the coral environment. The noise
as shown in Fig. 8(b). In the preprocessing step, the was removed with the help of filters and Lyzenga’s algo-
MicaSense RedEdge-M multispectral calibrate the re- rithm was applied to extract features from the bathy-
flectance panel converts raw pixel values into reflectance. metric map. The database was generated for training
This was done to correctly describe the target consist- to find the accuracy. The thematic map had five dif-
ing of the object of interest from the image pixel values. ferent classes. This was obtained by visual study and
In the absence of calibration, the data collected will fieldwork.
not correctly compare for change identification. Using Classifier Ensembles:
Structure from motion (SfM) imaging photogrammet- A classifier ensemble consists of two sets of SVM
ric database was generated. The Agisoft Metashape is and a minimum distance classifier to increase the result
used so that the values of image pixels correctly identify of pattern recognition. The hybrid classifier was used
the region of interest by adjusting the atmospheric and either in series or parallel. In this work, 50 samples were
light conditions For the canopy spectral study and tex- built to train and test the machine in the pre-processing
ture study, DTMs, DSMs, and orthophotos were given stage. The area covering each pixel was reviewed to get
as input with a resolution of 0.05m. For the spectral an accuracy of bottom type related to testing pixels.
study, 9 VI and 13 TA variables were used. After this bi- Refinement Phase 01 (SVM):
temporal data was combined with texture data. Then In this stage, various binary classifiers were built
supervised machine learning technique was employed to and validated. For the training set, 35 samples were con-
catch the reflectance values within a buffer area of 2m sidered, and a polynomial kernel was employed. For the
radius from each treetop. Using the same buffer area classification purpose, three spectral bands and bathy-
effect of crown shape and vague pixels from the over- metric were applied as input to all the machines. The
lapped crown areas were removed. The SVM algorithm output was given to the first decision unit which re-
delivered good results and could achieve the correct and sulted in the formation of two data subsets. One of the
reliable classification of tree species. subsets consisted of positive results whereas the other
Henriques et al. [23] evaluated a methodology to one consisted of ambiguous results. In the second data
classify coral reefs using a set of hybrid support vector subset list related to the classes was defined. The non-
machine (SVM) classifiers with the help of Landsat- ambiguously classified subset was given as an input to
7/ETM orbital images acquired from the satellite plus the combination unit. The next stage consists of the
the multi-spectral images. The performance results of classification process, ambiguous classification was per-
the Hybrid Support vector machine (Hybrid SVM) were formed. The use of the binary classifier produced sep-
compared with the support vector machine classifier arate classification for the classes which were on the
voting scheme methods. In all the tests performed, the list. The second decision unit was given the output of
Hybrid SVM achieved a greater classification accuracy 10 binary classifiers. The same process was repeated as
of 89.33% than that achieved by the other 2 algorithms. earlier. A new list was created from the second data
The 2 algorithms: One-Against-the-Rest showed an over- subset. The next processing unit received the ambigu-
all accuracy of 78.66% while the One-Against-One clas- ously classified pixels as well the newly created list of
sifier showed an overall accuracy of 70.67%. classes whereas the classified pixels were directly sent
to the combination unit.
2.3.3 Block diagram of coral reef classification Refinement phase 02 (MDC):
The minimum distance classifier (MDC) algorithm
Generation of data: which is based on the class centroids is considered in
As shown in Fig.8(c) the study area is an underwa- the second stage. A prototype vector that signifies the
ter area that had a depth of 20m. The ETM/Landsat- pattern class is produced by the average values of the
7 orbital images had ETM+1, ETM+2, and ETM+3 descriptor vector related to the validation data. The
spectral bands captured from the satellite. Out of the first step in this process was mapping the pixel val-

Fig. 7 A total of 6 medicinal plant leaves. The original images are transformed into gray scale with the region of interest that
allows classification of these leaves into Tulsi, Peppermint, Bael, Lemon Balm, Catnip and Stevia. It also shows 5 colourful
regions of observation.[21]
ues to the input space. After this, the algorithm was Figure 11(b) shows the uncalibrated and calibrated or-
implemented by considering the list of positive classes thomosaics(peach orchard) at the Neka study site.
received by the earlier phase. The output of this phase
was directly sent to the combination unit. Lastly, the 2.4.1 Block diagram for plant disease detection in
partial results of the outputs of different classification citrus and peach orchard
phases were added to produce the final classified image.
The classified image was geo-referenced which resulted In this work, two test areas were considered the Fasa
in the formation of a thematic map of the coral forma- test area consisting of 557 images and the Neka test
tion study site in the mapping stage. area having 683 images (refer Fig. 9 (a)). The images
were acquired with the help of a UAV.
Band-to-band registration:
2.4 Disease Detection In this step, the registration of the images was done
related to spectral bands. There was a geometric dis-
Pourazar et al. [24] applied aerial multi-spectral im- placement in the images due to the multi-lens feature.
agery for plant disease detection. Diseased trees in the This was removed with the help of the R band. The R
orchard were identified with the help of radiometric band was the master and the remaining bands acted as
calibration. Two ways were used to perform a radio- slaves and were registered to the master. In this way
metric correction on disease detection in plants. One registration of the bands with the images was done.
of the ways included the use of t-test and entropy dis- The results show RMSE value as 0.4 pixels which were
tances comparison to separate the healthy plants from achieved related to the measurements of the object.
the diseased ones. The second way, the accuracy of the Alignment of image and generation of ortho-
classification was calculated with the help of the ra- mosaic:
diometric calibration effect. The OA and kappa values In this step, capturing of the geometric tie points
were 96 %, 0.941,96.57 %, and 0.942 respectively for un- and the estimation of the internal and external fea-
calibrated and calibrated orthomosaic (citrus orchard) tures of the image was done. After aligning the band
classifications at the Fasa study site. Whereas the OA registered image a dense point cloud was generated for
and kappa values were 95.58 %, 0.913,95.56 %,0.913 re- orthomosaic creation.
spectively for uncalibrated and calibrated orthomosaic Radiometric normalization:
(peach orchard) classifications at the Neka study site. As various illumination states are considered dur-
Figure 11(a) shows the uncalibrated and calibrated or- ing image acquisition there is a need to generate ho-
thomosaics(citrus orchard) at the Fasa study site and mogenous orthomosaic. So radiometric normalization

12 Shilpa Gaikwad
Fig. 8 Different classification strategies using multispectral imaging on the plant data. (a)For collecting multi-spectral and
digital image databases, a computer vision-based experimental setup was used. The images were transferred into the grey level
format by cropping the leaf region exactly. A Sobel filter was used for edge/line detection of seed intensity. Five region of
observation were drawn on each image and fused features were extracted from the data set. The chi-square feature selection
technique reduces the feature vector space and selects the best features for medicinal plant leaves classification. 5 machine
learning classifiers were employed on the medicinal plant leaves database to classify them into 6 different classes. (b) Block
diagram of disease detection in rice. A UAS bi-temporal aerial imagery was used which incorporates a high-resolution 5-band
spectrum to capture an image of tree species. In the preprocessing step, the MicaSense RedEdge-M multispectral CRP converts
raw pixel values into reflectance. The Agisoft Metashape is used so that the values of image pixels correctly identify the region of
interest by adjusting the atmospheric and light conditions. Using SfM imaging photogrammetric database was generated. The
topographic information was extracted which includes DTM’s, DSM’s, and orthophotos. For the spectral study, 9 vegetation
indices and 13 TA variables were used. After this bi-temporal data was combined with texture data. Then machine learning
technique was employed to catch the reflectance values within a buffer area of 2m radius from each treetop. The SVM algorithm
was able to recognize healthy, unhealthy, and dead trees affected by bark beetle infection. (c) The study area is an underwater
area that had a depth of 20m. In the data generation, the ETM/Landsat-7 orbital images which consist of ETM+1, ETM+2,
and ETM+3 spectral bands are captured from the satellite. In the data extraction step, out of the total 6 images considered
in the study, only the first 3 bands were selected to be more specific in wavelength interval. A thematic map is a digital map
that consists of the bottom types. The fourth kind of data was a bathymetric map taken from multispectral images of the
study area. In the preprocessing stage, 3 images were taken to classify the coral environment. A classifier ensemble consists of
two sets of SVM and a minimum distance classifier to increase the result of pattern recognition. In Refinement phase 1, for the
classification purpose,3 spectral bands and bathymetric were applied as input to all the machines. The output was given to
the first decision unit which resulted in the formation of 2 data subsets. In Refinement phase 2, the first step was mapping the
pixel values to the input space. After this, the algorithm was implemented by considering the list of positive classes received by
the earlier phase. The output of this phase was directly sent to the combination unit. Lastly, the partial results of the outputs
of different classification phases were added to produce the final classified image. The classified image was geo-referenced which
resulted in the formation of a thematic map of the coral formation study site in the mapping stage.
was done on the image block for acquiring features and the reflective-based values the main DN values of both
classification purposes. BF and AF were used. From these new values, a new
Radiometric calibration: orthomosaic was generated for both the study areas.
In this step, reference images were acquired from Spectral feature study:
the white panel before and after each flight. For both In this work, 5 spectral bands and 12 VI’s were
the study databases maximum Digital number (DN) considered in feature space. For the differentiation of
were acquired from the white panel before and after healthy from the unhealthy tree classes, the t-test was
the flight (BF and AF). There is a direct relation of performed and the entropy method was employed for
sensor radians with the DN value. At the Neka site, the which information was gathered from the calibrated
rate of the reduction is much higher due to the varia- and uncalibrated data for Neka and Fasa study sites.
tion in illumination state during the flight. To calculate After this step, the acquired feature values from the

Fig. 9 A generalized approach for disease detection using multispectral imaging. (a) Plant disease detection for citrus and
peach orchard. In-band to band registration step, the registration of the images was done related to spectral bands to remove
geometric displacement in the images. The image alignment consists of capturing the geometric tie points and the estimation
of the internal and external features of the image. After aligning the band registered image, a dense point cloud was generated
for orthomosaic creation. Radiometric normalization was done on the image block for acquiring features and classification
purposes. In radiometric calibration, the reflective-based values were calculated from the main DN values of both before and
after the flight ( BF and AF). From these new values, a new orthomosaic was generated for each study area. In feature space
generation, 5 spectral bands and 12 VI’s were considered in feature space. For the differentiation of healthy from the unhealthy
tree classes, the T-test was performed and the entropy method was employed. In the classification step, the images were
classified for the selected training data which resulted in health map generation. A random forest classifier was employed in
which 100 decision trees were considered. In the disease map, the final accuracy is tested for detecting the diseased and the
healthy tree by evaluating the error matrices of the test information. (b) To capture the imagery data from the field plot,
a UAV along with a higher resolution multispectral camera was employed. In image preprocessing to acquire colour features
from the captured images and conversion of these images into various colour spaces, ENVI software was employed. 5 types of
VI’s were measured from the captured multispectral images. After this NDVIs-change maps of various severity degrees of the
disease were produced so that the image data could find the indication and the ShB disease severity on a rated scale. Gathering
of ground truths was done by the ground NDVI readings taken with the help of the GreenSeeker handheld crop sensor. 134
average NDVI readings were taken from the areas of 67 plots that were inoculated and uninoculated with pathogens. After
this ShB severity was marked on a scale of 0±9 for disease indications. In the data evaluation and drawing step for UAV image
processing Pix4D mapper was used.ArcGIS 9.1 was used for mapping and geospatial study. For statistical evaluation PASW
Statistic 18, software was used.
ground truth data were compared for calibrated and un- Zhang et al. [25] used multispectral imaging for the
calibrated orthomosaics. Both the study areas showed detection of rice sheath blight (ShB) disease with the
different values in comparison. In this way, sensor cal- help of an unmanned aerial system. The results indicate
ibration for disease detection in vegetation indices was that between ground-measured Normalized Difference
done. Vegetation Index (NDVIs) and image-extracted NDVI’s
Classification: there was a strong correlation. The determination coef-
The images were classified with respect to the se- ficients (R2 ) is 0.907 and the RMSE is 0.0854. Whereas
lected training data which resulted in health map gener- the results indicate that between image extracted ND-
ation. A random forest classifier was employed in which VIs and disease severity there is a good correlation with
100 decision trees were considered. A T-test was per- a value of (R2 ) = 0.627 and the RMSE = 0.0852.ND-
formed which helped in differentiation between healthy VIs calculated from the images achieved an accuracy
and unhealthy classes. After this, the entropy method of 63% which could measure various levels of the dis-
and low correlation were combined in the random forest ease in the plots. The study included Original RGB and
classifier. hue,lightness and saturation (HLS) transformation im-
ages of 67 field areas. Fig. 11 (c-f) shows the 67 field
Disease map:
plots of rice cultivars. Fig. 11 (c) The RGB image, Fig.
By computing, the error matrices of the test data 11 (d) The False colour image, Fig. 11 (e) The HLS
the final accuracy was calculated to detect diseased and image, and Fig. 11 (f) The NDVI image.
healthy trees. The following parameters represent the
error matrix- omission error, commission error, OA, or
kappa values.

14 Shilpa Gaikwad
Fig. 10 Block diagram of black oat seeds: With the help of the VideometerLab4 instrument, 19 multispectral images of 200
non-infested seeds as well as 200 infested seeds were acquired. These seeds were infested with D. Avenae fungi for inoculation
intervals of 24, 72 and 120 hours. The nCDA algorithm was employed to transform the images. For every seed, the region of
interest (ROI) was recognized and a mask was created to segment the seeds from the surrounding. Various colour description
models were employed to transform the multispectral data into colour data. By calculating the individual spectral bands, the
texture features were obtained. The mean reflectance of each seed lying within the range of the spectral bands was extracted
with the help of the MultiColorMean feature. A trimmed mean of the transformed pixel values was calculated using RegionMSI
Mean. Similarly, the percentage of the blob region that had a transformation value higher than the threshold was calculated
using RegionMSIthresh. LDA method was used to develop two models that could differentiate between infested and non-
infested seeds from the spectral information. The reflectance data obtained for each seed was used to create the first model.
Similarly, colour and texture features were used to create the second model. The data was split into two sets with 60% of each
sample used for training the model and the remaining 40% of each sample for testing the model. The result of the model was
evaluated using the kappa coefficient and accuracy. The VideometerLab4 and R software were used for statistical evaluation.
2.4.2 Block diagram for rice ShB disease detection ENVI software was employed. In this work, digital im-
ages of various severity levels of ShB were converted
Unmanned aerial system (UAS): to lightness, hue and saturation. After this, the mean
To capture the imagery data from the field plot, a values of HLS were taken.
UAV along with a higher resolution multispectral cam- VI calculation:
era was employed (refer Fig. 9 (b)). The Micasense 5 types of VI’s- NDVI, Ration Vegetation Index
RedEdgeTM can capture 12-bit raw images in 5 nar- (RVI), Difference Vegetation Index (DVI),Normalized
row bands. This camera was used to calculate the re- Difference Water Index (NDWI), and Red Edge (RE)
flectance of the plant which contains data about stress were measured from the captured multispectral images.
in crops more correctly than the 3 channel camera. After this NDVIs-change maps of various disease sever-
Image capture: ity stages were produced, so that the image information
To cover all 67 field plots the camera was directed at could find the indications and the ShB disease severity
the nadir at two heights, 27 meters above the ground to scale it.
and 5.5. meter to cover 4 plots for every image dur- Ground truth information gathering:
ing the flight. This data was collected in the Support The ground NDVI readings were taken with the help
Information File i.e. S1 file. of the GreenSeeker handheld crop sensor. Various mea-
Image preprocessing: surements were recorded in every field plot so that the
To acquire color traits from the captured images accuracy of NDVI values increased which signified the
and conversion of these images into various color spaces, severity of the ShB disease. In this work, 134 average

Fig. 11 (a) The uncalibrated orthomosaics in the Fasa field area are shown on the left side and the calibrated orthomosaics in
the Fasa field area are shown on the right side. These orthomosaics show the classification map for the healthy trees that are
represented with a green colour, infected trees represented with a yellow colour and non-trees represented with a red colour.
(b) The uncalibrated orthomosaics in the Neka field area are shown on the left side and the calibrated orthomosaics in the
Neka field area are shown on the right side. These orthomosaics show the classification map for the healthy trees that are
represented with a green colour, peach leaf infected tree represented with a yellow colour and non-trees represented with a red
colour [24]. (c) The RGB image of 67 field plots of rice cultivars[25]. (d) The False image of 67 field plots of rice cultivars[25].
(e) The HLS image of 67 field plots of rice cultivars.[25] and (f) The NDVI image of 67 field plots of rice cultivars[25].
Fig. 12 (a) Images of black oat seeds: Raw images and corresponding gray scale and nCDA at wavelength 365 nm for seeds free
from fungus and seeds exposed to D. Avenae fungi for 24, 72 and 120 h. The blue colour defines healthy tissues, intermediate
contamination is defined by green and yellow colours, and the red colour defines large contamination in fungi in the images
transformed by the nCDA algorithm[26]. (b) The False Color Composite image which depicts the healthy and Fire Blight
infected areas of the pear tree crown[27].

16 Shilpa Gaikwad
Fig. 13 Block diagram of disease detection. a) Fire blight infected pear tree. For this study, FB infected pear tree orchard
was used. To collect data, 75 trees were chosen. Five leaves from each tree were taken from various parts which consist of HEL,
NSL and SDL leaves. Ground multispectral imagery was employed to capture images to detect FB infection levels in leaves and
tops of the trees. The PixelWrench2 software was used in the preprocessing step to re-build the color of the raw images. Leaf
samples were put in numbered plastic bags after ground imaging. FCC of the acquired imagery with NIR-R-G band composite
was developed. The spectral data were obtained from the images. The VI which is receptive to changes in leaf health from
earlier studies were measured from each leaf sample. After VI calculation, ANOVA was implemented to find spectral indices
which are sensitive to detect FB disease and to differentiate between diseased and healthy leaves. The UAV was flown 20 m
above ground level at a certain location. When the UAV reached the center of the treetop, 3 images were acquired of the pear
tree. After acquiring images and obtaining the data from the camera the preprocessing step was performed. The ENVI software
was utilized for the supervised classification of the aerial images. The SVM with the RBF was applied as it was more accurate
than other methods of classification. b) HLB diseases in citrus. In this study, the citrus leaf image was acquired. For the color
study, histograms of H, S, I and grey image were used. Histogram LBP was used for extraction of the texture feature. LBP is
useful texture information. The LBP gives the feature vector for the window considered. Symptoms of HLB are irregular color
patterns and malformation of leaves in some cases. In the texture feature, the data is represented in grey, hue color images.
The 6 features obtained from histograms of H, S, I and grey images are mean, SD, uniformity, 3rd moment, entropy frequently
and smoothness. In the next step, 48 dimension features from each image are extracted and finally the classification is done.
c) Frass on tomato. For this study, 112 tomatoes were considered. The next step was to capture hyperspectral fluorescence
line-scan images of tomatoes. Images of 8 tomatoes, with and without stem, were chosen for obtaining sample pixels that define
7 classes. Out of the 16 images of tomatoes, 5820 sample pixels were chosen. The hyperspectral data collected was evaluated
to find important wavelengths for generating the multispectral imaging algorithm. This algorithm was implemented on the 96
tomatoes which were left. In the frass-detection algorithm of tomatoes, 5 wavelengths and 3 ratio functions were used. To each
non-background pixel of the tomato image, the algorithm was applied. The pixel was considered as frass-contaminated if the
ratio function values were less than or equal to the threshold value. This step was done with respect to all the 3 ratio functions
used to a single pixel.
NDVI readings were taken from the areas of 67 plots ture parameters from seeds incubated for 120 h.Fig. 10
that were inoculated and uninoculated with pathogens. shows the raw image and relative gray scale and nCDA
After this ShB severity was rated on a scale of 0±9 with of black oat seeds. Results show high performance of
respect to disease symptoms. the model with an accuracy of 0.86. This indicates that
Data evaluation and drawing: the Multispectral imaging method was capable of recog-
nizing D.Avenae fungi in black oat seeds. The accuracy
For UAV image processing Pix4D mapper was used. achieved was 0.86 for the colour and texture feature
ArcGIS 9.1 was used for mapping and geospatial study. which was satisfactory.The correct accuracy achieved
For statistical evaluation PASW Statistic 18, software for black oat seeds inoculated for 24, 72 and 120 hours
was used. RMSE and determination coefficient were was 0.78, 0.83 and 1.00 respectively. This shows that
used to find the accuracy of the correlation model. multispectral imaging can be an adequate method that
Françca et al. [26] developed a model for finding will assure that the black oat seeds do not contain any
D.Avenae fungi in black oat seeds using Multi-spectral diseases or fungi.
images. The study was performed using colour and tex-

2.4.3 Block diagram to detect D.Avenae fungi in black posite image for healthy and Fire Blight infected sec-
oat seeds tions of a pear tree (crown). With the help of the SVM
classifier using 5-fold cross-validation, the sample OA’s
Data Acquisition: achieved are 98.7 % for Healthy, 91.4 % for Diseased,
With the help of the VideometerLab4 instrument, 93.9 % for Nutrient Deficiency.
19 multispectral images of 200 non-infested seeds as Deng et al. [28] investigated a detection method of
well as 200 infested seeds were acquired (Fig.12 c)). huanglongbing (HLB) disease in citrus production using
These seeds were infested with D. Avenae fungi for an ViS image processing and Cost-Support Vector Classifi-
inoculation interval of 24, 72 and 120 hours. 19 high- cation (C-SVC). In the experiment (penalty parameter
resolution images were acquired at a time in an inter- C, γ coefficient as the default), the performance of the
val of 5 seconds.VideometerLab4 software version was Radial basis function was excellent with an accuracy of
used to analyze the data. The nCDA algorithm was em- 88 %. Whereas the Sigmoid function performance was
ployed to transform the images. This was basically done the worst, with an accuracy of 60 %. In this study Par-
so that there was a maximum separation between the ticle Swarm Optimization (PSO) algorithm which pro-
classes and minimum distance within a specific class. vides an excellent solution by using iteration was used
For every seed, the region of interest (ROI) was rec- [29]. The best C acquired by PSO for non-PCA data
ognized and a mask was created to segment the seeds is 1600.286. Whereas for PCA data the value of C is
from the surrounding.The blob database was used to 518.047. The best c of both PCA and non-PCA data is
collect the seeds and from each seed, 36 variables were 0.1. To find the accuracy these two parameters(C and
obtained. This also involved colour components such as c) in C-SVC are used to cross-validate (5-fold) the in-
hue and saturation. The earlier acquired 19 multispec- formation. The average accuracy achieved was 91 % for
tral images were then used to obtain the colour fea- non-PCA and 88 % for PCA.
tures. Various color description models were employed Yang et al. [30] developed an algorithm for the iden-
to transform the multispectral data into colour data. By tification of frass on ripened tomatoes using a multi-
calculating the individual spectral bands, the texture spectral imaging algorithm. The hyperspectral fluores-
features were obtained. The 19 spectral wavelengths in cence images of tomatoes were captured using hyper-
the range of 365 nm to 970 nm. The mean reflectance of spectral line-scan imaging set up [31]. The wavelength
each seed lying within the range of the spectral bands range was approximately 460–800 nm. This resulted in
was extracted with the help of the MultiColorMean fea- the selection of 5 wavelengths that were used in 3 ratio
ture. A trimmed mean of the transformed pixel values functions, R515/640, R724/690, and R664/690. The re-
was calculated using RegionMSI Mean. Similarly, the sults of the algorithm show strong detection of 100 % of
percentage of the blob region that had a transforma- 0.2 kg/L and 0.1 kg/L frass contamination spots on the
tion value higher than the threshold was calculated us- surface of tomatoes. For non-contaminated tomatoes,
ing RegionMSIthresh. the rate of frass detection was in the range between 86
Data Evaluation: % and 95 %.
LDA method was used to develop two models that
could differentiate between infested and non-infested 2.5 Fruit or Flower Detection
seeds from the spectral information. The reflectance
data obtained for each seed was used to create the first Vasas et al. [32] developed a technique for edge detec-
model. Similarly, colour and texture features were used tion in bees in which multispectral images of flowers
to create the second model. The data was split into show the importance of utilizing long-wavelength sen-
two sets with 60% of each sample used for training the sitive receptors. With the help of a multispectral im-
model and the remaining 40% of each sample for test- age data set of flowering plants, SNR ratios of long-
ing the model. The result of the model was evaluated wavelength (L) receptor reply were four times higher
using the kappa coefficient and accuracy. The Videome- than the short (S) and medium (M) wavelength recep-
terLab4 and R software were used for statistical evalu- tors depending on the specific conditions. The band-
ation. pass filters included the full wavelength range of bee
Bagheri et al. [27] introduced the detection of FB in- vision which had transmission peaks at 340, 400, 460,
fected pear trees using aerial remote sensing technology. 520, 580, 640 and 700 nm. As shown in Fig. 14 honey-
Aerial imaging of 25 healthy trees, 25 symptomatic dis- bees and bumblebees collect visual data with the help
eased trees and 25 trees with nutrient deficiencies was of 3 types of photoreceptors [34]. However, only a single
done for this study to differentiate FB infected trees type of receptor is used that responds to longer wave-
from other trees. Fig. 12(b) shows the False-colour com- lengths that are used for movement and edge detection.

18 Shilpa Gaikwad
Fig. 14 A series of monochromatic filters having different peak transmissions are used to capture black-and-white images.
Further, a false-colour (FC) image is produced from these photos to correct the white balance from the images. These FC
images are later tinted with black and white images to get a proper hue. (a) The 3 layers i.e. RGB are combined for human
vision. (b) Transformation of yellow, blue and UV into red, green and blue is done for the bees. The receptor response is
calculated for the (c) short (d) medium (e) and long-wavelength receptors which explained the electrical signals that the
photoreceptors in bees generate in response to the image [32].
The results show peak sensitivity at 544 nm which cor- were manually selected for comparing the reflectance
responds to green for long-wavelength receptors which spectra. A total of 52 images were taken for this study.
give the most steady signals in response to the natural 1000 pixels were randomly selected from the different
objects. parts of the flower using which the standard deviation,
mean and SNR were calculated. To understand the dif-
ference between the vision of humans and bees, false
2.5.1 Block diagram to detect long wavelength receptor coloured images were added to the database. The black
responses of bees: and white images were tinted to the hue set by the filter
wavelength with the help of ImageJ. The resultant im-
The database consisting of 53 images was scanned and ages were combined with the RGB colour images. The
processed in ImageJ as shown in (Fig. 15). Java plugin combination of RGB layers produces an original colour
colour correction was used to adjust the white balance. in human vision. However, for bees, the UV, Blue (B)
The normalized cross-correlation (NCC) method was and yellow (Y) layers are tinted with B, Green (G) and
used to separate the images of the same flower. Different Red (R) to produce false colour images.
parts of the flowers such as petals, centres and leaves

Fig. 15 Block diagram of Fruit/Flower detection - Long-wavelength sensitive receptor for bees. The database consisting of
53 images were scanned and processed in ImageJ. Java plugin colour correction was used to adjust the white balance. The
normalized cross-correlation (NCC) method was used to separate the images of the same flower. Different parts of the flowers
such as petals, centres and leaves were manually selected for comparing the reflectance spectra. A total of 52 images were taken
for this study.1000 pixels were randomly selected from the different parts of the flower using which the standard deviation,
mean and SNR were calculated. To understand the difference between the vision of humans and bees, false-coloured images
were added to the database. The black and white images were tinted to the hue set by the filter wavelength with the help
of ImageJ. The resultant images were combined with the RGB colour images. The combination of RGB layers produces an
original colour in human vision. However, for bees, the UV, blue and yellow layers are tinted with blue, green and red to
produce false colour images. For calculating receptor responses the SD and mean were calculated for the quantum catches (P)
and relative receptor responses (E) which are functions of receptor peak sensitivities. To understand the bee’s vision, heat
maps were included which show the receptor responses of the long, medium and short wavelength for each multispectral image.
Calculating Receptor Responses: ated. These results were combined and given as input to
the null model. To get a more clear picture of the bee’s
The Standard deviation (SD) and mean were calcu- vision, heat maps were included which show the recep-
lated for the quantum catches (P) and relative receptor tor responses of the long, medium and short wavelength
responses (E) which are functions of receptor peak sen- for each multispectral image.
sitivities. Lastly, the effects of reflectance functions, illu-
mination spectrum, background reflectance spectrum, A real-time system was developed by Lohumi et al.
the shape of the sensitivity of the receptor were evalu- [33] for the identification of foreign materials (FMs)

20 Shilpa Gaikwad
Fig. 16 Colour images of cabbage samples mixed with foreign materials (FM) are shown in the top row. Classification images
based on fluorescence are shown in the middle row. Classification images based on RGB are shown in the bottom row [33].
mixed with fresh-cut vegetables, with the help of flu- Fresh cut samples and FMs:
orescence and colour imaging. The setup included a In this work, cabbage and green onion samples were
multi-spectral fluorescence imager merged with an Liq- used in response to the fluorescence and change in colour.
uid crystal tunable filter (LCTF) to capture desired The FMs that were added to the freshly cut vegetables
band images sequentially of fresh-cut vegetables. Fig. are plastic pieces, peel residues, small metal pieces, and
16 shows the colour images of FM with cabbage sam- various kinds of woods pieces, toothpicks, etc.
ple. The detection accuracy (average) of FMs in cab- Real time detection system design:
bage and green onion sample was considered. The aver- White LED illumination was used to capture the
age of the total detection accuracy for 4 repetitions was colour images of the samples on the conveyor unit. Flu-
calculated to get the detection accuracy, which outper- orescence and colour imaging units were placed in such
formed 95%. A processing unit for vegetables that were a manner that they did not interfere with each other.
fresh-cut was installed in an industrial environment to The movement of the conveyor belt was controlled using
further test the real-time detection system. The per- a motion controller board and a computer unit and syn-
formance of the developed real-time system was almost chronized with the sensing unit. This was done to cap-
the same in the industrial condition. The system can ture the fluorescence images of the samples placed on
scan an area of a maximum of 24 × 24 cm2 in a time the conveyor unit. This was followed by, capturing the
duration of approximately 1.5 s. color images of the same area when the sample moved
one step ahead, which corresponds to approx. 24 cm, of
2.5.2 Block diagram to identify FM in fresh cut the Field of View (FOV) of the color camera.
vegetables Classification using optimal waveband selec-
tion:
As shown in Fig. 17 the LCTF based spectral imager The PCA and ANOVA methods were used to reduce
was employed to focus a UV-A beam light onto the tar- the volume of data and select the optimal wavelength.
get area. A built-in white light was used to provide an In this study, differentiation of vegetables and FMs on
alternative for the changeover. Thus, both reflectance the conveyor belt as well as the differentiation of the
and fluorescence images could be performed with this vegetable samples from the FMs was required. Thus,
setup. the best wavelengths were chosen to visually inspect

Fig. 17 Detection of foreign material in vegetables. The LCTF based spectral imager was employed to focus a UV-A beam
light onto the target area. A built-in white light was used to provide an alternative for the changeover. Thus, both reflectance
and fluorescence images could be performed with this setup. Cabbage and green onion samples were used in response to the
fluorescence and change in colour. The Foreign material (FMs) that were added to the freshly cut vegetables are plastic pieces,
peel residues, small metal pieces, toothpicks, etc. In a real-time detection system, design white LED illumination was used to
capture the colour images of the samples on the conveyor unit. Fluorescence and colour imaging units were placed in such
a manner that they did not interfere with each other. The PCA and ANOVA methods were used to decrease the volume of
data and select the optimal wavelength for classification. The maximum peak intensity centered at a wavelength of 465 nm
was chosen from the fluorescence spectra of the conveyor belt. The fluorescence peak intensity value was either 465 nm or
615 nm for different FMs. In fluorescence image-based detection two tests were performed with an algorithm to detect the
FMs. A threshold value was set in each case of the two tests by examining the smallest number of resulting false positive
and false negative pixels and the classification was performed. In colour imaging-based detection initially, the RGB image was
transformed into a gray scale image and the image was subtracted from each RGB channel. The threshold values which were
set were applied to each subtracted image to detect the small FMs of corresponding colours. The combination of fluorescence
and colour imaging gives better accuracy.

22 Shilpa Gaikwad
Fig. 18 Block diagram of Internal fruit parameter: Strawberry fruit (firmness and TSS). For the preparation of the sample
and multispectral imaging, 210 strawberry fruits were considered (70 unripened and 40 riped). Multispectral images were
captured for all the fruits. To generate the overripe fruit variety, the riped fruits were kept at room temperature for two days.
A texture analyzer was used to perform penetration tests on the skin of the fruit with a depth of 7 mm. The firmness of
the fruit was also calculated. The TSS was measured with a hand-held refractometer. The measurement of each fruit was
done thrice immediately after the multispectral image measurement. The video metric lab equipment was used for capturing
the multispectral images of the strawberry fruit. The nCDA algorithm was used to remove the background information and
segmented using a simple threshold. The images of strawberries, after removal of the background, were transformed into spectra
depending on the calculations of the mean. Each image was contributing its spectra for model calibration. Data evaluation the
model was built by considering the acquired spectra and the quality features like TSS and firmness. The SVM, PLS, BPNN
methods were used to develop the prediction model.
the fluorescence spectra of the conveyor belt, fresh-cut ing technique. By counting the mutually independent
vegetables and FMs. The maximum peak intensity cen- detected FMs from the RGB-based and fluorescence-
tered at a wavelength of 465 nm was chosen from the based classification images, the combined accuracy was
fluorescence spectra of the conveyor belt. The fluores- calculated. In this work, combined detection accuracy
cence peak intensity value was either 465 nm or 615 nm (average) for 4 repetitions of the test exceeded 95%.
for different FMs. The fluorescent FMs have been recog-
nized using these 2 band wavelengths. But the 465 nm
waveband created a problem for fluorescent FMs as the 2.6 Internal Fruit Parameters
images were partly covered by the fluorescent signals of
fresh-cut vegetables. To solve this problem, a 435 nm Changhong Liu et al. [35] proposed the use of Multi-
wavelength image was used where a little variation is spectral Imaging to find quality parameters and ripeness
observed between the fluorescent intensity of fresh-cut phases of the strawberry fruit. For this study, a total of
vegetables and the FMs. Later these two band images 210 fruits were included. For developing models, the cal-
were averaged (435 nm and 465 nm) and given a thresh- ibration set included 162 fruit which comprised of un-
old to check the existence of fluorescent FMs. ripe, ripe and overripe samples with 54 samples each.
Flourescence image-based detection: For model validation, 48 fruit were considered which
In this work, two tests were performed with an al- comprised unripe, ripe and overripe samples with 16
gorithm to detect the FMs. A threshold value was set samples each. This was done so that the calibration
in each case of the two tests by examining the small- model could verify the prediction performance. The re-
est number of resulting false positive and false negative sults show that for predicting firmness and TSS content,
pixels. Based on the threshold, the classification of the the back-propagation neural network (BPNN) perfor-
FMs mixed with fresh-cut vegetables was done accu- mance was good as compared with partial least squares
rately with minimum error. (PLS) and SVM using multispectral imaging. The SVM
Color imaging-based detection: and PCA-BPNN models were tested to classify fruit
Initially, the RGB image was transformed into a based on the maturity stage with the help of multi-
gray scale image and the image was subtracted from spectral imaging. The SVM model achieved a classifi-
each RGB channel. The threshold values which were cation accuracy of 100 %.
set were applied to each subtracted image to detect the
small FMs of corresponding colours. For the smallest 2.6.1 Block diagram to find internal fruit parameters
number of resulting false positive and negative pixels, (TSS and firmness) in Strawberry fruit
the best threshold value was selected in each case. It
was observed that the accuracy of the colour imaging- Sample preparation and Multispectral imaging:
based detection method is less than that of fluorescence In this work, 210 strawberry fruits were considered
imaging. (70 unripened and 40 riped) (Fig. 18). Multispectral
Combination of Color and fluorescence imag- images were captured for all the fruits. To generate the
ing: The combination of fluorescence and colour imag- overripe fruit variety, the riped fruits were kept at room
ing gives better accuracy. This technique gives much temperature for two days. A texture analyzer was used
higher accuracy as compared to that of the single imag- to perform penetration tests on the skin of the fruit

with a depth of 7 mm. The firmness of the fruit was These techniques were employed to compute the ho-
also calculated. The TSS was measured with a hand- mogeneity criteria related to the co-occurrence matrix.
held refractometer. The measurement of each fruit was These methods also helped to find the level of similar-
done thrice immediately after the multispectral image ity in a set of pixels. The comparison between all three
measurement. The video metric lab equipment was used methods was performed to select the one which gave the
for capturing the multispectral images of the straw- best results for measuring the ripeness of the banana.
berry fruit. Segmentation of the images into proper re-
gions was done for image evaluation. The nCDA algo-
rithm was used to remove the background information 2.8 Fruit Yield
and segmented using a simple threshold. Otsu adap-
tive thresholding method was used to emphasize the Cao et al. [38] developed a Unmanned aerial vehicle
required features. This was followed by segmentation. (UAV) multi-spectral images for monitoring growth in-
The images of strawberries after removal of the back- dicators of sugar beet. For this purpose, a wide dynamic
ground were transformed into spectra depending on the range vegetation index (WDRVI) is used. The values
calculations of the mean. Each image was contributing achieved for the determination coefficients (R2 ) for leaf
its spectra for model calibration. area index (LAI), Fresh Weight of leaves (FWL), and
Data evaluation: Fresh Weight of Roots (FWR) models of the sugar beet
The model was built by considering the acquired created using the WDRVI were 0.957, 0.950, and 0.963,
spectra and the quality features like TSS and firmness. respectively. The accuracy of growth indicators of sugar
The SVM, PLS, BPNN methods were used to develop beet can be improved from 1.05 % to 5.07 % using the
the prediction model. WDRVI index model. The relation between the ground
biomass of sugar beet and the growth indicator helps
in finding the accuracy of the growth indicator mon-
2.7 Fruit Ripeness or Maturity itoring. The increase in the ground biomass of sugar
beet reduces beet growth indicator, improving the ac-
With the help of Multi-spectral Imaging, a method of
curacy of growth monitoring by using WDRVI satura-
the Ripening Process in Bananas is proposed by San-
tion. There is huge biomass in sugar beet and thus, this
toyo et al. [36]. The brown spots on the banana peel
technique proves to be useful.
can be separated accurately by employing the Hotelling
transform. Two optical filters that were visible in the
range of 410-690 nm and NIR in the range of 820-910 2.8.1 Block diagram to find sugar beet growth
nm were used for this purpose. The texture homogene-
ity criteria were used to specify the growth of the brown A small UAV was used for aerial remote sensing in this
spots present in process of ripening with the fusion of work as shown in Fig. 22 a). This comprised of MicaS-
spectral image. Fig.19 shows the 7 phases of the banana ense Red Edge-M multi-spectral sensor and a passive
ripening process. For this study, 7 samples of bananas light optical sensor. The images of sugar beet were cap-
were used in the 7 days. For each sample, 10 images tured during the crucial growth period which consists
were taken every day, one for each optical filter result- of seedlings, leaf plumps growth, root growth and accu-
ing in a total of 490 images. mulation period of sugar. The pre-processing of mul-
tispectral data was done to remove noise, distortion
2.7.1 Block diagram to find banana ripeness in sensor, removal of background information and ra-
dial correction. 20 points were considered in the test
In this work, seven specimens of bananas were taken. area. In addition to this, a light sensor, GPS, inertial
Each day 10 images were taken for every specimen (one measurement unit was utilized to get the exact spa-
for every optical filter) (as shown in Fig. 20). This pro- tial data required. WGS84 was used as a geographic
cess resulted in 490 images. The banana specimens were coordinate system. By adding the weight coefficient to
kept for 7 days. 10 different optical filters were used NDVI, four WDRVI indices were calculated. This was
in this experiment. The flat-field correction technique done to evaluate the LAI, FWL and FWR of the study
was applied to remove the unwanted signals and retain plots. The sensitivity evaluation was performed based
the useful data for each multispectral image. To esti- on the five vegetation indices and three growth indi-
mate the maturity of the banana, quantification of the cators when NDVI > 0.8. From the results, it was ob-
brown spots growing on the peel was carried out. Three served that the WDRVI index is more sensitive than
quantifying methods namely Fourier fractal analysis, NVDI. It was also seen that the correlation coefficient
Hotelling transform, and texture analysis were used. between the beet growth index and the WDRVI index

24 Shilpa Gaikwad
Fig. 19 (a) Seven phases of the ripening process of the banana. (b) Example of the images used during the 1 week which
shows the colour image of the sample. (c) Example of the images used during the 1 week that shows gray scale images from a
specified spectral region for the same sample [36].
Fig. 20 Block diagram of Fruit ripeness: banana fruit. Seven specimens of bananas were taken for measuring the ripeness.
Each day 10 images were taken for every specimen (one for every optical filter). This process resulted in 490 images. The banana
specimens were kept for 7 days. 10 different optical filters were used in this experiment. The flat-field correction technique
was applied to remove the unwanted signals and retain the useful data for each multispectral image. To estimate the maturity
of the banana, quantification of the brown spots growing on the peel was carried out. Three quantifying methods namely
Fourier fractal analysis, Hotelling transform, and texture analysis were used. These techniques were employed to compute the
homogeneity criteria related to the co-occurrence matrix. These methods also helped to find the level of similarity in a set of
pixels. The comparison between all three methods was performed to select the one which gave the best results for quantifying
the ripeness of the banana.
is higher than that of NDVI. The models were devel- using an Unmanned aerial vehicle (UAV). The classifi-
oped using the data from the years 2018 and 2019. 70% cation results with the help of multispectral imaging for
of the data was used to develop the model and 25% the rice plants with different window sizes are shown
was used for verifying the model. RMSE, relative root in Fig. 21. The results show that texture features and
mean square error (RRMSE), R2̂ were calculated. The spectral features (TS) of decision tree (DT). The clas-
results show that the WDRVI1 index with a weighted sification accuracy achieved by TS-DT was high. The
coefficient of 0.05 accurately evaluated the LAI, FWL, accuracy achieved was in the range 5.2-26.7 % in OA
and FWR. WDRVI1 evaluated the FWL in the early and KC in the range of 0.06–0.40, irrespective of various
growth phase in the best way. growth phases and UAV images.
Zheng et al. [16] used RGB, NIR-G-B and multi-

spectral images for early growth detection of rice plants

Fig. 21 Multispectral images with windows of different sizes. Results before the canopy closure are shown in the top row.
Results after the canopy closure are shown in the bottom row. The rice plants are represented by the green area while the
soil/water background is represented by the black area [37]
2.8.2 Block diagram to detect rice plant growth The first step in optimal spectral feature involved
setting a threshold for classifying the UAV images. The
Image capturing using UAV and sensors: In this threshold varied for every spectral feature. The average
work, the UAV was used to capture images under spe- (C.V) value was calculated for the entire image and
cific weather conditions (Fig. 22 (b)). However, the spa- then the different ranges for the threshold were com-
tial resolution varied for each image. The ground spa- puted which had different values. For example, 0.1 C.V
tial distance (GSD) for the RGB image, NIR image (mean) to 2.0 C.V (mean). Once the threshold was set
and the MS image was 1.3 cm, 1.4 cm and 5.4 cm for every spectral feature, the value remained the same
respectively. Due to ultra-high spatial resolution, the over various growth phases.
processing speed for RGB and NIR images was quite Texture feature calculation:
low. The image processing efficiency was enhanced by The optimal window size for texture evaluation was
re-sampling the images. The spatial resolution of RGB estimated by calculating texture features (related to
images was increased by 3 times while for NIR images spectral feature) with help of 25 window sizes for the
it increased by 4 times. images of UAV. The classification accuracy was com-
pared to find the best window size and best spectral
Image processing:
feature for the respective UAV image. After this, the
The orthophotos were generated by processing the classification of the image was done in three classes -
RGB and NIR images with Agisoft PhotoScan Pro- rice plant, road and background (water and soil).
fessional 1.2. Four vegetation indices were calculated Classification of mixed class and pure class:
to classify using normalized digital numbers (DN). An Initially, the classification of images was done into
empirical line correction method was used to process two classes- mixed class and pure class. This was done
the MS images. This was followed by a computation by comparison of the texture feature (C.V) in each
of NDVI for MS image classification. In this work, the moving window. The mixed class had a higher (C.V)
image classification was done by calculating the mean value than that in the pure class. The classification of
and variance (C.V). For the MS images, the DN and the mixed class was done into rice plants and back-
reflectance images were employed for texture evalua- ground by comparison of every pixel value with the
tion whereas only the DN was employed for RGB and mean value in the moving window. For the mixed class,
NIR images. The Environment for visualizing images when the mean value was greater than the pixel spec-
(ENVI) software package was utilized for texture eval- tral value, the image was grouped into the background
uation and processing. class. Whereas when the mean value was lower than
Spectral feature and window size determina- the pixel spectral value, the image was classified into
tion: the rice plant class. Finally, the comparison of the DN

26 Shilpa Gaikwad
Fig. 22 Block diagram of Fruit yield (a) Sugar beet growth indicator. A small UAV was used for aerial remote sensing.
This comprised of MicaSense Red Edge-M multi-spectral sensor and a passive light optical sensor. The pre-processing of
multispectral data was done to remove noise, distortion in sensor, removal of background information and radial correction. 20
points were considered in the test area. In addition to this, a light sensor, GPS, inertial measurement unit was utilized to get
the exact spatial data required. WGS84 was used as a geographic coordinate system. By adding the weight coefficient to NDVI,
4 WDRVI indices were calculated. This was done to evaluate the LAI, FWL and FWR of the study plots. The sensitivity
evaluation was performed based on the 5 VI’s and 3 growth indicators when NDVI > 0.8. From the results, it was observed
that the WDRVI index is more sensitive than NVDI. Linear regression was used to build models using tested data from 2018
and 2019. 70 % of the data was used to develop the model and 25 % was used for model verification. The results show that
the WDRVI1 index with a weighted coefficient of α = 0.05 accurately evaluated the LAI, FWL, and FWR. (b) Rice plant
growth. For image processing the orthophotos were generated by processing the RGB and NIR images with Agisoft PhotoScan
software. 4 vegetation indices were calculated to classify using normalized digital number (DN). An empirical line correction
method was used to process the MS images. This was followed by a computation of NDVI for MS image classification. The
image classification was done by calculating the mean and variance (C.V). For the MS images, the DN and reflectance images
were employed for texture evaluation, whereas only the DN was employed for RGB and NIR images. The ENVI software
package was utilized for texture evaluation and processing. For optimal spectral feature and window size determination, the
first step in optimal spectral feature involved setting a threshold for classifying the UAV images. The average (C.V) value was
calculated for the entire image and then the different ranges for the threshold were computed which had 20 different values.
In texture feature calculation the best window size for texture evaluation was estimated by calculating texture features with
help of 25 window sizes for the images of UAV. The classification accuracy was compared to find the best window size and
best spectral feature for the respective UAV image. After this, the image was differentiated into 3 classes- rice plant, road and
background (water and soil). Initially, the classification of images was done into two classes- mixed class and pure class. This
was done by comparison of the texture feature (C.V) in each moving window. The mixed class had a higher (C.V) value than
that in the pure class. Finally, the comparison of the DN and reflectance value in the red band was considered to classify the
pure class into rice plant and road [37].

Fig. 23 Block diagram of Tomato Yield: The field plot consisted of 4 individual tomato plants with 3 different planting dates.
For the yield prediction model and estimation purpose, the total yield of the 3 harvests was considered. A structure from
motion (SfM) was used to produce orthomosaic images from UAS data. The UAS platform and sensors were used to capture
RGB images as well as 3 bands which consisted of green, red, and NIR spectra. The developed system collected RGB and
multispectral images. In UAS data collection, an Iris platform was used to acquire RGB images for 3 months whereas the
multispectral UAS images were taken every other week. The SfM algorithm is used to produce the ground sampling distance
of the orthomosaic image which is the resolution. To produce DSM and orthomosaic images from the UAS raw images, the
PhotoScan Pro software was employed. 3D point clouds, DSMs and orthomosaics were produced using the SfM algorithm. The
PhotoScan Pro software produced the DSM and orthomosaics of RGB as well as multispectral images. To compute VI with
high accuracy, radiometric calibration was done. The empirical line method (ELM) was employed to transform the DN in the
multispectral orthomosaic image into the reflectance value. To capture the field-based phenotypic data, rectangular polygons
were created as field borders for each tomato variety. Dimension of 1 × 2 meters was used to create the rectangles. The field
boundary covered all 4 individual plants. To capture the phenotypic information, the pixels inside the polygon were selected
of an individual variety. A total of 81 polygons (9 cultivars x 3 replications × 3 mulches) was considered. From the RGB,
multispectral orthomosaic images and DSMs, CCs, CV, ExG, NDVI were produced. Advanced phenotypic features were used
as multi-temporal phenotypes. To model the growth structure of each variety, time-series calculations of these phenotypes
were done. From the time-series UAS measurements, 22 phenotypic features were taken from growth and progression models.
The correlation coefficients were calculated for phenotypic features and the actual yield in the 81 field plots. To estimate the
tomato yield, linear regression was used for 81 tomato varieties. A yield estimation model was developed with the help of a
total of 22 phenotypic features from UAS data.
and reflectance value in the red band was considered to variables selection value is 0.79, forward selection value
classify the pure class into rice plant and road. When is 0.7, and backward selection value is 0.77. The two
the DN value or reflectance value in the R band was important factors which are dependent on high-yield
high, the image was grouped as road. For lower DN varieties are growth speed and timing. The yield pre-
or reflectance value, the image was classified as a rice diction model selected is five time-series phenotypes.
plant. The variance achieved is 65 % in the real harvest. The
phenotypic features which are extracted from RGB im-
Chang et al. [39] proposed the tomato yield estima- ages play a major part in prediction yield. The average
tion method using an unmanned aircraft system (UAS) tomato yield for the rest of the varieties was increased
and High-Throughput Phenotyping (HTP). The linear by 53% using UAV data using the proposed method.
regression model shows that the R2 values achieved to
be high by using different variable selection methods.
The values obtained for the 3 selection methods are: all

28 Shilpa Gaikwad
2.8.3 Block diagram to detect tomato yield variety. Dimension of 1 X 2 meters was used to create
the rectangles. The field boundary covered all four indi-
Study Area: vidual plants. To capture the phenotypic information,
The field plot consisted of 4 individual tomato plants the pixels inside the polygon were selected of an indi-
with 3 different planting dates (Fig.23). For each field vidual variety. A total of 81 polygons (9 cultivars x 3
plot, the tomatoes were harvested three times. For the replications × 3 mulches ) was considered.
estimation of yield prediction model, the total yield of Geospatial Product Generation:
the 3 harvests was considered. A total of 9 Ground From the RGB, multispectral orthomosaic images
control points (GCPs) were deployed for this study. A and DSMs ,Canopy cover (CC), canopy volume (CV),
structure from motion (SfM) was used to produce or- excessive greenness index (ExG), NDVI were produced.
thomosaic images from UAS data. Feature Extraction using Time-Series UAS
UAS Platforms and Sensors: Data:
The UAS platform and sensors were used to capture Advanced phenotypic features were used as multi-
RGB images as well as three bands which consisted temporal phenotypes. From these phenotypes,time-series
of green, red, and NIR spectra. The developed system calculations were performed to model the growth struc-
collected RGB and multispectral images. ture of each variety.
UAS Data Collection: Phenotypic Feature Extraction :
The operation of RGB and the multispectral UAS From the time-series UAS measurements, 22 pheno-
system took place individually with the help of two typic features were taken from growth and progression
flight teams. As the specifications vary for the field of models.
view (FOV) and focal length for both RGB and mul- Calculation of Correlation Coefficient:
tispectral cameras. Another Iris platform was used to The correlation coefficients of phenotypic features
acquire RGB images for 3 months whereas the multi- and the actual yield in the 81 field plots were calculated.
spectral UAS images were taken every other week. The Yield Estimation Modelling:
SfM algorithm is used to produce the ground sampling To estimate the tomato yield, linear regression was
distance of the orthomosaic image which is the resolu- used for 81 tomato varieties. The model was developed
tion. by considering the dependent variable as the actual
UAS-Based HTP: tomato yield and the independent variable as the cor-
To produce DSM and orthomosaic images from the responding multi-temporal phenotypic feature.
UAS raw images, the PhotoScan Pro software was em- Factor study for UAS based Phenotypic Fea-
ployed. 3D point clouds, digital surface models (DSM) tures:
and orthomosaics were used to produce the SfM algo- A yield estimation model was developed with the
rithm. However, this algorithm has a drawback called help of a total of 22 phenotypic features from UAS
the bowl effect. To solve this problem, GCPs with high data. Two feature selection methods were used: the for-
accuracy were used. The PhotoScan Pro software pro- ward feature selection method and the backward fea-
duced the DSM and orthomosaics of RGB as well as ture elimination method. The value of R2 was greater
multispectral images. than 0.7 between the estimated and actual yields for the
Radiometric Calibration for Multispectral Im- yield estimation models. The factor evaluation method
age: picked 9 phenotypic features which were given as an in-
To compute vegetation indices with high accuracy, put to the linear regression model to find the tomato
radiometric calibration was done. The empirical line yield. The final model employed 5 phenotypic features
method (ELM) was employed to convert the DN in the and a forward feature selection method. The tomato
multispectral orthomosaic image into the reflectance yield is calculated using the formula:
value. For this purpose, four reflectance panels were
mounted in the tomato field. After this, a comparison T omatoY ield = 1.841 ∗ F 10 + 0.875 ∗ F 18 − 2.546
of the actual reflectance value which was measured in (1)
∗ F 3 + 2.026 ∗ F 7 − 0.902 ∗ F 8
the lab was done with and the average DN of every re-
flectance panel of the multispectral orthomosaic image. where, F3 = days after planting (DAP) when the
The EL converted the whole pixel value of the multi- canopy cover(CV) growth rate was maximum F7 =
spectral orthomosaic image into reflectance. DAP when the canopy volume (CV) growth rate was
Product Generation from UAS Data: maximum F8 = duration of the half maximum CV pe-
To capture the field-based phenotypic data, rectan- riod F10 = DAP when the ExG value was maximum
gular polygons were created as field borders for each F18 = triangle area of the decreasing ExG slope

3 Results 245 spectra of cream samples were considered. 18 dif-

ferent wavelengths in the range of 405 nm to 970 nm
The results are summarised with the help of tables [Ta- were taken. The results show fresh samples with TVC
ble 1 to Table 5] which concludes all the 8 objectives of ≤ 2.0 log CFU/g and spoiled samples with TVC ≥ 6.0
multispectral imaging of this review paper in terms of log CFU/g. Overall classification accuracy is 91.7% for
their database size, number of colour spectrums, accu- 2 classes in model validation [20].
racy and other factors.For varietal identification in rice Classification
seeds 250 rice seed samples was considered and 19 dif- For the classification of Medicinal plant leaves, the
ferent wavelengths were used. The lowest classification size of the digital image data set of 100 × 6 = 600
accuracy achieved was 62% whereas the highest classi- coloured images was used. Six different varieties of fresh
fication accuracy was 94% with PLS-DA and LS-SVM leaves of medicinal plants are taken. Five bands with a
model respectively [1]. range of 460 nm to 1560 nm are known as Red (R), Blue
Varietal identification (B), Green (G), (NIR) and Spectral Bands Shortwave
For varietal identification of tomato seeds, 2 sample Infrared (SWIR). The results show distinct classifica-
sets were considered. Sample set 1 had 50 to 55 seeds tion accuracy by the MLP classifier on 6 medicinal plant
whereas Sample set 2 had 188 to 243 seeds. The low- leaves as Tulsi = 99.10% ,Peppermint = 99.80%,Bael =
est classification accuracy achieved was 82% for PLS- 98.40%,Lemon balm = 99.90%,Catnip = 98.40%, and
DA whereas the highest classification accuracy achieved Stevia = 99.20%.The MLP classifier has an accuracy of
was 96% for Step wise PLS-DA models on high (seven) 99.01%[21].For tree species classification into healthy,
sensitivity cultivars model [13]. unhealthy, and dead trees affected by infection of bark
Fruit Quality beetle two data sets were considered.The first data set
For fruit quality of pomegranate 100 samples were (spring 2019) had 153 images and the second data set
considered. The wavelengths used were 700, 800, 900, (summer 2019) had 144 images. Spectral bands with
and 1000 nm. The results show that the multiple linear centre wavelength (nm): B-475 ,G-560, R-668, NIR-
regression model predicted TSS with r = 0.97, RMSEC 840, RE-717. The results show tree species classifica-
= 0.21°Brix and RPD=6.7°Brix[19].For the quality of tion with an OA of 81.18% and Kappa value of: 0.70
Jatropha curcas seeds, 3 lots were considered with n = and health state evaluation with an OA of 84.71% and
100 seeds per lot. 19 different wavelengths were taken. Kappa value of 0.66 [22]. For the classification of a coral
The results show high accuracy in the range of 96% to reef using hybrid classifier ensembles, training data of
98%. The Crude fat (F) content greatly differed among 50 samples per class were built to allow training and
seed lots with F = 15.37, df = 2, 6, crude protein con- validation of the model. The ETM+1, ETM+2, and
tent (P)= 0.004.Lot 2 results show 51% and 35% more ETM+3 spectral bands in visual wavelength band were
fat than Lot 1 and Lot 3 respectively [14]. To find out captured using the satellite. The Multi-class voting scheme
optical reflection changes in wok-fried vegetables,358 techniques showed the One-Against-the-Rest classifier
carrot samples and 389 celeriac samples were consid- accuracy of 78.66%, and One-Against-One classier ac-
ered. Whereas 20 different wavelengths in the range of curacy of 70.67% [23].
430 to 970 nm were taken. The results show that signifi- Disease detection
cant changes were observed in the reflectance spectrum For plant disease detection in citrus and peach or-
for both celeriac and carrots. And significant changes chards,2 databases were considered.Database-1 is cit-
were observed for the celeriac continuing until day 14 rus orchard at Fasa site which consists of 494 images.
[15].For crop hail damage evaluation in potato, the field Whereas Database-2 is peach orchard at Neka site which
was divided into 18 crop rows having a width of 86 cm consists of 760 images. Five wavelength spectral bands:
with two varieties such as Norkotah and Ranger planted R, G, B, RE, and NIR were taken.The results are as
at 25 cm apart and 20 cm deep. In 2014, the wave- follows-Citrus orchard:Overall accuracy (uncalibrated)
lengths used were NIR, G, and B while in 2015, the = 96.49 %, kappa value (uncalibrated) = 0.941, Over-
the wavelengths used were R, G and NIR. UAV-based all accuracy (calibrated) = 96.57 %, kappa value (cali-
multispectral imagery was used to evaluate the dam- brated) = 0.942. Peach orchard: Overall accuracy (un-
age severity in potato under 4 hail damage stages of calibrated) = 95.58 %, kappa value (uncalibrated) =
0%, 33%, 66% and 99% at 3 growth phases of TI, EB, 0.913, Overall accuracy (calibrated) = 95.56 %, kappa
and LB, in 2 potato types. The results show that VI value(calibrated) = 0.913 [24]. For the detection of rice
(GNDVI, NDVI, SAVI) at the EB had a good correla- sheath blight disease,67 field plots were considered. 5
tion of r > 0.77 with crop yield [16]. For the classifica- narrow bands - R, B, G, RE to NIR were taken. NDVIs
tion of vanilla cream into two microbiological quality, measured from the multispectral images could quan-

30 Shilpa Gaikwad
Table 1 Varietal Identification and Fruit Quality

Sr. Reference Number of
Database size Accuracy (%) / Results
No. paper colour spectrums
Varietal Identification
19 wavelengths :
405, 435, 450, 470, 505, LS-SVM model:94 %
1. Liu et al. [35] 250 rice seed samples 525, 570, 590, 630, 645, PLS-DA:62 %
660, 700, 780, 850, 870, PCA-BPNN: 84 %
890, 910, 940, and 970 nm
Classification accuracy
19 wavelengths : 375,
PLS-DA: 82 %
Sample set 1: 50 to 55 seeds 405, 435, 450, 470, 505,
Stepwise PLS-DA models on
2. Shrestha et al. [13] Sample set 2: 188 to 243 525, 570, 590, 630, 645,
high (7) sensitivity cultivars : 96 %
seeds 660, 700, 780, 850, 870,
Stepwise PLS-DA models on poor
890, 940 and 970 nm
(4) sensitivity cultivars: 86 %
Fruit Quality
TSS with r = 0.97
Khodabakhshian Wavelengths: 700, 800, RMSE of calibration = 0.21°Brix
1. 100 Samples
et al. [19] 900, and 1000 nm Ratio performance
deviation (RPD)= 6.7 °Brix.
Accuracy = 96% to 98%.
19 wavelengths: Crude fat (F) = 15.37,
365, 405, 430, 450, 470, df = 2, 6 , crude protein content
2. Bianchini et al. [14] n = 100 seeds per lot 490,515, 540, 570, 590, ( P )= 0.004 ,
630, 645, 660, 690, 780, Lot 2 results show 51% and
850,880, 940, and 970 nm 35% more fat than Lot 1 and
Lot 3 respectively.
Significant changes were observed
in the reflectance spectrum for both
20 different wavelengths
Clemmensen 358 carrot samples celeriac and carrots.
3. in the range of 430 to
et al.[15] and 389 celeriac samples Significant changes were observed
970 nm.
for the celeriac continuing until
day 14.
In 2014, the wavelengths were Vegetation indices
NIR, G, and B. (GNDVI, NDVI, SAVI) in
4. Zhou et al.[16] 18 crop
In 2015, the wavelengths were EB phases had a
NIR, R, and G. r >0.77 with crop yield.
Fresh samples: TVC ≤ 2.0 log CFU/g
18 different Spoiled samples: TVC ≥ 6.0 log CFU/g
245 spectra of cream
5. Lianou et al.[20] wavelengths 405 nm Overall classification
samples.
to 970 nm accuracy = 91.7 % for 2 classes
in model validation.
tify various degree of the disease in the field area with eased trees = 91.4%,OA for trees with Nutrient Defi-
an accuracy of 63% [25]. For detection of D.Avenea ciency = 93.9%. Producer’s accuracy for the healthy
fungi in black oat seeds, multispectral images of 200 trees = 87.7%,Producer’s accuracy for FB infected trees
non-contaminated seeds and 200 contaminated seeds = 54.9%,Producer’s accuracy for the trees with nutri-
with D.Avenae for each time interval of inoculation at ent deficiency = 98.8%,Operator’s accuracy for healthy
24, 72 and 120 h were taken at 19 wavelengths from trees = 72.2%,Operator’s accuracy for FB infected trees
UVA to NIR. The colour and texture feature accu- = 74.2%,Operator’s accuracy for trees with nutrient
racy was 86% . Correct accuracy of 78%, 83% and defiencies = 98.2% [27]. For the identification of frass
100% was achieved for classification of seeds inoculated on red tomatoes, a total of 5820 sample pixels were
for 24, 72 and 120 h,respectively [26]. For detection of chosen. The fluorescence intensities at 5 wavelengths -
fire blight infected pear trees, healthy trees-25, asymp- 515,640,664,690,and 724 nm were considered. The al-
tomatic diseased trees-25, nutrient deficiencies trees- gorithm detected 100% of 0.2 kg/L and 0.1 kg/L frass
25 were selected for aerial imaging.In Vis-NIR spec- contamination spots on the surfaces of tomatoes. The
tra, 9 wavelengths were taken. The results show Over- rate of successful identification was in the range of 86%
all accuracies for SVM classification using 5-fold cross- to 95% for non-contaminated tomatoes [30].
validation as OA for Healthy trees = 98.7%,OA for Dis-
Fruit and Flower Detection

Table 2 Classification
Sr. Reference No of colour
Database size Accuracy
No. paper spectrums
The distinct classification
accuracy by the MLP
classifier on 6
medicinal plant leaves are
Five bands with range of 460
Size of digital image dataset of Tulsi = 99.10% ,
nm to 1560 nm known as Red
100 × 6 = 600 colored images were Peppermint = 99.80%,
1. Naeem et al.[21] (R), Blue (B), Green (G),
used. 6 different varieties of fresh Bael = 98.40%,
(NIR) and Spectral Bands
leaves of medicinal plants are taken. Lemon balm = 99.90%,
Shortwave Infrared (SWIR).
Catnip = 98.40%, and
Stevia = 99.20%.
The MLP classifier has an
accuracy of 99.01%.
Tree species classification with
Spectral bands with center an OA of 81.18%
First dataset (spring 2019): 153
Abdollahnejad wavelength (nm): B-475 , and Kappa coefficient= 0.70.
2. Second dataset (summer 2019) : 144
et al.[22] G-560 , R-668 , Health status evaluation
images
NIR-840 , RE-717 with an OA of 84.71%
and Kappa coefficient=0.66.
The Multi-class voting
scheme techniques:
ETM+1,ETM+2,and ETM+3
One-Against-the-Rest
3. Henriques et al.[23] Sets of 50 samples per class spectral bands
classifier accuracy = 78.66%,
captured using the satellite.
One-Against-One
classifier accuracy =70.67%
When using long-wavelength-sensitive receptors for For the ripening process in banana, 7 samples of
edge detection in bees, from each of the 53 images, 49 bananas were used in the seven days. Ten images were
species in the multispectral database was chosen. From taken every day, from each sample, one per optical filter.
this database, the number of leaf images was 34, the So a total of 490 images were processed. The wavelength
number of petals images were 52 and the number of cen- range was 270-1000 nm. The results show the Least
tres images were 38. The band-pass filters collectively square method with Correlation factor (r) for the worst-
included the full wavelength range of bee vision peak case as 0.74 (approx.) and Correlation factor (r) for
transmissions at 340, 400, 460, 520, 580, 640 and 700 the best case as 0.84 (approx.).The Hotelling transform
nm. The results show that the SNR of long-wavelength with correlation factor (r) for the worst-case as 0.84 and
receptor responses is 4 times higher than the short and the correlation factor (r) for the best-case is 0.74[36].
medium wavelength receptors [32]. For the detection Fruit Yield
of FMs in fresh-cut vegetables, 6 individual test sets
For monitoring sugar beet growth indicators, the
were performed for each fresh-cut vegetable. The flu-
test was performed thrice using 15 field areas. The wave-
orescence images of all fresh-cut vegetables and FMs
length (nm) selected was B-475, G-560, R-668, NIR-
were acquired in the 420–730 nm wavelength range.
840, RedEdge-717. The WDRVI1 index model achieved
The detection accuracy (average) of FM in cabbage and
accuracy in the range of 1.05% to 5.07% of sugar beet
green onion samples, measured by averaging the total
growth indicators. [38]. For early growth detection of
detection accuracy for 4 replications, was greater than
rice plants, a total of 36 plots/treatments were consid-
95 % [33].
ered. The spectral bands (nm) used were - for RGB
Internal Fruit Parameter camera R, G, B.For NIR-G-B camera: B (430-500 nm),
To find the quality parameters of the strawberry G (500-570 nm),NIR (670-770 nm). For Multispectra
fruit, a total of 210 fruits were included. 19 different camera: B (490 nm), G (550 nm), R (680 nm), RE
wavelengths were considered. The SVM model achieved (720 nm), NIR1 (800 nm), NIR2 (900 nm). The results
the classification accuracy of 100% [35]. show that Texture features which are mean and vari-
Fruit Ripeness ance (C.V) and spectral features (TS) of decision tree

32 Shilpa Gaikwad
Table 3 Disease Detection

Database size Accuracy (%)
No. paper spectrums
Citrus orchard:
OA (un-calibrated) = 96.49%,
kappa value (un-calibrated) = 0.941
Database 1: Citrus orchard OA (calibrated) = 96.57%,
Pourazar at Fasa site = 494 5 spectral bands: kappa value (calibrated) = 0.942
1.
et al.[24] Database 2: Peach orchard R, G, B, RE, and NIR Peach orchard:
at Neka site = 760 OA (un-calibrated) = 95.58%,
kappa value (un-calibrated) = 0.913
OA (calibrated) = 95.56%,
kappa value (calibrated) = 0.913
5 narrow bands:
2. Zhang et al.[25] 67 field plots NDVIs overall accuracy = 63%.
R, B, G, RE to NIR
19 wavelengths: The color and texture feature
Non-infested seeds = 200 365, 405, 430, accuracy = 86 %
Franca-Silva
3. Infested seeds with 450,470,490, 515, 540, 570, Correct accuracy of 78%, 83%
et al.[26]
D.Avenae fungi = 200 590, 630, 645, 660, 690, 780, and 100% for classes of seeds inoculated
850, 880, 940, 970 nm. for 24, 72 and 120 h, respectively.
OAs for SVM classification
using 5-fold cross validation:
OA for Healthy trees = 98.7%
OA for Diseased trees = 91.4%
OA for trees with Nutrient
Deficiency = 93.9%.
Producer’s accuracy for the
Healthy trees :25
healthy trees = 87.7%
Symptomatic diseased 9 Wavelengths: 500, 550,
Producer’s accuracy for FB infected
4. Bagheri et al.[27] trees: 25 570, 670, 680, 700,750, 800
trees = 54.9%
Nutrient deficiencies and 870 nm wavelengths.
Producer’s accuracy for the trees
trees: 25
with nutrient deficiency = 98.8%
Operator’s accuracy for healthy
trees = 72.2%
Operator’s accuracy for FB infected
trees = 74.2%
Operator’s accuracy for trees with
nutrient deficiencies = 98.2%.
Histogram of Hue (H),
Saturation (S),Intensity (I)
Average accuracy of non-PCA = 91.92%
5. Deng et al.[28] 898 samples and gray scale image was used
Average accuracy of PCA = 88.20%.
for color. LBP was used for texture
feature extraction.
The algorithm detected 100%
of 0.2 kg/L and 0.1 kg/L frass
The fluorescence intensities at contamination spots on the surfaces
6. Yang et al.[30] 5820 sample pixels 5 wavelengths - 515, 640, of tomato.
664, 690, and 724 nm The rate of successful identification was
in the range of 86% to 95%
for non-contaminated tomatoes.
(DT) i.e. TS-DT are OA for RGB is 91.25%, OA for 4 Discussion

NIR is 2.88%, OA for MS is 93.53%, KC for RGB is
0.86, KC for NIR is 0.86 and KC for MS is 0.88 [37].
For the tomato yield estimation method, 81 polygons
(9 cultivars × 3 mulches × 3 replications) were con- The main focus of this manuscript was to present an
sidered. Three wavelength bands Green, Red and NIR overview of how Multispectral imaging can be used for
were taken. The variety selection process using UAV various applications in fruits and vegetables. The refer-
data increased the tomato yield (average) of the rest of ence papers which are reviewed have applied technolo-
the varieties by 53 %. gies like chemometric methods, deep learning, machine
learning, imaging and cost-support vector classification

Table 4 Fruit/ flower detection, Internal fruit parameter, Fruit ripeness/ maturity.
Database size Accuracy ( % )/Results
No. paper spectrums
Fruit/ flower detection
From each of the 53 images
49 species in the multispectral SNR of long wavelength receptor
database were chosen. 340, 400, 460, 520 responses = 4 times higher
1. Vasas et al [32]
Number of Leaf images = 34 580, 640 and 700 nm. than the short and
Number of petals images = 52 medium wavelength receptors
Number of Centers images = 38
6 individual test sets were
2. Lohumi et al.[33] performed for each fresh-cut 420-730 nm spectral range. Accuracy greater than 95%.
vegetable.
Internal fruit parameter
19 wavelengths:
405,435, 450, 470,
505, 525, 570, 590, 630,
1. Liu et al.[1] 210 fruits SVM model accuarcy was 100%.
645, 660,700, 780, 850,
870, 890, 910, 940 and
970 nm.
Fruit ripeness/ maturity
Least square method:
Correlation factor (r)
for the worst case = 0.74 (approx.)
Wavelength range: 270- for the best case = 0.84 (approx.)
1. Santoyo et al.[36] 490 images
1000 nm Hotelling transform:
for the worst case = 0.84
for the best case = 0.74.
Table 5 Fruit Yield

Database size Accuracy (%) /Results
No. paper spectrums
Wavelength (nm): B-475, G-560, WDRVI1 accuracy can be
1. Cao et al.[38] 15 plots R-668, NIR-840, in the range of
RE-717 1.05% to 5.07%.
Spectral bands (nm):
For RGB camera: R, G, B.
TS-DT showed:
For NIR-G-B camera:
OA for RGB = 91.25%
B (430-500 nm), G (500-570 nm),
OA for NIR = 92.88%
NIR (670-770 nm).
2. Zheng et al.[37] 36 treatments OA for MS = 93.53%
For Multispectra camera:
KC for RGB = 0.86
B (490 nm), G (550 nm),
KC for NIR = 0.86
R (680 nm), RE (720 nm),
KC for MS = 0.88
NIR1 (800 nm),
NIR2 (900 nm).
81 polygons
The averaged tomato yield
3. Chang et al.[39] (9 cultivars × 3 mulches 3 wavelength band - G, R and NIR.
was increased by 53%.
× 3 replications)

34 Shilpa Gaikwad
with multispectral imaging to find the fruit and veg- were compared. The results show p-values for the t-test
etable traits with high accuracy. with the null hypothesis, that there is no variation in
mean values between days. The grey background repre-
sents days having significant differences at a 5% level.
4.1 Varietal Identification The statistical tests show that for carrots, a significant
change in the mean from days 2 to 4 for all wavelengths
In varietal identification of rice seeds, multispectral imag- up to 890 nm can be verified. This pattern was similar
ing techniques combined with various chemometric meth- for other percentiles than the illustrated 5th percentile.
ods, including PCA, PLS-DA, LS-SVM, PCA-BPNN For carrots, after day 4 we note only a few significant
were employed to separate the rice seeds based on their changes in the mean of the visible wavelengths and from
varieties. nCDA and PCA were used to discriminate days 12 to 14 in the NIR wavelengths. However, for
and identify the tomato cultivars for parents and off- celeriac, we can significantly track a change every 4–6
spring study. Further, all the cultivars were classified days. However, the changes are not as clear since these
with the help of PLS-DA [13]. statistics are based on pure wavelengths rather than ra-
tios of wavelengths. As for carrots, the most significant
Polythene bags are permeable for oxygen molecules and
4.2 Fruit Quality
thus, the oxidation in vegetables causes the variation
in spectra. The celeriac turns brown/ grey due to ox-
The TA, TSS and pH were measured to assess the
idation. Similarly, the carrot turns paler and brown.
quality of pomegranate fruit. This was done using vis-
An increase in brown/grey colour is due to a varia-
ible/NIR spectroscopy in the range of 400–1100 nm.
tion in a wide range of the spectrum as well as the
PLS was used to study the spectral data. A multispec-
brightness. There are significant colour changes from
tral imaging system was successfully utilized that was
days 2 to 4 throughout the visual spectrum for carrots
related to effective wavelengths. The performance of
and in a large part of the visual spectrum for celeriac
the developed system was examined using the multi-
[15]. To evaluate the crop hail damage in potato crops,
ple linear regression models [19]. Multispectral imaging
Aerial multispectral imaging along with a Green nor-
and X-ray data were used to characterize the Jatropha
malized vegetation index was used. For 2 potato va-
curcas seed based on its quality. nCDA algorithm was
rieties, 3 replicate treatments at each stage were per-
applied to achieve spatial and spectral patterns on var-
formed followed by which the mean of GNDVI was cal-
ious seed lots. With the help of the reflectance data
culated. The treatment effects in the range of 0–99%
and X-ray data that was based on LDA, the models
hail damage concerning imaging intervals, days after
to perform classification were developed. These mod-
damage (DAD) at different potato growth stages were
els showed high accuracy when the reflectance was at
considered. [16].To classify the vanilla cream based on
940 nm when operated individually or combined. This
the microbiological quality, multispectral imaging was
X-ray information was used to find the quality features
employed. An unsupervised machine learning approach
like normal seedlings, abnormal seedlings and dead seeds
was followed to build an automated selection method
at the same reflectance [14]. For Multispectral Imag-
with the help of multispectral images. The Gaussian
ing of Wok-Fried Vegetables, the Quality Index results
mixture developed model included a lot of variabili-
show that the total sum of the assessor scores is rather
ties as close to a real-life case. The model development
low for both carrots and celeriac, even after 14 days
was related to 2 different product samples (fresh and
of storage at +5° C. For the ANOVA evaluation of
spoiled) supplied directly by the vanilla cream man-
the scores of each question, the effect of day on the
ufacturer that was conserved under laboratory state.
discolouration of celeriac had a p-value of 0.088. In
Samples from various retail stores as well as the ex-
comparison, the p-value of the effect of day on the
pired samples were considered for testing and valida-
discolouration of carrots was 0.46. The discolouration
tion of the model. 96% of the samples with TVC ≤ 2.0
scores are related as a function of the days [40]. Only
log CFU/g were classified accurately with the UOS ap-
one other effect was significant at a 10% level of signifi-
proach. Thus, multispectral imaging spectroscopy can
cance, namely, the interaction effect of assessor and day
be widely used in the dairy industry to perform qual-
for the off-taste scores of celeriac. The p-value of the test
ity control. The microbiological quality of vanilla cream
was 0.074. It is, thus, possible that the discolouration
can be estimated fast and hence, the product can be in-
noted occurs simultaneously with a slightly noticeable
stantly released to the market [20].
off-taste. A total of 358 carrot samples and 389 celeriac
samples distributed over six different sample days were
tested for group-wise differences. Only successive days

4.3 Classification spectral images in coral surroundings. A hybrid SVM

classifier was employed to classify the coral reefs. For
this purpose Landsat ETM+ multispectral images were
For Medicinal plant leaves classification, using multi-
used [41]. The Hybrid-SVM classifier results were com-
spectral and texture features, five multi-layer (ML) clas-
pared with the results of the One-Against-the-Rest and
sifiers were used. The classifiers used are multi-layer
One-Against-One classifier based on SVM. High accu-
perceptron (MLP), LogitBoost (LB), Bagging (B), Ran-
racy was achieved by the classifier set as compared to
dom Forest (RF), and Simple Logistics (SL), deployed
the other two algorithms for all the tests performed [23].
on fused optimized features data set for the classifi-
cation of medicinal plant leaves. The foundation of the
data set holds six types of medicinal plant leaves named
Tulsi, Lemon Balm, Bael, Stevia, Peppermint and Cat- 4.4 Disease detection
nip. The medicinal leaves based on fused features are
performed using cross-validation (10-fold) data split- For plant disease detection, radiometric calibration in
ting approach. Different testing parameters such as “Re- UAV-based multispectral imagery was explored. For this
ceiver Operating Characteristic” (ROC), “Kappa Statis- work, 2 study sites with varying conditions: weather,
tics”, “False Positive (FP), “Recall” (R), “True Posi- specific day, different species of trees and disease kinds
tive” (TP), and “F-Measure” is observed. Firstly, an ex- were picked. Multispectral images were acquired using
periment performed on a region of observation (ROO’s) a RE camera. The multi-lens camera design and the
size (220×220) for the classification of medicinal plant low flying height resulted in the band-specific misregis-
leaves and observed a well-organized accuracy which tration error which was not minor. So, in the first step,
is 95.87%, 95.04%, 94.21%, 93.38%, and 92.56% using the error was registered. In the next step, to produce
MLP, LB, B, RF, and SL, respectively. We started our stable orthomosaics, the alignment of the registered im-
experimentation with the size of ROOs 220 ×220. Af- ages and normalization in radiometric calibration was
ter that, we gradually increased the size of ROOs to done. Radiometric normalization caused uniformity in
achieve better accuracy. Finally, at the size of ROOs the image captured over a wide variety of environmen-
280 × 280, we observe the promising accuracy because tal states. This step was necessary to generate a ho-
it covers maximum useful information. Further increase mogeneous orthomosaic. For Radiometric calibration,
in the size of ROOs was causing a decrease in the accu- two cases are considered. In the first case, a t-test was
racy due to speckle noise [21]. The species of trees were performed and compared with entropy distances to sep-
classified and the status of the health was assessed for arate the healthy tree classes from the unhealthy ones.
a mixed Broadleaf Conifer forest. This was done using In the other case, the accuracy of tree classification
UAS multispectral imaging. The study was to find the was found based on the radiometric calibration effects.
ability of an economical UAS bi-temporal multispectral In this study, the best features were selected from the
measurement. For this purpose, the RedEdge-M sen- tree classification criteria and correlated with random
sor was used. The UAS can obtain data in ultra-high forest classifier to detect diseased trees [24]. To detect
spatial and temporal resolutions. This resulted in the rice sheath blight (ShB) disease, a combination of high
classification of the tree species and its current health resolution unmanned aerial systems and multispectral
state. In addition, the UAS can also help in checking imaging was used. UAV provided RGB and multispec-
the spread of disease, finding the way for detection at tral images with high resolution [42]. These images were
an early stage of the eruption of a new disease to nearby processed and studied to identify the growth of ShB and
healthy trees. This would help in accurate forest man- measure/evaluate various degree of the disease severity
agement and specific decision-making. By combining in the field area. From the available image, a data com-
bi-temporal spectral and texture information as well parison was done for ShB intensity and NDVIs. From
as the machine learning techniques, the reflectance val- RGB images, the hue value was obtained that sepa-
ues were focused on a buffer area with a radius of 2 rated the infected areas from the clean and healthy ar-
meters from every treetop. The effect of crown shape eas in the field. The combination of a UAV and a cam-
and unclear pixels from the interconnected crown parts era with high resolution is an advancement that has
was removed within the same buffer area. The results the ability for fast and accurate detection of ShB [25].
show accurate and reliable tree species classification. For detection of D. Avenae fungi in black oat seeds,
This study showed that SVM was highly efficient with a multispectral imaging and machine vision are used. In
limited number of training samples. A vast data of var- this study, an nCDA model which combines spectral
ious types can be combined using this technique [22]. information from various seed health groups with their
A hybrid classifier set was used to classify the multi- respective reflectance, colour, and texture features is

36 Shilpa Gaikwad
used. The result of taking out multispectral informa- algorithm was used. A hyperspectral line-scan imaging
tion from each seed was that an accurate classification system that acquired hyperspectral fluorescence images
model was built that could perform classification based of tomatoes was used. These images were in the wave-
on the health status of the seeds. The nCDA algorithm length range of 460–800 nm. Five wavelengths such as
and statistical modelling helped in detecting the wave- 515 nm, 640 nm, 664 nm, 690 nm, and 724 nm were
length which is crucial i.e.365 nm in finding uninocu- selected from this range. These wavelengths were fur-
lated seeds from inoculated seeds. This is not possible ther used in the 3 ratio functions in a frass-detection
with RGB images that fall under the visible light spec- algorithm [30].
trum. The number of fungal infections increased as the
seeds were kept in contact with the D. Avenae fungi for
a longer time interval. Reflectance features of black oat
4.5 Fruit/Flower detection
seeds varied depending on the absence and presence
of D. Avenae. The reflectance intensity was high for
The study shows the importance of long-wavelength
healthy seeds and decreased for increased inoculation
sensitive receptors for edge detection in bees using mul-
intervals for seeds infested with D.Avenae fungi [26].
tispectral images of flowers. The results show that the
For the detection of pear trees infected with fire blight,
response of the long-wavelength receptor plays a key
an aerial remote sensing technique was used. This is
role due to the following reasons: 1. The response value
done by comparing the changes in wavelength for dis-
is high and the standard deviation value is small over
eased and healthy leaves and estimating spectral VI in
a variety of flower types[45]. 2. The signal strength is
visible and NIR ranges [43]. As a result, it was possi-
steady/consistent and has the highest value of SNR ra-
ble to differentiate HEL, SDL and NSL leaves. In the
tio for every leaf, petal and centre of the flower. Changes
first phase of bacterial infection, there are no visible or
in the output of the photoreceptors and their SNR have
physical marks/signs on the tree. The NIR range can
considerable significance in subsequent visual process-
be used for early detection of the FB disease as in this
ing[46]. So signal strength is high in long-wavelength
phase, the light absorption in the NIR range is affected
for recognizing flower petals and for dividing a visual
by the changes in the internal parts of the leaf. A false-
image into areas. The short and medium wavelength re-
colour composite image is generated that captures the
ceptors have high variability which reveals that flower
healthy and infected leaves of the crown of the pear
colours extend in the colour spectrum of bees. This
tree. They show that there is a variation in the spectral
can signify crucial information about the species and
features of the leaves infected with FB as well as the
it can also highlight important information about the
healthy leaves. The healthy leaves in the FCC image are
nectar reward associated with the flower[47]. Thus, the
represented by the red colour while the infected leaves
long-wavelength channel helps in detecting the object
are represented by the white colour. From the aerial
which is proved by the honeybee’s eye’s resolution and
multispectral imaging of the crown of the pear tree, FB
in edge detection. The medium-wavelength sensitive re-
infected tree can be correctly identified. To differentiate
ceptors have high SNR’s as well, but only for petals. So
between the healthy trees, FB infected trees and trees
medium-wavelength receptors can detect the front ob-
with nutrient deficiency, multispectral imagery (aerial)
ject in honeybees [32]. By combining fluorescence imag-
was captured and processed by an SVM classifier [27].
ing with colour imaging along with image processing al-
For greening detection in citrus with the help of visible
gorithms, foreign materials can be detected in fresh-cut
spectrum imaging, the C-SVM method was used. This
vegetables. The development of a liquid crystal tunable
study focuses on detecting HLB infected citrus from
filter found on spectral imager with customized software
healthy leaves. It also detects other unhealthy leaves
was used for implementing the system with the help
that do not have HLB. The grey and hue colour his-
of the images of flowers[48]. The wavelength range of
togram of LBP was used for texture feature collection.
420–730 nm was used for capturing the fluorescence im-
Compared to other kernel functions, the RBF kernel
ages of the fresh-cut vegetables mixed with FMs. From
function performed best for C-SVC binary classifica-
this range, only 4 wavelengths were found to be best,
tion [44]. The PCA method was examined and other
for fast and real-time recognition of FMs in fresh-cut
methods such as quadratic discriminant analysis (QDA)
vegetables. Along with fluorescence imaging, a colour
and k-Nearest Neighbor algorithm (KNN) were stud-
camera was mounted for the identification of FMs. The
ied and compared. The results found that the C-SVC
set up which includes the spectral imager, camera, and
method is the best in deciding the HLB infected citrus
conveyor belt, was technically in sync for concurrent
as compared to other methods[28]. To detect frass on
capturing of the colour images and multispectral of the
ripened red tomatoes, a multispectral frass detection
specimen when they pass on the conveyor belt. [33].

4.6 Internal fruit parameter which consisted of normalized digital number (DN), re-
flectance and VI derived from various UAV images. The
To evaluate the quality features and maturity stage in best spectral features and window size were taken in the
strawberry fruit, a multispectral imaging system was DT for the NIR, RGB and MS image [37]. The TS-DT
used. In contrast to PLS and SVM models, the BPNN method can accurately separate rice plants from the
model performed well in predicting the firmness and surroundings. The new TS-DT method achieved higher
TSS content. 100% classification accuracy was achieved accuracy in Kappa coefficient value notwithstanding
for accurately identifying the maturity stage of the straw- the various growth phases and UAV images. So, the TS-
berry fruit with the help of the SVM model. The SVM DT method is a strong and practical method for the de-
model achieved higher accuracy than that of using the tection of crops with the help of a UAV image [37]. For
PCA-BPNN model. VIS range of the spectra was used estimation of tomato yield UAS based HTP was used.
for all models [35]. A new HTP framework was used to find tomato yield
using multi-temporal UAS information. Shape features
which include CC and CV were considered. The VI
4.7 Fruit Ripeness/Maturity found from the field match the mathematical graphs.
From the growth and growth rate graphs, the time-
For non-destructive quantification of the maturity pro- series phenotypes were computed. As the phenotypic
cess in banana, 3 methods- Fourier fractal study, Hotelling features are correlated with actual yield separately. Lin-
transform, and homogeneity texture was used. In this ear regression models have higher R2 values greater
study, detecting, recording, and quantifying the matu- than 0.7. The factor study reveals 2 important factors:
rity action of a banana at the 7th phase of the ripening growth speed and timing, which show strong relation
process was done. These methods are non-destructive with the yield performance of tomato varieties. To pre-
techniques related to computing multispectral images. dict the yield, 5 time-series phenotypes were taken. The
The texture study is related to the co-occurrence matrix results show the variance of 65 per cent of the actual
as well as the homogeneity criteria. With the separa- harvest. From the RGB images, the phenotypic features
tion of the brown spots, the texture evaluation gave the were derived to provide adequate information for pre-
most optimal results. In the developed experiment, the dicting the yield. The actual and estimated values of the
results show low variability as the brown spots increase yield are compared, to find the capability of UAS-based
all across the banana peel. A decreased homogeneity variety selection in the breeding scheme. With the help
causes the brown spots over the banana to show an in- of an estimator, the variety that showed high perfor-
crease in uniform differences. Later it covers the entire mance was not properly selected. Whereas the variety
area of the peel, which makes it possible to evaluate and with low performance showed similar results between
record the advancement of the maturity process when the actual yields and the UAS yields. An increase by 53
the brown spots appear on the surface of the banana % for remaining varieties was achieved for the average
peel [36]. tomato yield using UAV data for the variety selection
process [39].
4.8 Fruit Yield
5 Limitation
For evaluating the sugar beet growth indicators UAV,
multispectral images were used. For this purpose, weight 5.1 Varietal Identification
coefficients (α) were added to NDVI to compute 4 WDRVI
indexes. This was done to estimate the LAI, FWL and For varietal identification of rice seeds, PLS-DA and
FWR of the study area. A sensitivity study of 5 in- PCA-BPNN models achieved low accuracies. PLS-DA
dices and 3 growth indicators shows that when the and PCA-BPNN, even when tremendously used in chemo-
NDVI 0.8, the WDRVI index has high sensitivity than metrics, does not give good results as in LS-SVM, at
NDVI. In the sugar beet, seed growth stage interval of least for this case [1].
40 to 60 days, as compared to NDVI, there is varia-
tion in WDRVI. The early growth stage of sugar beet
has a correlation value of 0.01 between growth index 5.2 Fruit Quality
and growth indicator. The correlation value is greater
than 0.858 in WDRVI1. So, WDRVI1 can be used to For the evaluation of the quality of pomegranate fruit,
check beet growth [38].To detect rice plants, many UAV the spectroscopy method had some limitations as com-
images were utilized. The TS-DT method is proposed pared with hyperspectral imaging. The spectroscopy

38 Shilpa Gaikwad
method calculates the total amount of transmitted or the result of microbiological inspection with the multi-
reflected light from a certain part of a sample and does spectral images, the task of delivering the product in
not contain wavelength-specific data [49]. This limita- the market was completed [20].
tion was solved by using the hyperspectral imaging sys-
tem which spatially receives the wavelength responses
(pixel) of images of the fruit [19]. For Jatropha cur-
5.3 Classification
cas seed quality, spectral imaging is used, which is ma-
jorly related to physical, chemical and texture features
In the classification of Medicinal plant leaves, the fea-
related to the insect damages, fungal infections, etc.
ture selection (FS) process is the most crucial part of
The limitation in spectral imaging is to inspect internal
the ML-based classification. The objective of this work
tissues [50]. Whereas, X-ray imaging can gather data
is to identify the vital features and eliminate the extra
from internal composition like injury in embryos and
features which are not necessary for the classification
endosperm [51]. Also, X-rays are small electromagnetic
process. The goal of this research is to achieve better
waves in the range of 0.01 nm to 10 nm, with high
accuracy in less time. It is observed that without fea-
penetrating strength [52] [14]. For Wok-Fried Vegeta-
ture selection, the MLP classifier gives 98.81% accuracy
bles, in earlier work, a feature vector was extracted for
results where the size of ROO is 280 × 280, but it takes
each vegetable piece, consisting of several percentiles
a lot of time (4.83 Seconds) due to a large number of
in the probability distribution functions based on ra-
features. But when selected features are used, higher ac-
tios of pixel intensities in different wavelengths. This
curacy is obtained in less time. Other methods like the
limitation is solved as in the current study, reflection
PCA provides excellent results on the linearly separated
changes in pure wavelengths are examined. T-tests are
data set. The PCA method is an unsupervised approach
performed to evaluate the significance of reflection changes
and does not provide good results on labelled data, as
between days [53]. For each wavelength and each piece
in this work the medicinal plant leaves varieties data
of vegetable, a t-test was conducted to check for sig-
set is labelled [57]. This limitation in PCA is resolved
nificant differences on a 5% level in several percentiles
by ML-based supervised feature selection techniques:
of the light reflectance. The results show significant
chi-square feature evaluator (ranked search) method
changes from days 2 to 4 in the reflectance spectrum for
[58]. This method selects the best features from the
both the celeriac and carrots, significant changes con-
huge features vector space (FVS). Compared to PCA,
tinuing until day 14 were found [15].To evaluate crop
the current method can extract the sub-database with
hail damage in potatoes, the use of aircraft and satel-
the best features for this large database [21]. For tree
lites for aerial imaging is limited. The limitations are
species classification and health status evaluation in the
that the values of spatial and temporal resolution are
forest, major studies contribute to tree species identifi-
low and the cost is high that is unsuitable for fields of
cation and forest health evaluation from multispectral
small sizes. Nowadays, unmanned aerial vehicles (UAV)
images using UAS. These studies focused on rural and
are used as they are highly efficient in collecting data
urban forests instead of managed forests. They are also
in agriculture and phenomics at lower costs [54]. Com-
limited to only spectral study and usually incorporate
pared to remote sensing using satellites, UAV can cap-
RGB sensors. Work on plant nurseries and low vegeta-
ture high spatial resolution (cm) images with persis-
tion which has low tree volume and low levels of de-
tent frequency [55]. This shows that UAV can be em-
sign complexity are useful, but forest conditions create
ployed for the collection of high-resolution images in
problems for detection [59]. Compared to earlier stud-
plots, to examine crop destruction because of hail dam-
ies limitations, a more different and specific method-
age and other causes [16]. To classify the quality of
ological structure is introduced. SVM method was used
vanilla cream, the objective was to develop a model
with a combination of textural and spectral data. Later
which could correctly identify fresh or spoiled samples.
statistical methods for classifying tree species and iden-
A threshold value was set as TVC ≤ 2 to assess the qual-
tification of dead and unhealthy trees harmed by bark
ity of the vanilla cream samples in the dairy industry.
beetles are investigated [22]. To classify the multispec-
When this condition was satisfied, the product is deliv-
tral images in the coral surroundings, the task is to map
ered instantly to the market within a selected time of
underwater coastal conditions from orbital remote sens-
30 days. In earlier methods, finding the threshold limit
ing images. The limitation is the level of the water col-
was a costly, effort-intensive, and lengthy process. Thus,
umn that causes a low signal strength captured by the
multispectral imaging provides a less expensive, fast,
orbital sensor [60]. The various underwater floors usu-
and automatic method that helps in the quick judg-
ally have a single spectral mark, as a result of electro-
ment process for quality managers [56]. By comparing
magnetic energy dissipation and absorption processes

[61]. This results in very close spectral output, gener- of frass on mature red tomatoes, hyperspectral line-scan
ating inter classes that produce errors into the output imaging systems were used, which shows direct use to
in the process of classification. For this study, the hy- automated food processing units. Such systems are fast,
brid classifier is studied and the classification is done operate at high speeds which is a major requirement for
in 3 stages. Finally, combining the several partial clas- commercial food processing lines[64]. They collect im-
sifications generated, the desired accuracy is achieved ages at narrow, accurate (selected) bandwidths of light
[23]. in the wide range of the visible and NIR spectra. [65]
The use of specific bands which are large and inaccurate
will cause less effective separation and identification as
5.4 Disease Detection
more precise details and changes in spectral features
might be lost. Hyperspectral data that has high-quality
Radiometric calibration is a crucial step in UAV for
and full-wavelength bands can be studied to pick wave-
aerial multispectral imagery to detect plants with dis-
lengths important to the object examination task [66].
eases. The reference targets are by default measured
This results in the development of a multispectral in-
and executed in practical accordance methods [62]. But
spection algorithm best suited for fast real-time appli-
the effectiveness and need of this step are not clear.
cation [30].
This limitation is resolved by using UAV-based multi-
spectral images in this study[24]. The rice sheath blight
disease can be identified by using UAS, RGB, and mul-
tispectral imagery data. The purpose was to approxi-
mately and measurably identify the rice sheath blight 5.5 Fruit /Flower Detection
disease in plot areas. The colour features computed us-
ing the multispectral images and the RGB images could Multispectral images of flowers show that long-wavelength
hardly identify the variations in covering caused due sensitive receptors play a crucial part in edge detection
to the disease [63]. Also, the information available is in bees. The earlier work of foraging bees was a visual
not enough to classify various levels of sheath blight, task, to locate, recognize and identify reward producing
as the wavelength range is small and large bands re- flowers [67]. The general nature of plant-pollinator in-
lated to the RGB camera are used. This leads to the teraction is that bee-pollinated flowers show advanced
limitation of a UAV that is equipped with an RGB displays pointed at bees [68]. Also, the bees visual sys-
camera for disease detection at a business level and tem should be able to detect bee-pollinated flowers [69].
shows ShB resistant cultivars in fields for the rice cul- The multispectral images allow us to model the visual
tivation program. [25].For detection of D. Avenae fungi response which the bees photoreceptors give in response
in black oat seeds, currently, the detection is done by to the flower image. This, in turn, will help to differenti-
examining the dry seed visually. This method is hard, ate the function of the different wavelength channels in
slow, and require experts. So fast, reliable, and cor- colour processing. In this work, the estimation was done
rect techniques are required to find seed health status. on the following factors: the variation in the responses
Advanced sensors combined with object recognition by to 3 different receptor types (long, medium and short
computers that can automatically operate will result in wavelength) over a range of kinds of flowers the signal
a fast examination of seed health status by extraction information of each part such as petals of a flower and
of wavelength, texture features and colour [17]. Mul- leaves [32]. For real-time identification of foreign objects
tispectral imaging combines optical spectroscopy and mixed with fresh-cut vegetable processing, fluorescence
computer vision, which results in producing spatial and imaging was applied. The earlier methods: metal detec-
wavelength-specific data on different fungi species [26]. tion, X-ray inspection, and colour imaging techniques
For identification of fire blight infected pear trees, the are not satisfactory. The X-ray technique is not suit-
earlier research did not use remote sensing. For this able for recognizing soft foreign objects [70], whereas
study, aerial as well as ground remote sensing imagery metallic objects can only be detected by metal detec-
was used. This was done to identify the pear trees in- tors [71]. The limitation in Terahertz imaging is that
fected with the FB disease, in the early phase before it the spatial resolution is low and show low results in
could infect other parts. [27].For citrus greening detec- water media [72]. Also, the machine vision technology
tion, earlier research used various machine vision clas- with the colour camera is not useful for visibly opaque
sifiers. This resulted in the high machine cost. This lim- plastic FMs, and cannot detect the FMs that are similar
itation was solved by a low-cost image acquiring setup, in colour to fresh-cut vegetables. The limitations of the
to identify the HLB disease by using a visible spectral earlier work were solved by using fluorescence imaging
image of citrus leaf in this study [28]. For identification in this work [33].

40 Shilpa Gaikwad
5.6 Internal fruit parameter [78]. NDVI is a popularly used index that finds the
biophysical features of canopies [79]. The LAI and leaf
To determine the quality features in the strawberry cluster biomass found from NDVI is used to monitor
fruit, the limitation was that the NIR spectrometers growth, but this is not possible for medium and high
can only identify a small part of the fruit. So the re- bio-masses. A WDRVI developed uses NVDI NIR band
sultant spectrum does not represent the entire fruit at reflectance data, which causes lowering of the weights in
certain times. Hyperspectral imaging combines spec- such bands which belong to high and medium biomass
troscopy with standard imaging to acquires spatial as cases [80]. This results in increased linearity and re-
well as spectral data from an object concurrently [73]. duces saturation, which makes correct growth monitor-
The limitation of hyperspectral imaging is that prob- ing easier. In this work, an enhanced WDRVI is used
lems are faced while processing the data that make it to evaluate the early-phase growth of sugar beet [38].
difficult for commercial real-time applications [74]. This For classification of rice plants detection in the early
problem is solved using multispectral imaging. This tech- stages, UAV was used. This resulted in high resolution
nology is employed in the fast, nondestructive exami- remotely sensed images that were used for crop moni-
nation of the interior and exterior features in different toring economically and practically [81]. The UAV im-
types of vegetables and fruits [35]. ages give more comprehensive data for the study. The
limitation is more complex surrounding like ground, wa-
ter, and shadow of plant, which leads to difficulty in de-
5.7 Fruit Ripeness/Maturity tecting plants in the early season. [82]. If the surround-
ing is not eliminated, the accuracy of crop growth obser-
For estimation of the banana’s ripening process, the vation will reduce. So, correct classification of the rice
techniques involving optical sensing which include spec- field would be useful if the surrounding environment is
troscopy and imaging were used earlier [75]. But the removed and improve the accuracy of rice growth mon-
limitation is to collect sufficient spectral and spatial itoring. [37].
data to evaluate food and agricultural products. How-
For tomato yield estimation, in earlier studies crop
ever, the spectral data cannot be obtained with the
parameters were taken from UAS data, which lead to
help of standard imaging technique [76]. Whereas in
the development of various new methods directly. But
the spectroscopy method huge sample areas cannot be
there was variation in the measurements of the UAS
operated. Recently spectral imaging has appeared as an
based data which was related to data gathering con-
excellent technique for quality examination of various
ditions such as climate, sensor, time, and date [83].
agricultural products. The earlier methods aforemen-
The earlier work used time-based data for the whole
tioned are more complex in their algorithms or experi-
crop season and phenotypes [84]. From the UAS data
mental setups, which are not useful in real-time appli-
of the tomato plots, high phenotypic features can be ex-
cations. Other than this, earlier techniques never incor-
tracted. The growth and growth rate graphs were plot
porated multispectral imaging in their work. So, this
from phenotypic data using multi-temporal UAS data.
work focuses on multispectral imaging which helps in
This was used to extract crop features which include
finding, recording, and quantifying the banana ripening
growth speed and time in the crop season. Finally, a
process [36].
yield evaluation model was developed in the tomato
plot and then experimented to select the high-graded
varieties and reject the low-grade varieties [39].
5.8 Fruit Yield
For observation of sugar beet growth indicators, WDRVI

was used. Earlier studies used satellite platforms that
are confined to both height and orbit and cannot man- 6 Future Work
age the temporal, spatial or spectral resolution essential
for monitoring the growth [77]. The new UAV tech- 6.1 Varietal identification
nology can manage low-cost, crop growth observation
at high temporal, spatial, and spectral resolution. Re- For varietal identification of tomato, further studies will
cently, drones are used in the observation of crop growth. focus on the use of spectral data related to VIS-NIR ob-
The RGB images acquired by the UAV evaluated the tained from multispectral imaging to discriminate the
canopy estimated area of a tree and provided the best rice seeds of different varieties. This information will be
option. In some studies, vegetation indices are generally related to practical organic aspects of tomato and other
employed to estimate crop canopy design and growth crops [13].

6.2 Fruit Quality not operate in that wavelength range. Future work can
be on the application of thermal sensors which will en-
For quality evaluation of pomegranate fruit, future re- hance the classification and health status of different
search will concentrate on the execution of the devel- tree species. The combination of multispectral sensors
oped multispectral imaging by carrying out the real- and UAS can decrease the operating costs of tracking
time test [19]. For discrimination of Jatropha curcas forest areas in the small to medium range [22]. To clas-
seed quality, future work will be an evaluation of seed sify the multispectral images in the coral surrounding,
quality by reducing intrinsic distinction of seed test- future work can be done, for studying the orbital image
ing by using a combination of X-ray and multispectral with the help of an SVM classifier. With this technique,
imaging for fast, feasible, and safe evaluation [14]. For it was possible to classify images of submerged parts,
wok-fried vegetables of carrot and celeriac samples, the as the formation of reefs inflected to noises caused due
sensory quality of the wok-fried celeriac and carrots is to the weather effect and the water level [23].
assessed using the Quality index method. QIM is a sen-
sory method established on important sensory factors,
and it consists of a scoring system for each factor with
scores from 0 to 2 or 3 faulty points [85]. The scores
of all features are then added that results in an overall 6.4 Disease Detection
sensory score. This is called Quality Index. QIM indi-
cates scores of 0 for fresh products and larger scores
For aerial multispectral imagery employed in disease
in increasing order for spoiled products. This plan is
detection for citrus and peach orchard, radiometric cal-
observed by images and complete information of all
ibration shows small changes in the results. Removal of
factors. The sensory study requires that the samples
radiometric calibration will ease UAV-centered disease
are reheated before assessment, as otherwise, it is not
detection and agricultural monitoring on a small scale
possible to assess taste and smell [15]. For hail damage
[88]. For further study, wherein images from varying
evaluation in potatoes, the UAV-based imaging method
dates were taken or big areas are used for classification
provides the best solution. Further studies of UAV can
the accuracy has to be improved using radiometric cal-
be performed in other applications, like yield prediction
ibration [24]. For detection of rice sheath blight, using
in agriculture [16]. For vanilla cream, microbiological
a UAV can help in the cultivation of rice cultivars with
quality classification the Unsupervised Online feature
resistance to Sheath Blight disease. For future work, a
Selection (UOS) algorithm could manage a classifica-
new UAV system developed can be used for site-specific
tion error that was quite low in the test database. The
accurate fungicide application technique to control ShB
threshold limit which was set for the “fresh” samples
disease in rice [25]. For the detection of D. Avenae in
of vanilla cream, was a crucial factor for product deliv-
black oat seeds, multispectral imaging can be used in
ery to the market by the manufacturer. This example
future for separating seed-carrying fungi, affected by
gives a good outlook for the application of multispec-
other elements, but are firmly related to physical and
tral imaging in combination with the UOS algorithm in
chemical differences caused by fungi [26]. For the use
the quality management of the dairy industry [20].
of aerial remote sensing technique in the identification
of FB infected pear trees, the future work can be the
6.3 Classification ability of multispectral both aerial and ground, remote
sensing imagery to identify FB disease of other trees,
For the classification of medicinal plant leaves, this study usually in the early phase. This will help in preventing
is limited to six medicinal plant leaves while there are the spread of the FB disease so that necessary action
millions of types of medicinal plant/herbs in the world. can be taken [27]. For citrus greening detection, further
This is a pixel-based method and in the future, an study can be, collecting a full image data set of HLB
object-based method can be used [86]. In future, this positive and negative leaves of various types taken at
proposed method can be used on other medicinal plant different times. In this work, classification was done in
leaves. Further, the results can be improved using hy- two classes. In future work, the multi-class case can be
perspectral and 3D digital image data sets [21]. For explored to find every symptom in citrus [28]. For iden-
tree species classification and health status assessment, tification of frass on mature red tomatoes, further work
identifying physical stress in earlier stages or finding can be done on tomato samples at different maturity
small wavelength changes can be done with high relia- phases and to check the method in real-time. Future
bility and accuracy by using hyperspectral sensors [87]. work can focus on a high level of frass dilutions and
This was not possible in multispectral sensors as it does infection on different areas on the fruit [30].

42 Shilpa Gaikwad
6.5 Fruit/Flower Detection for tomato yield estimation, the results of this method
can be helpful in cultivation schemes to chose high-yield
For multispectral images of flowers, receptors with long and disease-resistant types for tomato fields. In the fu-
wavelengths for edge detection in bees was used. The ture, the Artificial Intelligence model can be utilized
results provide acceptable information for selecting the for yield estimation and variety identification in agri-
long-wavelength channel over the shorter-wavelength cultural products [39].
channels for clarifying a visual scene. This can initiate
the ways environment and prior knowledge can build
more resulting neural processes which can be the future
work [32]. For real-time detection of foreign material
mixed with fresh-cut vegetables, two types of fresh-cut
vegetables were tested. For future work, this method
can be experimented on different vegetables by pick-
ing the best fluorescence wavelengths(band) and using 7 Conclusion
a suitable algorithm for processing images of fresh-cut
vegetables [33].
In this review paper, we have focused on Multispec-
tral imaging which is the emerging technology for the
6.6 Internal Fruit Parameter grading of fruits and vegetables. Quality assessment of
vegetables and fruits is the need of the hour for the
To find quality features and ripeness phase in straw- food industry. To meet consumer demand and profit,
berry fruit, PLS, SVM, and BPNN models were em- the technologies developed should provide more con-
ployed. All the models’ used visible areas of the spec- cise, fast and accurate results. Some of these technolo-
trum. But, if the NIR spectra are also incorporated then gies have their merits and demerits. This paper has
the results could be marginally improved. The multi- reviewed different ways in which multispectral images
spectral imaging system can reduce the image retrieval could be used for fruit and vegetable assessment like
as well as operating time in contrast to the hyperspec- varietal identification which includes identification of
tral imaging system that allows online automated qual- rice seeds and tomato, fruit quality which includes wok-
ity observation systems [35]. fried vegetables(carrots and celeriac) , hail damage in
potato, quality of vanilla cream, etc, classification which
includes medicinal plant leaves, tree species and coral
6.7 Fruit Ripeness/Maturity
reefs, disease detection which includes plant disease de-
tection in citrus and peach orchard, rice sheath blight
For the ripening process in bananas, future work can be
disease, pathogen in black oat seeds, HLB disease in cit-
online applications to measure maturity level in other
rus, frass on tomatoes, fruit or flower detection which
fruits [36].
includes detection of foreign material in cabbage, to
find how the honey bees and bumble bees respond to
6.8 Fruit Yield the multispectral images of flowers for long,short and
medium wavelength receptors ,internal fruit parameters
For observation of sugar beet growth indicators, the fu- which includes to find firmness and TSS in strawberry,
ture work can be for growth tracking of potato, radish, sugar content in potatoes,Fruit ripeness which includes
and other underground crops to a large extent for reap- ripeness in banana, and Fruit yield which includes sugar
ing benefits [38]. For detection of rice plants in the beet growth,rice plants and tomato yield. The data ob-
early phase using NIR-G-B, RGB and multispectral im- tained by multispectral imaging can promote the eval-
ages from UAV, the current method can be improved uation of the classification of fruits and vegetables. The
by merging more data taken from multiple UAV im- study of multispectral imaging for fruit and vegetable
ages and working on complex conditions. High spatial will be beneficial to disease prevention, irrigation, and
resolution satellite images are easier to capture [89]. yield improvement. Furthermore, multispectral infor-
The TS-DT method, in the future, can work on vari- mation can be considered the basis for evaluating var-
ous factors in agriculture with satellite images, like rice ious characteristics and parameters of fruit, vegetable
planting evaluation, crop plant area survey. So, TS-DT and tree species and promoting planting research on
is a good technique for the detection of crops using fruit, vegetable and trees species. There are also certain
UAV images and have future prospects in satellite im- challenges like high cost, accuracy and time factors that
ages [37]. For Unmanned Aircraft System-based HTP are to be considered.

Declarations References
Funding information 1. W. Liu, C. Liu, F. Ma, X. Lu, J. Yang, and L. Zheng,

“Online variety discrimination of rice seeds using multi-
spectral imaging and chemometric methods,” Journal of
No funding was involved in the present work. Applied Spectroscopy, vol. 82, no. 6, pp. 993–999, 2016.
2. T. Bourlai and B. Cukic, “Multi-spectral face recogni-
tion: Identification of people in difficult environments,”
Conflicts of interest in 2012 IEEE International Conference on Intelligence
and Security Informatics. IEEE, 2012, pp. 196–201.
3. M. Moroni, A. Mei, A. Leonardi, E. Lupo, and F. L.
Authors S. Gaikwad declare that there has been no con-
Marca, “Pet and pvc separation with hyperspectral im-
flict of interest. agery,” Sensors, vol. 15, no. 1, pp. 2205–2227, 2015.
4. S. Aggarwal, “Principles of remote sensing,” Satellite re-
mote sensing and GIS applications in agricultural mete-
Code availability orology, vol. 23, pp. 23–28, 2004.
5. H.-W. Chen, J.-H. Lee, B.-Y. Lin, S. Chen, and S.-T. Wu,
“Liquid crystal display and organic light-emitting diode
Not applicable. display: present status and future perspectives,” Light:
Science & Applications, vol. 7, no. 3, pp. 17 168–17 168,
2018.
Authors’ contributions 6. J. D. Vincent, S. Hodges, J. Vampola, M. Stegall, and
G. Pierce, Fundamentals of Infrared and Visible Detector
Conceptualization was done by Shilpa Gaikwad (SG). Operation and Testing. John Wiley and Sons, 2015.
7. G. A. Shaw and H. K. Burke, “Spectral imaging for re-
All the literature reading and data gathering were per- mote sensing,” Lincoln laboratory journal, vol. 14, no. 1,
formed by SG. The formal analysis was performed by pp. 3–28, 2003.
SG. Manuscript writing- original draft preparation was 8. A. Goldberg, B. Stann, and N. Gupta, “Multispectral,
done by SG. Review and editing was done by SG. Vi- hyperspectral, and three-dimensional imaging research at
the us army research laboratory,” ARMY RESEARCH
sualization work was carried out by SG. LAB ADELPHI MD, Tech. Rep., 2003.
9. I. Makki, R. Younes, C. Francis, T. Bianchi, and M. Zuc-
chetti, “A survey of landmine detection using hyperspec-
Ethics approval tral imaging,” ISPRS Journal of Photogrammetry and
Remote Sensing, vol. 124, pp. 40–53, 2017.
10. S. Baronti, A. Casini, F. Lotti, and S. Porcinai, “Mul-
All authors consciously assure that the manuscript ful-
tispectral imaging system for the mapping of pigments
fills the following statements: 1) This material is the in works of art by use of principal-component analysis,”
authors’ own original work, which has not been previ- Applied optics, vol. 37, no. 8, pp. 1299–1309, 1998.
ously published elsewhere. 2) The paper is not currently 11. D. Sivakumar, Y. Jiang, and E. M. Yahia, “Maintaining
mango (mangifera indica l.) fruit quality during the ex-
being considered for publication elsewhere. 3) The pa- port chain,” Food Research International, vol. 44, no. 5,
per reflects the authors’own research and analysis in a pp. 1254–1263, 2011.
truthful and complete manner. 4) The paper properly 12. B. Zhang, W. Huang, J. Li, C. Zhao, S. Fan, J. Wu,
credits the meaningful contributions of co-authors and and C. Liu, “Principles, developments and applications
of computer vision for external quality inspection of fruits
co-researchers. 5) The results are appropriately placed and vegetables: A review,” Food Research International,
in the context of prior and existing research. vol. 62, pp. 326–343, 2014.
13. S. Shrestha, L. C. Deleuran, M. H. Olesen, and R. Gis-
lum, “Use of multispectral imaging in varietal identifica-
Consent to participate tion of tomato,” Sensors, vol. 15, no. 2, pp. 4496–4512,
2015.
14. V. d. J. M. Bianchini, G. M. Mascarin, L. C. A. S. Silva,
This article does not contain any studies with animals V. Arthur, J. M. Carstensen, B. Boelt, and C. B. da Silva,
or humans performed by any of the authors. Informed “Multispectral and x-ray images for characterization of
consent was not required as there were no human par- jatropha curcas l. seed quality,” Plant Methods, vol. 17,
ticipants. All the necessary permissions were obtained no. 1, pp. 1–13, 2021.
15. L. H. Clemmensen, B. S. Dissing, G. Hyldig, and H. Løje,
from Institute Ethical committee and concerned au- “Multispectral imaging of wok-fried vegetables,” Journal
thorities. of Imaging Science and Technology, vol. 56, no. 2, pp.
20 404–1, 2012.
16. J. Zhou, M. J. Pavek, S. C. Shelton, Z. J. Holden, and
Consent for publication S. Sankaran, “Aerial multispectral imaging for crop hail
damage assessment in potato,” Computers and Electron-
ics in Agriculture, vol. 127, pp. 406–412, 2016.
Authors have taken all the necessary consents for pub- 17. S. Cubero, N. Aleixos, E. Moltó, J. Gómez-Sanchis, and
lication from participants wherever required. J. Blasco, “Advances in machine vision applications for

44 Shilpa Gaikwad
automatic inspection and quality evaluation of fruits and imagery,” Postharvest biology and technology, vol. 76, pp.
vegetables,” Food and bioprocess technology, vol. 4, no. 4, 40–49, 2013.
pp. 487–504, 2011. 32. V. Vasas, D. Hanley, P. G. Kevan, and L. Chittka, “Multi-
18. N. Kondo, “Automation on fruit and vegetable grading spectral images of flowers reveal the adaptive significance
system and food traceability,” Trends in Food Science & of using long-wavelength-sensitive receptors for edge de-
Technology, vol. 21, no. 3, pp. 145–152, 2010. tection in bees,” Journal of Comparative Physiology A,
19. R. Khodabakhshian, B. Emadi, M. Khojastehpour, M. R. vol. 203, no. 4, pp. 301–311, 2017.
Golzarian, and A. Sazgarnia, “Development of a multi- 33. S. Lohumi, B.-K. Cho, and S. Hong, “Lctf-based multi-
spectral imaging system for online quality assessment of spectral fluorescence imaging: System development and
pomegranate fruit,” International Journal of Food Prop- potential for real-time foreign object detection in fresh-
erties, vol. 20, no. 1, pp. 107–118, 2017. cut vegetable processing,” Computers and Electronics in
20. A. Lianou, A. Mencattini, A. Catini, C. Di Natale, G.- Agriculture, vol. 180, p. 105912, 2021.
J. E. Nychas, E. Martinelli, and E. Z. Panagou, “Online 34. A. G. Dyer, J. Spaethe, and S. Prack, “Comparative psy-
feature selection for robust classification of the microbi- chophysics of bumblebee and honeybee colour discrim-
ological quality of traditional vanilla cream by means of ination and object detection,” Journal of Comparative
multispectral imaging,” Sensors, vol. 19, no. 19, p. 4071, Physiology A, vol. 194, no. 7, p. 617, 2008.
2019. 35. C. Liu, W. Liu, X. Lu, F. Ma, W. Chen, J. Yang, and
21. S. Naeem, A. Ali, C. Chesneau, M. H. Tahir, F. Jamal, L. Zheng, “Application of multispectral imaging to deter-
R. A. K. Sherwani, and M. Ul Hassan, “The classifica- mine quality attributes and ripeness stage in strawberry
tion of medicinal plant leaves based on multispectral and fruit,” PloS one, vol. 9, no. 2, p. e87818, 2014.
texture feature using machine learning approach,” Agron- 36. M. Santoyo-Mora, A. Sancen-Plaza, A. Espinosa-
omy, vol. 11, no. 2, p. 263, 2021. Calderon, A. I. Barranco-Gutierrez, and J. Prado-
22. A. Abdollahnejad and D. Panagiotidis, “Tree species Olivarez, “Nondestructive quantification of the ripening
classification and health status assessment for a mixed process in banana (musa aab simmonds) using multispec-
broadleaf-conifer forest with uas multispectral imaging,” tral imaging,” Journal of Sensors, vol. 2019, 2019.
Remote Sensing, vol. 12, no. 22, p. 3722, 2020. 37. H. Zheng, X. Zhou, J. He, X. Yao, T. Cheng, Y. Zhu,
23. A. P. Henriques, A. D. D. Neto, and R. F. Amaral, “Clas- W. Cao, and Y. Tian, “Early season detection of rice
sification of multispectral images in coral environments plants using rgb, nir-gb and multispectral images from
using a hybrid of classifier ensembles,” Neurocomputing, unmanned aerial vehicle (uav),” Computers and Elec-
vol. 73, no. 7-9, pp. 1256–1264, 2010. tronics in Agriculture, vol. 169, p. 105223, 2020.
24. H. Pourazar, F. Samadzadegan, and F. Dadrass Javan, 38. Y. Cao, G. L. Li, Y. K. Luo, Q. Pan, and S. Y. Zhang,
“Aerial multispectral imagery for plant disease detection: “Monitoring of sugar beet growth indicators using wide-
radiometric calibration necessity assessment,” European dynamic-range vegetation index (wdrvi) derived from
Journal of Remote Sensing, vol. 52, no. sup3, pp. 17–31, uav multispectral images,” Computers and Electronics
2019. in Agriculture, vol. 171, p. 105331, 2020.
25. D. Zhang, X. Zhou, J. Zhang, Y. Lan, C. Xu, and 39. A. Chang, J. Jung, J. Yeom, M. M. Maeda, J. A. Lan-
D. Liang, “Detection of rice sheath blight using an un- divar, J. M. Enciso, C. A. Avila, and J. R. Anciso, “Un-
manned aerial system with high-resolution color and mul- manned aircraft system-(uas-) based high-throughput
tispectral imaging,” PloS one, vol. 13, no. 5, p. e0187470, phenotyping (htp) for tomato yield estimation,” Journal
2018. of Sensors, vol. 2021, 2021.
26. F. França-Silva, C. H. Q. Rego, F. G. Gomes-Junior, 40. J. Camo, A. Lorés, D. Djenane, J. A. Beltrán, and P. Ron-
M. H. D. d. Moraes, A. D. d. Medeiros, and C. B. d. calés, “Display life of beef packaged with an antioxidant
Silva, “Detection of drechslera avenae (eidam) sharif active film as a function of the concentration of oregano
[helminthosporium avenae (eidam)] in black oat seeds extract,” Meat science, vol. 88, no. 1, pp. 174–178, 2011.
(avena strigosa schreb) using multispectral imaging,” 41. Ö. Kayadibi, “Evaluation of imaging spectroscopy and
Sensors, vol. 20, no. 12, p. 3343, 2020. atmospheric correction of multispectral images (aster and
27. N. Bagheri, “Application of aerial remote sensing tech- landsat 7 etm+),” in Proceedings of 5th International
nology for detection of fire blight infected pear trees,” Conference on Recent Advances in Space Technologies-
Computers and Electronics in Agriculture, vol. 168, p. RAST2011. IEEE, 2011, pp. 154–159.
105147, 2020. 42. A. Matese and S. F. Di Gennaro, “Practical applications
28. X. Deng, Y. Lan, T. Hong, and J. Chen, “Citrus green- of a multisensor uav platform based on multispectral,
ing detection using visible spectrum imaging and c-svc,” thermal and rgb high resolution images in precision viti-
Computers and Electronics in Agriculture, vol. 130, pp. culture,” Agriculture, vol. 8, no. 7, p. 116, 2018.
177–183, 2016. 43. R. Devadas, D. Lamb, S. Simpfendorfer, and D. Back-
29. A. Engelbrecht, “Particle swarm optimization: iteration house, “Evaluating ten spectral vegetation indices for
strategies revisited,” in 2013 BRICS Congress on Com- identifying rust infection in individual wheat leaves,”
putational Intelligence and 11th Brazilian Congress on Precision Agriculture, vol. 10, no. 6, pp. 459–470, 2009.
Computational Intelligence. IEEE, 2013, pp. 119–123. 44. Y. Bao, T. Wang, and G. Qiu, “Research on applicabil-
30. C.-C. Yang, M. S. Kim, P. Millner, K. Chao, B.-K. Cho, ity of svm kernel functions used in binary classification,”
C. Mo, H. Lee, and D. E. Chan, “Development of multi- in Proceedings of International Conference on Computer
spectral imaging algorithm for detection of frass on ma- Science and Information Technology. Springer, 2014,
ture red tomatoes,” Postharvest Biology and Technology, pp. 833–844.
vol. 93, pp. 1–8, 2014. 45. A. Fitter, R. Fitter, I. Harris, and M. Williamson, “Re-
31. B.-K. Cho, M. S. Kim, I.-S. Baek, D.-Y. Kim, W.-H. Lee, lationships between first flowering date and temperature
J. Kim, H. Bae, and Y.-S. Kim, “Detection of cuticle de- in the flora of a locality in central england,” Functional
fects on cherry tomatoes using hyperspectral fluorescence Ecology, pp. 55–60, 1995.

46. J. C. Anderson and S. B. Laughlin, “Photoreceptor per- radar satellite data for underwater topography estima-
formance and the co-ordination of achromatic and chro- tion in coastal areas,” Ocean Dynamics, vol. 61, no. 12,
matic inputs in the fly visual system,” Vision research, pp. 2099–2120, 2011.
vol. 40, no. 1, pp. 13–31, 2000. 62. K. Linnet, “Performance of deming regression analysis
47. E. Meléndez-Ackerman, D. R. Campbell, and N. M. in case of misspecified analytical error ratio in method
Waser, “Hummingbird behavior and mechanisms of se- comparison studies,” Clinical chemistry, vol. 44, no. 5,
lection on flower color in ipomopsis,” Ecology, vol. 78, pp. 1024–1031, 1998.
no. 8, pp. 2532–2541, 1997. 63. S. Thomas, M. T. Kuska, D. Bohnenkamp, A. Brug-
48. W. Wang, C. Li, E. W. Tollner, G. C. Rains, and R. D. ger, E. Alisaac, M. Wahabzada, J. Behmann, and A.-
Gitaitis, “A liquid crystal tunable filter based shortwave K. Mahlein, “Benefits of hyperspectral imaging for plant
infrared spectral imaging system: Calibration and char- disease detection and plant protection: a technical per-
acterization,” Computers and Electronics in Agriculture, spective,” Journal of Plant Diseases and Protection, vol.
vol. 80, pp. 135–144, 2012. 125, no. 1, pp. 5–20, 2018.
49. A. Ghita, P. Matousek, and N. Stone, “High sensitivity 64. K. Chao, C.-C. Yang, Y. Chen, M. Kim, and D. Chan,
non-invasive detection of calcifications deep inside biolog- “Hyperspectral-multispectral line-scan imaging system
ical tissue using transmission raman spectroscopy,” Jour- for automated poultry carcass inspection applications for
nal of biophotonics, vol. 11, no. 1, p. e201600260, 2018. food safety,” Poultry science, vol. 86, no. 11, pp. 2450–
50. T. Arnold, M. De Biasio, and R. Leitner, “High- 2460, 2007.
sensitivity hyper-spectral video endoscopy system for 65. J. Qin, M. S. Kim, K. Chao, D. E. Chan, S. R. Delwiche,
intra-surgical tissue classification,” in SENSORS, 2010 and B.-K. Cho, “Line-scan hyperspectral imaging tech-
IEEE. IEEE, 2010, pp. 2612–2615. niques for food safety and quality applications,” Applied
51. B. Gagliardi and J. Marcos-Filho, “Relationship between Sciences, vol. 7, no. 2, p. 125, 2017.
germination and bell pepper seed structure assessed by 66. S. Mahesh, D. Jayas, J. Paliwal, and N. White, “Hyper-
the x-ray test,” Scientia Agricola, vol. 68, no. 4, pp. 411– spectral imaging to classify and monitor quality of agri-
416, 2011. cultural materials,” Journal of Stored Products Research,
52. L. Weinstein, “Electromagnetic waves,” Radio i svyaz’, vol. 61, pp. 17–26, 2015.
Moscow, 1988. 67. L. Chittka and N. E. Raine, “Recognition of flowers by
pollinators,” Current opinion in plant biology, vol. 9,
53. P. R. Mills, S. C. Tomkins, and L. J. Schlangen, “The
no. 4, pp. 428–435, 2006.
effect of high correlated colour temperature office lighting
68. S. J. Hegland, A. Nielsen, A. Lázaro, A.-L. Bjerknes, and
on employee wellbeing and work performance,” Journal
Ø. Totland, “How does climate warming affect plant-
of circadian rhythms, vol. 5, no. 1, pp. 1–9, 2007.
pollinator interactions?” Ecology letters, vol. 12, no. 2,
54. Y. Shi, J. A. Thomasson, S. C. Murray, N. A. Pugh, W. L.
pp. 184–195, 2009.
Rooney, S. Shafian, N. Rajan, G. Rouze, C. L. Morgan,
69. Z. Bukovac, A. Dorin, V. Finke, M. Shrestha, J. Garcia,
H. L. Neely et al., “Unmanned aerial vehicles for high-
A. Avargues-Weber, M. Burd, J. Schramme, and A. Dyer,
throughput phenotyping and agronomic research,” PloS
“Assessing the ecological significance of bee visual detec-
one, vol. 11, no. 7, p. e0159781, 2016.
tion and colour discrimination on the evolution of flower
55. J.-C. Padró, F.-J. Muñoz, J. Planas, and X. Pons, “Com- colours,” Evolutionary Ecology, vol. 31, no. 2, pp. 153–
parison of four uav georeferencing methods for environ- 172, 2017.
mental monitoring purposes focusing on the combined 70. R. P. Haff and N. Toyofuku, “X-ray detection of defects
use with airborne and satellite remote sensing platforms,” and contaminants in the food industry,” Sensing and In-
International journal of applied earth observation and strumentation for Food Quality and Safety, vol. 2, no. 4,
geoinformation, vol. 75, pp. 130–140, 2019. pp. 262–273, 2008.
56. G. ElMasry, N. Mandour, S. Al-Rejaie, E. Belin, and 71. A. H. Trang, P. V. Czipott, and D. A. Waldron, “Char-
D. Rousseau, “Recent applications of multispectral imag- acterization of small metallic objects and nonmetallic an-
ing in seed phenotyping and quality monitoring—an tipersonnel mines,” in Detection and Remediation Tech-
overview,” Sensors, vol. 19, no. 5, p. 1090, 2019. nologies for Mines and Minelike Targets II, vol. 3079.
57. R. Gopalan, R. Li, and R. Chellappa, “Domain adapta- International Society for Optics and Photonics, 1997, pp.
tion for object recognition: An unsupervised approach,” 372–383.
in 2011 international conference on computer vision. 72. D. Mittleman, M. Gupta, R. Neelamani, R. Baraniuk,
IEEE, 2011, pp. 999–1006. J. Rudd, and M. Koch, “Recent advances in terahertz
58. S. Picek, A. Heuser, A. Jovic, and L. Batina, “A sys- imaging,” Applied Physics B, vol. 68, no. 6, pp. 1085–
tematic evaluation of profiling through focused feature 1094, 1999.
selection,” IEEE Transactions on Very Large Scale Inte- 73. R. Lu and Y.-R. Chen, “Hyperspectral imaging for
gration (VLSI) Systems, vol. 27, no. 12, pp. 2802–2815, safety inspection of food and agricultural products,” in
2019. Pathogen Detection and Remediation for Safe Eating,
59. S. E. Sesnie, P. E. Gessler, B. Finegan, and S. Thessler, vol. 3544. International Society for Optics and Pho-
“Integrating landsat tm and srtm-dem derived variables tonics, 1999, pp. 121–133.
with decision trees for habitat classification and change 74. T. Adão, J. Hruška, L. Pádua, J. Bessa, E. Peres,
detection in complex neotropical environments,” Remote R. Morais, and J. J. Sousa, “Hyperspectral imaging: A
Sensing of Environment, vol. 112, no. 5, pp. 2145–2159, review on uav-based sensors, data processing and appli-
2008. cations for agriculture and forestry,” Remote Sensing,
60. M. L. Zoffoli, R. Frouin, and M. Kampel, “Water col- vol. 9, no. 11, p. 1110, 2017.
umn correction for coral reef studies by remote sensing,” 75. P. Rajkumar, N. Wang, G. EImasry, G. Raghavan, and
Sensors, vol. 14, no. 9, pp. 16 881–16 931, 2014. Y. Gariepy, “Studies on banana fruit quality and matu-
61. A. Pleskachevsky, S. Lehner, T. Heege, and C. Mott, rity stages using hyperspectral imaging,” Journal of food
“Synergy and fusion of optical and synthetic aperture engineering, vol. 108, no. 1, pp. 194–200, 2012.

46 Shilpa Gaikwad
76. Z. Takats, J. M. Wiseman, and R. G. Cooks, “Ambi-

ent mass spectrometry using desorption electrospray ion-
ization (desi): instrumentation, mechanisms and appli-
cations in forensics, chemistry, and biology,” Journal of
mass spectrometry, vol. 40, no. 10, pp. 1261–1275, 2005.
77. W. H. Maes and K. Steppe, “Perspectives for remote
sensing with unmanned aerial vehicles in precision agri-
culture,” Trends in plant science, vol. 24, no. 2, pp. 152–
164, 2019.
78. D. Haboudane, J. R. Miller, E. Pattey, P. J. Zarco-
Tejada, and I. B. Strachan, “Hyperspectral vegetation
indices and novel algorithms for predicting green lai of
crop canopies: Modeling and validation in the context of
precision agriculture,” Remote sensing of environment,
vol. 90, no. 3, pp. 337–352, 2004.
79. A. A. Gitelson, “Wide dynamic range vegetation index
for remote quantification of biophysical characteristics of
vegetation,” Journal of plant physiology, vol. 161, no. 2,
pp. 165–173, 2004.
80. L. Wang, Q. Chang, F. Li, L. Yan, Y. Huang, Q. Wang,
and L. Luo, “Effects of growth stage development on
paddy rice leaf area index prediction models,” Remote
Sensing, vol. 11, no. 3, p. 361, 2019.
81. J. Bendig, M. Willkomm, N. Tilly, M. Gnyp, S. Bennertz,
C. Qiang, Y. Miao, V. Lenz-Wiedemann, and G. Bareth,
“Very high resolution crop surface models (csms) from
uav-based stereo images for rice growth monitoring in
northeast china,” Int. Arch. Photogramm. Remote Sens.
Spat. Inf. Sci, vol. 40, pp. 45–50, 2013.
82. S. L. Ozesmi and M. E. Bauer, “Satellite remote sensing
of wetlands,” Wetlands ecology and management, vol. 10,
no. 5, pp. 381–402, 2002.
83. V. Döpper, T. Gränzig, B. Kleinschmit, and M. Förster,
“Challenges in uas-based tir imagery processing: image
alignment and uncertainty quantification,” Remote Sens-
ing, vol. 12, no. 10, p. 1552, 2020.
84. S. Varela, T. Pederson, C. J. Bernacchi, and A. D. Leakey,
“Understanding growth dynamics and yield prediction of
sorghum using high temporal resolution uav imagery time
series and machine learning,” Remote Sensing, vol. 13,
no. 9, p. 1763, 2021.
85. E. Costell, “A comparison of sensory methods in quality
control,” Food quality and Preference, vol. 13, no. 6, pp.
341–353, 2002.
86. P. Aplin and G. Smith, “Advances in object-based image
classification,” The International Archives of the Pho-
togrammetry, Remote Sensing and Spatial Information
Sciences, vol. 37, no. B7, pp. 725–728, 2008.
87. A. Lowe, N. Harrison, and A. P. French, “Hyperspectral
image analysis techniques for the detection and classifica-
tion of the early onset of plant disease and stress,” Plant
methods, vol. 13, no. 1, pp. 1–12, 2017.
88. N. Delavarpour, C. Koparan, J. Nowatzki, S. Bajwa, and
X. Sun, “A technical study on uav characteristics for pre-
cision agriculture applications and associated practical
challenges,” Remote Sensing, vol. 13, no. 6, p. 1204, 2021.
89. Q. Yu, P. Gong, N. Clinton, G. Biging, M. Kelly, and
D. Schirokauer, “Object-based detailed vegetation clas-
sification with airborne high spatial resolution remote
sensing imagery,” Photogrammetric Engineering & Re-
mote Sensing, vol. 72, no. 7, pp. 799–811, 2006.

SSRN Id3905180

Uploaded by

Copyright:

Available Formats

SSRN Id3905180

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id3905180

Uploaded by

Copyright:

Available Formats

Noname manuscript No.

(will be inserted by the editor)

Literature Review on Multi-spectral imaging for fruits and

Received: date / Accepted: date

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Zheng et al. [16] used RGB, NIR-G-B and multi-

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

3 Results 245 spectra of cream samples were considered. 18 dif-

Electronic copy available at: https://ssrn.com/abstract=3905180

Table 1 Varietal Identification and Fruit Quality

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Table 3 Disease Detection

(DT) i.e. TS-DT are OA for RGB is 91.25%, OA for 4 Discussion

Electronic copy available at: https://ssrn.com/abstract=3905180

Table 5 Fruit Yield

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

4.3 Classification spectral images in coral surroundings. A hybrid SVM

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

For observation of sugar beet growth indicators, WDRVI

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Funding information 1. W. Liu, C. Liu, F. Ma, X. Lu, J. Yang, and L. Zheng,

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

Electronic copy available at: https://ssrn.com/abstract=3905180

76. Z. Takats, J. M. Wiseman, and R. G. Cooks, “Ambi-