The Pharma Innovation Journal 2021; SP-10(10): 761-767

ISSN (E): 2277- 7695

ISSN (P): 2349-8242
NAAS Rating: 5.23 Predicting yield attributes of maize through image
TPI 2021; SP-10(10): 761-767
© 2021 TPI processing
www.thepharmajournal.com
Received: 16-08-2021
Accepted: 18-09-2021 Sandhya P, Patil SG, Radha M, Djanaguiraman M, Dheebakaran GA and
Sandhya P Gunasekaran
Research Scholar, Department of
Physical Sciences & Information
Abstract
Technology, AEC & RI, TNAU,
Currently, the yield attributes of maize ears (such as ear length, kernel count, kernel weight, and so on)
Coimbatore, Tamil Nadu, India
are often assessed by hand throughout the breeding process, which necessitates a large number of
Patil SG workers. Furthermore, subjective mistakes are difficult to prevent, and manual measuring efficiency is
Assistant Professor, Department quite poor. The method described in this work can quickly assess yield attributes linked to breeding traits
of Physical Sciences & of numerous maize ears, significantly improving maize variety evaluation efficiency. From photographs
Information Technology, AEC & of ears taken from field trial plots, a low-cost ear digital imaging system was developed that offers
RI, TNAU, Coimbatore, estimations of ear and kernel characteristics such as ear number and size, kernel number and size, and
Tamil Nadu, India kernel weight. Image J, an open-source program, is used here to process the images using a script that
runs in batch mode. The total kernel number was determined from the number of visible kernels on the
Radha M picture and the average kernel size was used to calculate kernel weight. Ground truth measurements and
Assistant Professor, Department data obtained by image processing have an excellent agreement in terms of accuracy and precision. The
of Physical Sciences & procedure also entails utilizing a mobile camera to picture scattered kernel samples and counting them
Information Technology, AEC &
using the software. Results demonstrate that this is a fast (less than a minute per sample) and reliable
RI, TNAU, Coimbatore,
approach that may be extensively used for estimating yield attribute and kernel counting.
Tamil Nadu, India

Djanaguiraman M Keywords: maize, yield attributes, image-processing, estimation, correlation

Assistant Professor, Department
of Crop Physiology, AC&RI, 1. Introduction
TNAU, Coimbatore, Prediction of crop yield in agriculture is a paramount snag as the yield essentially hangs on
Tamil Nadu, India
weather conditions, pesticides, and many more. Decision-making concomitant to agricultural
Dheebakaran GA risk management and future conjecture needs unerring counsel concerning crop yield’s history
Assistant Professor, Agro (Hajir Almahdi, 2020). To meet the requirements of enormous population growth there is a
Climatic Research Centre, requirement for field-level statistics as it is essential to plan at micro-level and crop insurance
TNAU, Coimbatore, exceptionally (Anup, 2005). Accurate and timely assessment of crop yield is an essential
Tamil Nadu, India
process to ensure the adequacy of the food supply. In the past, estimates of crop yield were
Gunasekaran done, in general, from the expertise of farmers or, as claimed by Geipel, Link and Claupein,
Associate Professor, Department (2014) “better estimations can be drawn from destructive sampling procedures in
of Mathematics, PSG College of representative areas”.
Arts & Science, Coimbatore, For yield estimation, basic mathematical or statistical relationships based on agronomic and
Tamil Nadu, India
meteorological data were established (Dadhwal & Ray, 2000), as well as crop development
models (Thorp, De Jonge, Kaleita, Batchelor, and Paz, 2008); Later remote sensing techniques
evolved as an effective tool for assessing and monitoring crop yield at a lower cost than other
methods providing a spatial and periodic, comprehensive view of real crop state (Geipel et al.,
2014). But these methods are expensive, time-consuming, and laboriousand require trained
personnel and expertise. So it is mandatory to identify crop growth monitoring and prior
estimation of yield using unbiased, systematized, and quicker methods like image processing
than going for traditional methods.
Maize or corn (Zea mays L.) is an important annual cereal crop of the world belonging to the
family Poaceae. It is the third leading crop of the world after rice and wheat (Sandhu, Singh, &
Malhi, 2007). The world production of maize was 967 million metric tons (MMT) and in
India, its production was 23 MMT in 2013–14 (India maize summit, 2014). It is considered a
Corresponding Author staple food in many parts of the world. The measurement of kernel traits is important for maize
Sandhya P breeding and the evaluation of its yield (Lovemore Chipindu, 2020) [2].
Research Scholar, Department of Ears are a primary agricultural product of maize, which has led the majority of previous
Physical Sciences & Information
phenotyping efforts to focus on aspects of yield, such as ear size, kernel row number, and
Technology, AEC & RI, TNAU,
Coimbatore, Tamil Nadu, India kernel structure and dimensions.
~ 761 ~
The Pharma Innovation Journal http://www.thepharmajournal.com

Additional information about the crop might be obtained and were then imaged separately under the same setup as cobs and
farmers could make projected kernel weight could be saved in a single folder by their cob number which is then
quantified fast and precisely. Recently, image processing, batch processed.
machine learning, and deep learning have shown great
potential in progressing the digital capabilities needed for the 2.2 Image Processing
future of agriculture. These techniques have shown to be Image analysis was conducted in Fiji [an extension of ImageJ
reliable in high-throughput phenotyping and in enabling software-which has a collection of plugins that make
farmers to make a real-time decision, something that was scientific picture analysis easier]. It is open-source software
previously not possible. (Saeed Khaki, 2020) [1]. that will help us to extract yield component parameters (i.e.
Few method sallow the extraction of ear and kernel features ear and kernel attributes) using a series of steps as per
through image processing.In the perspective of breeding, Makanza R. et al. (2018) [16].
research has indicated that yield components have a higher The following steps were performed using built-in ImageJ
heritability than overall yield. [23, 24]; allowing for independent plugins and the process was recorded as a macro to run
selection for these qualities and then combining the subsequent images of cob. The images were pre-processed at
responsible genetic loci to generate a genotype with higher first to obtain an even dark background using the background
performance or developing a selection index through trait subtraction method with a threshold of 100. Then for
combinations [24]. According to Miller et al. [6], more could be segmenting individual kernels, CLAHE (Contrast Limited
learned about the genetic underpinnings of yield components Adaptive Histogram Equalization) process is selected which
and how to improve those utilizing current and future maize helps in overcoming the noise problem by reducing noise by
genetic resources if its ear and kernel traits could be setting the parameters of the CLAHE plugin as (I) Block Size
automatically evaluated with higher objectivity and precision. to 29 (II) No. of histogram bins to 250 and (III) Maximum
Using ear digital imaging, this study presents a simple, high- slope to 5.
throughput, and robust approach for collecting yield After this using the edge enhancement process the images
components (ear and kernel properties) from harvested maize were sharpened with an unsharp mask plugin with a radius of
ears (EDI). 15 and mask as 0.70 followed by conversion of images to 8-
bit format. The images of low contrast were dealt with an auto
2. Materials and Methodology local threshold method called Phansalkar whose parameters 1
Maize Hybrid Co (MH) – 6 was chosen for the research. The & 2 were kept to default values as an increase in their default
experiment was carried out at the Eastern block (75) in Tamil values proved to be of no effect while the radius of the local
Nadu Agricultural University, Coimbatore [11.0087° N, domain was set to 15 and the option of a white object on
76.9404° E and altitude 300m]. For analysis 100 plants were black background was selected to threshold the image.
randomly selected from the field and the data on cob The images were then binarized using convert to binary
attributes and their related plant characters were recorded for plugin and fill holes option was selected to avoid kernel split
each selected plant using the standard method of during adjustable watershed step followed later with a
measurements. tolerance of 3. After the successful segmentation of kernels,
the measurement of yield attributes was initialized.
2.1 Photo acquisition The threshold is calculated using the equation
Harvested ears of randomly selected plants were collected
from the field, numbered accordingly, and were dried upto s(x, y)
12-18% moisture content. First, the weight of cobs with and T(x, y) = m(x, y)[1 + pe−qm(x,y) ] + k ( ) − 1 (1)
R
without sheath was recorded using Infra Digi precision
balance with a resolution of 0.01gm. Then manual Where, m = mean,
measurements of sheath length, sheath weight, cob length, p & q = Phansalkar’s exponential constants,
cob width, and number of rows per ear, number of kernels per k = constant in the range of 0.2-0.5,
row, kernel length, kernel width, and kernel weight were s = standard deviation of pixel intensities,
taken. R = dynamic range of standard deviation which is 0.5 for
First, the images were acquired by placing each selected cob normalized images.
one by one on a black chart using a mobile camera (16 MP
rear camera) mounted on an Arm Stand Holder at a height of 2.3 Kernel counts and attributes
60 cm from the mobile camera to the table at which maize is The segmented images were then analyzed using analyze
placed and positioned at nadir. The process of photo particles plugin after setting its size to 0.03-1.00 pixels2 and
acquisition under diffused lighting conditions was setup in our the circularity to 0.05-1.00 which helps in identifying the
computer lab. At the same height as cobs, an image of a ruler kernels visible on the image. Then the kernel length and width
was taken to convert pixel values to known measurements. of the cob image were measured as the distance between two
The images were numbered according to cob number and points along the major and minor axis of ear and kernel which
saved in a single folder as raw RGB images without any requires setting the scale to a known distance at first. The
conversion for batch processing. To validate kernel counting, particle analysis plugin also helps us in measuring the total
based on image versus manually counted kernels the kernels kernel area, average kernel area, and average perimeter of the
of selected ears were shelled using Mini hand corn thresher, ear in addition to kernel count, kernel weight, and total kernel
winnowed, kept separately in plastic bags, and were number.
numbered respectively to the ears. The number of kernels per
cob was counted manually and the kernel weight of each cob 2.4 Ear count and attributes
was measured separately using Electronic Compact Digital The analysis can be done not only on single cob images but
Kitchen SF-400A Weighing Scale. The kernels of each cob we can also analyze several cobs on a single image. To find
~ 762 ~
The Pharma Innovation Journal http://www.thepharmajournal.com

the number of cobs in an image, the images were filtered out Where kn = number of visible kernels on the image.
via a Gaussian blur filter with the sigma set to 10 for The kernel weight model was created using a linear regression
smoothing. Then they were binarized followed by the fill model between average kernel length (𝑘𝑙 ̅ ) and average kernel
holes plugin and watershed process and analyze particles weight (total kernel weight divided by the total number of
plugin was selected to count the number of cobs on a kernels), both of which were physically measured using a
particular image. Ear length and width were measured as the digital balance with a precision of 0.01 g. Kernel weight was
distance between two points along the major and minor axis recorded at moisture content levels ranging from 11 to 13%.
of a single ear. This was done with 100 ears of different kernel sizes. The
visible region of the segmented ear was used to calculate the
2.5 Modeling of Kernel weight and Kernel Count average kernel length. To build a model that correlates kernel
To count kernels via image the shelled kernels of each cob ̅ ) was plotted against the average
length into kernel weight, (𝑘𝑙
were spread to an area of 60 cm wide and 40 cm long on a measured kernel weight for each ear measured manually (Fig.
black chart and images was taken as the same setup as cobs 6a). The model was then tested, and it appeared to be correct
and the following steps were carried out through ImageJ built- in estimating the kernel weight (Fig. 6b).
in plugins to count kernels and were recorded as a macro for
subsequent analysis. The color threshold of the image was ̅ ∗ 0.7435 – 0.155) (3)
Average Kernel Weight (g) = (𝑘𝑙
first adjusted, then binarized, and was processed using the
watershed plugin to split individual kernels from others. Then 2.6 Estimation of Kernel Weight
to get kernel count, analyze particles plugin is selected and Given that Eq. 2 gives the total kernel number and Eq. 3 gives
the size and circularity of objects are adjusted to get an the average kernel weight, the total kernel weight (Eq. 4) is
absolute count. calculated by multiplying the two equations:
From the number of visible kernels on the image (kn) (Eq. 2,
r=0.98***), a linear regression model for predicting the total ̅ ∗ 0.7435 – 0.155) (4)
Total Kernel Weight (g) = (2.4051 * kn – 6.7334) * (𝑘𝑙
kernel number on each ear was constructed. The association
between estimated and measured kernel parameters was The predicted total kernel weight was confirmed using 100
evaluated using Pearson's correlation coefficient r. cob images taken under controlled lighting conditions.

Total Kernel Number = 2.4051 * kn – 6.7334 (2)

Graph 1: Steps involved in cob image processing and mining of data

~ 763 ~
The Pharma Innovation Journal http://www.thepharmajournal.com

Graph 1: Steps associated with image processing of kernel

Table 1: Correlation between Actual and Estimated maize yield attributes

Variables Correlation Coefficient (ρ) R² Regression Equation
Image-based Kernel Count Vs Manual Kernel Count 0.9906 0.9815 y = 0.9794x + 0.0019
Estimated Ear Length (cm) Vs Measured Ear length (cm) 0.9553 0.9127 y = 0.9698x + 1.0451
Estimated Ear Width (cm) Vs Measured Ear Width (cm) 0.9535 0.9092 y = 0.8989x + 0.3615
Kernel Weight (g) Vs Kernel Length (cm) 0.9997 0.9995 y = 1.4194x + 0.2359
Estimated Kernel Weight (g) Vs Measured Kernel Weight(g) 0.9906 0.9815 y = 0.9794x + 0.0019

Fig 1: Regression modeling and Correlation between image-based Kernel Count and manual Kernel Count

Fig 2: Regression modeling and Correlation between measured and estimated Ear Length
~ 764 ~
The Pharma Innovation Journal http://www.thepharmajournal.com

Fig 3: Regression modeling and Correlation between measured and estimated Ear Width

Fig 4: Regression model for predicting Kernel Weight from Kernel Length

Fig 5: Regression modeling and Correlation between measured and estimated Kernel Weight

3. Result & Discussion data. The identical ears that were utilized for kernel count
3.1 Kernel Count and Ear attributes validation were used to compare manual ear length and width
As indicated in the methodology section, the kernel count measurements to those generated by the image processing
model was evaluated using 100 ears collected from randomly method. For both attributes, the analysis shows a linear
selected plants in the field. The calculated kernel count from correlation of (r > 0.95, p0.001) between the two procedures.
entire ears using the model and the actual count of detached
kernels exhibited a linear connection (ρ=0.98, p<0.001) in the

~ 765 ~
The Pharma Innovation Journal http://www.thepharmajournal.com

3.2 Kernel Weight Estimation conditions in the field, the logistics of using it in the field (i.e.
Data were obtained from 100 ears (as detailed in the the need for a computer) and the limited number of ears (3–5)
procedure) to validate the kernel weight estimate approach. that can be scanned at one time make it unsuitable for
The weight of the measured kernels was then compared to the evaluating thousands of ears in a breeding trial.
weight of the estimated kernels. The results show that, on The EDI technique also predicts kernel weight based on
average, the calculated kernel weight and the measured kernel kernel size, allowing for a low-cost yield performance
weight are in relatively excellent agreement, with a evaluation, especially in cases where ear shelling and kernel
correlation of r = 0.99. weighing are too expensive or when the necessary equipment
The number of harvestable kernels with their weight can be is unavailable. It's worth noting that this technique doesn't
used to predict maize grain yield. Kernel number, out of such account for kernel moisture (as the kernel weight model was
two yield factors, typically explains the most variation and is designed for a range of kernel moisture between 11 and 13
highly connected to ear size. Kernel qualities, despite their percent), which can have a big impact on the final weight if
importance, are difficult to assess quickly and correctly, it's not taken into account. Furthermore, the EDI technique
mainly because of the necessity for ear threshing before they does not account for kernel depth when estimating weight,
can be measured. The number of rows in one length of the ear which may result in a minor underestimate of the real kernel
can be counted manually and multiplied by the number of weight in some situations.
kernels in one length of the ear to get the total kernel count of
cobs. These manual yield component measurements have 4. Conclusion
shown to be effective and were utilized in a divergent The EDI technique has been demonstrated to be a viable
selection study of the connection between ear length and alternative to standard ear phenotyping methods in this study.
yield, for example. Hand measures, which generally use calipers and manual
The issue with these approaches is the lack of consistency counting, are more reliable, especially for large numbers of
inherent in the way data is gathered (which is dependent on ears that are routinely assessed in breeding experiments. The
the training and appreciation of the people committed to that accuracy of this approach is primarily dependent on the
work), as well as the time and related expense, which makes camera's resolution; however, this is no longer a serious
them most appropriate for very small trials. According to a problem due to recent considerable improvements in the
preliminary evaluation, depending on the intended resolution of all camera types, including smartphone and
measurement, the suggested EDI technique can be twice tablet cameras. As the results show a good correlation with
(example: ear count) to five-fold (example: ear dimensions) the images taken using a mobile camera. The technique will
quicker than manual methods. Manual techniques are labor be especially beneficial to breeding operations with limited
demanding, making them more expensive than the EDI operating resources. The ability to combine ear and kernel
approach. Because of differences in labor costs, the cost characteristics might aid in the development of cultivars with
difference would vary depending on the location/country. desired farmer qualities such as ear or kernel size.
Automated measures that are more reliable, quick, and low-
cost might be used in yield component research and crop 5. References
improvement choices. 1. Khaki S, Pham H, Han Y, Kuhl A, Kent W, Wang L.
Miller et al. [6] presented a kernel counting imaging technique Convolutional neural networks for image-based corn
based on individual kernel areas. The technique assesses kernel detection and counting. Sensors 2020;20(9):2721.
kernel size (width and depth) on separated kernels alone. 2. Chipindu L, Mupangwa W, Mtsilizah J, Nyagumbo I,
While this procedure is quite accurate, it necessitates the Zaman-Allah M. Maize kernel abortion recognition and
removal of the kernels from the ears, which may be classification using binary classification machine learning
inconvenient when dealing with a large number of ears. algorithms and deep convolutional neural networks.
Similarly, Liang et al. [4] have developed a technique for AI 2020;1(3):361-375.
scoring maize kernel characteristics based on line-scan 3. Sayago S, Bocco M. Crop yield estimation using satellite
imaging, which is not appropriate for field evaluation due to images: comparison of linear and non-linear models. Agri
time and cost constraints. Scientia 2018;35(1):1-9.
The suggested EDI technique has the benefit of generating ear 4. Liang X, Wang K, Huang C, Zhang X, Yan J, Yang W. A
and kernel characteristics data from intact ear pictures. Grift high-throughput maize kernel traits scorer based on line-
et al. [19] created a machine vision-based method to count scan imaging. Measurement 2016;90:453-460.
maize kernels on the ear inside a quasi-cylindrical mid-section 5. Zhang X, Liu J, Song H. Corn ear test using SIFT-based
and ear maps, which is comparable to our methodology. panoramic photography and machine vision technology.
While their approach is intriguing to some extent, the imaging Artificial Intelligence in Agriculture 2020;4:162-171.
is performed in a soft box with a light reflector and a high- 6. Miller ND, Haase NJ, Lee J, Kaeppler SM, De Leon N,
quality diffused lighting environment. The throughput of this Spalding EP. A robust, high‐throughput method for
sort of imaging setup is its constraint. In the case of ear size, computing maize ear, cob, and kernel attributes
the EDI technique revealed excellent agreement between automatically from images. The Plant Journal
humanly measured ear measurements and automated image 2017;89(1):169-178.
processing findings (Fig. 8). Miller et al. [21] observed similar 7. Warman C, Sullivan CM, Preece J, Buchanan ME,
findings. The primary difference between the two systems is Vejlupkova Z, Jaiswal P et al. A cost‐effective maize ear
that the one described by Miller et al. [21] acquires ear pictures phenotyping platform enables rapid categorization and
using flatbed document scanners, whereas the EDI method quantification of kernels. The Plant Journal
employs an RGB image acquired by the camera in our work, 2021;106(2):566-579.
it is the mobile camera. Furthermore, while the flatbed 8. Warman C, Fowler JE. Custom-built scanner and simple
scanner has the advantage of being able to control lighting image processing pipeline enables low-cost, high-

~ 766 ~
The Pharma Innovation Journal http://www.thepharmajournal.com

throughput phenotyping of maize ears. Bio Rxiv 2019,

780650.
9. Liang X, Ye J, Li X, Tang Z, Zhang X, Li W et al. A
high-throughput and low-cost maize ear traits
scorer. Molecular Breeding 2021;41(2):1-16.
10. Sayago S, Bocco M. Crop yield estimation using satellite
images: comparison of linear and non-linear models. Agri
Scientia 2018;35(1):1-9.
11. Komyshev E, Genaev M, Afonnikov D. Evaluation of the
SeedCounter, a mobile application for grain
phenotyping. Frontiers in plant science 2017;7:1990.
12. Tolley SA, Singh A, Tuinstra MR. Heterotic Patterns of
Temperate and Tropical Maize by Ear
Photometry. Frontiers in Plant Science 2021;12:1117.
13. Gage JL, Miller ND, Spalding EP, Kaeppler SM, De
Leon N. TIPS: a system for automated image-based
phenotyping of maize tassels. Plant Methods
2017;13(1):1-12.
14. Shyu CR, Green JM, Lun DP, Kazic T, Schaeffer M, Coe
E. Image Analysis for Mapping Immeasurable
Phenotypes in Maize [Life Sciences]. IEEE Signal
Processing Magazine 2007;24(3):115-118.
15. Makanza R, Zaman-Allah M, Cairns JE, Eyre J,
Burgueño J, Pacheco Á et al. High-throughput method
for ear phenotyping and kernel weight estimation in
maize using ear digital imaging. Plant Methods
2018;14(1):1-13.
16. Ma Q, Jiang J, Zhu D, Li S, Mei S, Lü Y. Rapid
measurement for 3D geometric features of maize ear
based on image processing. Transactions of the Chinese
Society of Agricultural Engineering 2012;28(1):208-212.
17. Wu D, Cai Z, Han J, Qin H. Automatic kernel counting
on maize ear using RGB images. Plant Methods
2020;16:1-15.
18. Grift TE, Zhao W, Momin MA, Zhang Y, Bohn MO.
Semi-automated, machine vision-based maize kernel
counting on the ear. Bio systems engineering
2017;164:171-180.
19. Severini AD, Borrás L, Cirilo AG. Counting maize
kernels through digital image analysis. Crop Science
2011;51(6):2796-2800.
20. Miller ND, Haase NJ, Lee J, Kaeppler SM, De Leon N,
Spalding EP. A robust, high‐throughput method for
computing maize ear, cob, and kernel attributes
automatically from images. The Plant Journal
2017;89(1):169-178.
21. Hausmann NJ, Abadie TE, Cooper M, Lafitte HR,
Schussler JR. U.S. Patent No. 8,073,235. Washington,
DC: U.S. Patent and Trademark Office 2011.
22. Messmer R, Fracheboud Y, Bänziger M, Vargas M,
Stamp P, Ribaut JM. Drought stress and tropical maize:
QTL-by-environment interactions and stability of QTLs
across environments for yield components and secondary
traits. Theoretical and Applied Genetics
2009;119(5):913-930.
23. Peng B, Li Y, Wang Y, Liu C, Liu Z, Tan W et al. QTL
analysis for yield components and kernel-related traits in
maize across multi-environments. Theoretical and applied
genetics 2011;122(7):1305-1320.
24. Robinson HF, Comstock RE. Genotypic and phenotypic
correlations in corn and their implications in selection.
North Carolina State University. Dept. of Statistics 1950.

~ 767 ~