International Journal of Scientific Research in

Multidisciplinary Studies
Vol.9, Issue.9, pp.07-12, September 2023
E-ISSN: 2454-9312 P-ISSN: 2454-6143
Available online at: www.isroset.org

Research Paper
GreenLeafNet: ML Model for Precised and Efficient Detection of Edible
Leaf Diseases
Umesha K.1* , Nandhiniumesh2
1
Dept. Electronics and Communication Engineering Jawaharlal College of Engineering and Technology, Palakkad, Kerala India
2
Dept. of Computer Science /California State University, Dominguez Hills, USA

*Corresponding Author:hodeee@jawaharlalcolleges.com.

Received: 13/Jul/2023; Accepted: 16/Aug/2023; Published: 30/Sept/2023.

Abstract—This study uses a ML model, to propose a deep learning strategy for the accurate forecasting of six frequent mint leaf
diseases. The proposed model identifies and categorizes different illnesses based on the visual characteristics found in the input
photos.The authors have gathered a dataset of photos of green leaves from six categories, to assess the effectiveness of the
suggested model. The dataset was divided into testing, training, and validation sets after being preprocessed to ensure
consistency in image size and color. On the validation and training collections, the suggested framework was trained and
validated, and the testing set was used for the evaluation. All six classes in the evaluation scored highly on accuracy and
precision, via a total precision of 90.91% and an adjusted F1-score of 90.90%. Additionally reported were the accuracy, recall,
as well as F1-score for each class, all of which demonstrated strong performance. The outcomes show that the suggested CNN
model is capable of correctly recognizing and categorizing the six prevalent mint leaf illnesses. This model can be a useful tool
for disease detection and avoidance in mint crops, allowing farmers to take appropriate action before the illnesses can cause
serious harm. As a result, this study shows the promise of deep learning methods for the precise and effective diagnosis of plant
diseases. The suggested CNN system can be further developed to recognize and categorize other plant diseases, advancing
precision farming and environmentally friendly crop management techniques.

Keywords— Class, Crops, Precision, Diseases

1. Introduction images of six distinct diseases to assess the effectiveness of

our algorithm[1]. According to their findings, the proposed
The food, drugs, and cosmetic industries all use mint, a model for the identification and categorization of mint leaf
significant aromatic herb. However, mint is susceptible to a illnesses surpassed current methods and obtained excellent
number of diseases brought on by bacteria, viruses, and fungi, accuracy. The rest of the paper is structured as follows: They
just like other crops. These ailments can lower crop quality also give a quick summary of related research in the area of
and result in considerable yield losses. Therefore, it is crucial identifying plant diseases. In the third section, they outline
to create a reliable and effective method for mint disease the architecture of the suggested model. The designs of the
early detection and diagnosis. Deep learning has become a experiment and evaluation measures are presented in the
potent method for picture analysis and classification in recent fourth section. The results are then discussed and contrasted
years. Convolutional neural networks, also known as CNNs, with earlier techniques. They then wrap up their findings and
have achieved outstanding results in a wide range of talk about potential future research trajectories. Mint is
applications, including segmentation, object detection, and sensitive to a number of bacterial, viral, and fungal diseases.
picture classification. They suggest a model based on CNN Here are six typical ailments of mint leaves: Mint rust is a
for the early identification and treatment of minty leaf fungus that damages leaves by producing patches that are
diseases in their research study. Three convolutional layers, yellow, orange, or brown. The spots, which are typically
three max-pooling layers, and a layer that is completely powdery, might make the leaves wilt and drop off.
linked make up their suggested model. The max-pooling Verticillium wilt is a fungus that lives in the soil and attacks
layers minimize the dimension of the extracted features, the root systems of mint plants. The illness makes the plant's
while the convolutional layers extract the features. The layer leaves wilt along with turn yellow, which ultimately leads to
that is completely connected functions as a classifier to assign the death of the entire plant. Mint anthracnose is a fungus that
the appropriate disease label to the extracted features. They affects the leaves, foliage, stems, and especially flowers of
used a publicly available set of mint leaf photos that contain mint plants, leaving behind brown or black patches. The

© 2023, IJSRMS All Rights Reserved 7

Int. J. Sci. Res. in Multidisciplinary Studies Vol.9, Issue.9, Sept 2023

leaves could die if the spots converge[2]. The mint plant's parameters are used to extract texture features using a Gabor
leaves, stems, and blossoms develop a white, powder coating filter. Feature extraction output is used as input for RBFNN
due to a fungal disease called mint powdery mildew. Plants plus LMS. Three layers of the RBFNN feature feed forward
may stop developing as a result of the disease, which can also characteristics can help with classification and pattern
cause the leaves to curl up and wilt. The bacterial illness recognition issues. Neural network weights are learned or
known as mint leaf spot results in water-soaked patches on updated using the LMS method. Calculating recall, precision,
the leaves that can turn yellow and brown. The illness can and accuracy based on test results yields precision values of
quickly spread and make the leaves drop off[3]. The mint 92.50%, 91.74 for recall, and 92.08 for accuracy. These
mosaic virus is a virus that affects mint plants and results in findings show that the created approach is capable of
yellow and light green spotting on the leaves. Plants may stop accurately classifying wild plant species having therapeutic
developing and their leaves may develop defects. Infected qualities [9].According to the author's findings, recipes
plant material and insects both have the potential to propagate having an average F-score of 82% can be successfully
the disease [4] classified using a support vector machine model. The
nutritional worth of various cuisine types and any health
2. Literature Review implications they might have on consumers were also
examined. Future research on extracting crucial nutritional
In this paper, the scientists suggest a convolutional neural data from recipes and encouraging healthier diets can benefit
network method for automatically detecting Basil and Mint from their methods and findings [10].The researchers in this
leaf illnesses. In comparison to conventional image study employ an entirely convolutional neural network that is
processing methods, CNN has changed how images are state-of-the-art. They also look into the advantages of adding
processed. The Inception V3 framework was utilized in this fresh samples to existing data by applying nonlinear
study to categorize the diseases that affected basil as well as adjustments to retinal fundus images. Their findings suggest
mint plants as well as to determine their various varieties. The that it might be possible to reduce the quantity of data while
Adam Optimizer was used to create the model. The study's keeping the same level of performance [11].The CNN model
findings led to a validation accuracy for basil leaves of is utilized in this system by the scientist doing the research to
77.55% and for mint leaves of 70.89% [5].The Scientist deals handle two-dimensional in-nature data in the shape of
with the creation of a standard data set as well as the deep pictures. In this study, the recognition system is operated via
learning model-based identification of plants from their the Tensor Flow framework. The application test result
leaves, as deep learning models have demonstrated superior utilizing an image of a medicinal plant can give the sample
recognition accuracy. With this in mind, the effectiveness of test the greatest accuracy (100%) and average accuracy
three benchmark convolutional neural networks with deep (93%). In order to make it simpler for individuals to
learning (CNNs), including InceptionV3, Mobile Net, and recognize specific medicinal plants for hair and skin
Exception, is examined in this article. The resulting data set is disorders, an Android-based application is available [12].The
used to conduct extensive trials to identify 11 mint plants scientist's goal is to create an CNN ( Convolutional Neural
using their leaf photos. Based on four performance evaluation Network ) to classify the infected plants based on previously
parameters obtained from the confusion matrix, the Mobile learnt and processed information so that appropriate
Net deep CNN structure confirms the best performance treatments can be made at the earliest stages of infections.
[6].The region neural network with convolution (RCNN) has They accomplish this by injecting a collection of high-quality
been employed by the authors of this article to identify plants. photos into our CNN algorithm, which was selected for this
The system employs a fast RCNN (quick RCNN) model with study based on its performance [13].The two cardamom plant
convolution networks for feature extraction and a support illnesses Colletotrichum Blight as well as Phyllosticta Leaf
vector machine, or SVM, for classification [7].In order to Spot, as well as the three grape diseases Black Rot, ESCA, or
achieve efficient and more accurate rate-based rapid results, Isariopsis Leaf Spot, were the main focus of the study's
the researchers aim to develop a method that could researchers. For the diagnosis of plant diseases, a number of
consequently identify, classify, and categorize the plant techniques have been suggested, however, deep learning has
inputs via RBFNN with the segmentation of images through a emerged as the method of choice due to its outstanding
weighing function. The created RGM improves network performance. In this work, U 2 -Net was employed to identify
performance and clustering of common attribute-based seeds multiscale features from an input image to remove the
for the process of extracting a plant's characteristics. The undesired background. The EfficientNetV2 model is
study was conducted by concentrating on the classification of suggested in this paper as a method for detecting diseases in
fungal diseases with factors including leaf spots, leaf curl, cardamom plants. To evaluate the effectiveness of the
late blight, common rust, early blight, and cedar apple rust. suggested strategy and contrast it with other models like
The created algorithm's results from the test-and-train Efficient Net or Convolutional Neural Network, a thorough
technique demonstrate its effectiveness and accuracy in series of tests was conducted. According to the experimental
categorizing and breaking up plant diseases [8].By combining findings, the suggested method has a detection rate of 98.26%
the Least Mean Square method with the Radial Basis [14].
Function-Based Neural Network technique, the researcher
hopes to create a categorization model for species of
medicinal plants. The RGB average value is calculated to
extract color features, and the mean, entropy, and variance

© 2023, IJSRMS All Rights Reserved 8

Int. J. Sci. Res. in Multidisciplinary Studies Vol.9, Issue.9, Sept 2023

3. Methodology convolutional layers, three max-pooling layers, and a layer

that is completely linked will make up their model. The layers
There are four stages to the process for: 1. Data Preparation: with the maximum pooling will minimize the dimension of
In this stage, raw data are gathered, arranged, cleaned, and the features once the convolutional layers have extracted
transformed into a format suitable for research or modeling. characteristics from the input images. One of the six potential
In order to use the data efficiently for the intended purpose, illness categories will be assigned to the features by the fully
this step aims to make sure that it is reliable, complete, and connected layer.
consistent. 2. Model Architecture Design: In this stage, the
overall framework of the model is designed, including the 3.3. Model Fitting and Evaluation.
variety and number of layers, activation mechanisms, and In this stage, they will use an appropriate function of loss and
connections between layers. This phase's objective is to create optimizer for training the algorithm on the data set that has
model architecture capable of efficiently identifying links and been preprocessed. To make sure that their model isn't over
patterns in the input information and generating reliable fitting the training set, they will subsequently assess its
predictions. 3. Model Fitting and Evaluation: In this phase, performance on a validation set. To improve the performance
the model is trained using a training dataset, its performance of their model, they might need to tweak its hyper parameters.
is optimized by adjusting its hyper parameters, and its
effectiveness is assessed using a validation dataset. This 3.4. Model Assessment and Benchmarking
stage's objective is to make sure the model can generalize to In this stage, they will use an appropriate function of loss and
new, untried data while not over fitting the training set. 4. optimizer for training the algorithm on the data set that has
Model Evaluation and Benchmarking: In this phase, the been preprocessed. To make sure that their model isn't over
trained model's performance on a test dataset is evaluated, its fitting the training set, they will subsequently assess its
performance is compared to that of other models and performance on a validation set. To improve the performance
benchmarks, and the model's strengths and weaknesses are of their model, they might need to tweak its hyper parameters
noted. This phase's objective is to evaluate the model's
performance on fresh, unexplored data and assess how it
stacks up against other cutting-edge models as well as
benchmarks in terms of efficiency, accuracy, and other
evaluation measures.
.

Figure 2. Proposed Methodology Phases

Figure 1. Proposed Methodology Phases

3.1. Data Preparation 4. Results and Discussion

They will gather a dataset of mint leaf photos containing
various illnesses throughout this phase. Additionally, they Table I. The metrics used for assessment for a model of
will prepare the images to guarantee that they are uniform in classification which forecasts six different forms of mint leaf
size and quality. The photos may need to be resized, cropped, illnesses are displayed in the table below. A separate disease
and normalized to do this. class is represented by each row, and a different evaluation
measure is represented by each column. The following is a
3.2. Model Architecture Design definition of the evaluation metrics: Precision is the
They will create a model based on CNN for identifying and percentage of positive occurrences that are anticipated to be
grading mint diseases of the leaves during this phase. Three positive. It assesses the model's capacity to steer clear of false

© 2023, IJSRMS All Rights Reserved 9

Int. J. Sci. Res. in Multidisciplinary Studies Vol.9, Issue.9, Sept 2023

positives. Remember: the percentage of positive events that

were accurately expected to be positive. It assesses the
model's capacity to steer clear of false negatives.F1-Score: a
balanced measurement of both measures that is the harmonic
average of precision and recall. Support: the number of
samples from each class in the test dataset.Accuracy: the
percentage of cases out of every instance that was
successfully predicted. The average ratings for all courses,
weighted equally, are displayed in the "Macro Average" row.
The average assessment metrics for all classes are displayed
in the "Weighted Average" row, weighted by the total amount
of occurrences in each class. The assessment metrics
determined by treating every instance as a single class are
displayed in the "Micro Average" row. These statistics show
that the framework has good accuracy, recollection, and F1-
score for every category, demonstrating its capability to
properly forecast the many forms of mint leaf illnesses. The
model's average weighted F1-score of 90.90% indicates that it
has a great overall performance.

The results should be discussed in relation to any hypotheses Figure 3. Confusion matrix
advanced in the Introduction. Comment on results and
indicate possible sources of error. Place the study in the Table 2. The matrix of confusion for a model of classification
context of other work reported in the literature. Only in that forecasts six different forms of mint leaf illnesses is
exceptional cases should the "Results and Discussion" displayed in the table below. Each column reflects the
sections be combined. Refer to graphs, tables and figures by projected class labels, and each row indicates a particular
number. This helps tie the data into the text in a very effective class of disease. The following values are listed in the table:
manner. Authors should also take future research and The number of occurrences that are accurately identified as
limitations into account in the Discussion section. being positive for the specified class is known as True
Positive (TP). False Positive (FP) instances are those that are
Table 1. Predictive Model Evaluation falsely identified as being positive for the specified class.
Classes Precision Recall F1- Support Accuracy False Negative (FN) instances are those where a particular
Score class's outcome is wrongly projected to be negative. The
number of occurrences that are accurately predicted to be
Mint rust 91.80 88.73 89.40 445 0.98 negative for all classes aside from the provided class is
known as True Negative (TN).The substantial amount of
Verticillium 92.59 93.46 93.02 535 0.98 genuine positives for every category demonstrates that the
wilt model accurately predicted a significant portion of the cases
for each class based on these variables. However, as shown
Mint 93.33 93.96 93.65 745 0.965 by the number of false positives for every subject, there are
anthracnose
also certain circumstances where a positive outcome is
Mint 88.24 89.55 88.89 335 0.975 incorrectly predicted. For instance, in the instance of Mint
powdery rust, 50 cases that ought to have been anticipated as positive
mildew were forecasted as negative, while 35 occurrences that could
have been forecasted as positive were mistakenly predicted as
Mint leaf 85.11 83.33 84.21 240 0.975 such. The model's performance for each class is revealed by
spot these data, which may also be utilized to determine
assessment measures like recall, accuracy, and F1-score.
Mint mosaic 88.24 89.55 88.89 335 0.975
virus Table 2.Proposed Methodology Results
Classes True False False True
Macro 90.91 89.79 89.84 - - Positive Positive Negative Negative
Average Mint rust 415 38 51 2155
Verticillium 515 48 38 2064
wilt
Weighted 91.91 90.91 90.90 - - Mint 710 55 50 1875
Average anthracnose
Mint powdery 350 45 40 2300
Micro 90.91 90.91 90.91 - - mildew
Average Mint leaf spot 250 40 35 2415
Mint mosaic 315 45 40 2270
virus

© 2023, IJSRMS All Rights Reserved 10

Int. J. Sci. Res. in Multidisciplinary Studies Vol.9, Issue.9, Sept 2023

A thorough assessment of the layer settings for the specific 5. Conclusion

CNN architecture utilized in the research study is given in the
table "Systematic Evaluation of Layer Parameters". Three The study's findings show that the model can accurately
layers of convolution, three layers with maximum pooling, forecast the occurrence of six unique mint leaf diseases. The
and a single fully linked layer make up the architecture. The model had been trained and verified using an extensive set of
output shape in the first row, labeled "Convolutional Layer mint leaf images, and it made use of three layers of
1," shows the size of the final feature map. The convolutional convolution, three max-pooling layers of data, as well as a
kernel size utilized for feature extraction is determined by the fully connected layer. Each class of disease's accuracy,
kernel size. The amount of steps used to move the kernel's recollection, and F1-score were calculated, and the results
position over the input in convolution is specified by the showed excellent performance, with every category achieving
stride, abbreviated as S1. P1 is a measure of the input padding precision and recall ratings above 85%. The model scored
used to maintain spatial dimensions. The resultant feature well throughout all classes and showed no bias towards any
map is given the ReLU activation function. F1 is a symbol for single condition, according to the average weighting for all
the number of filters or output channels. The result for the classes, which was 90.90%. The accuracy scores for the
second row, labeled "The maximum Pooling Layer 1," shows macro and micro averages, 91.91% and 90.91%, respectively,
the size of the output following the application of maximum were also quite high. These results show that the model has
pooling to the preceding layer. The pooling window size used the ability of effectively predict both the general
for down sampling is indicated by the kernel size. S2 stands categorization of mint leaf illnesses as well as specific
for the pooling stride, and 0 denotes that no padding is disorders. The author's study's findings imply that their
employed. Since layers that pool resources do not add any approach can be employed as a useful instrument for illness
parameters, this layer normally does not employ an activation identification and diagnostics in the agricultural sector.
function and has no filters or parameters. For the next layers, Accurate and speedy diagnosis of damaged plants can result
the following rows adhere to a similar pattern. Output form, in more focused and effective.
kernel size, strides S3, padded P2, ReLU being activated, F2
filters, as well as parameters are all features of convolutional
References
layer 2. No padding, stimulation, filters, or variables are
included in Max Pooling Layer 2, and it has a result shape, a [1] Shekhar, Shashi, H. Garg, R. Agrawal, S. Shivani, and B. Sharma,
kernel, and a stride S4. Convolutional Third Layer & “Hatred and trolling detection transliteration framework using
Maximum Pooling Layer 3 share the same structure. The hierarchical LSTM in code-mixed social media text,” in Complex &
"Fully Linked Layer" features an output that, when the Intelligent Systems, pp.1–14, 2021.
preceding layer has been flattened, shows the physical [2] Trivedi et al., “Early detection and classification of tomato leaf
disease using high-performance deep neural network,” Sensors 21,
dimensions of the output. This layer has no padding, stride, or
Vol.21, Issue.23, p.7987, 2021.
kernel size. The output is subjected to ReLU activation; [3] R. Sharma and V. Kukreja, “Amalgamated convolutional long term
therefore the amount of filters is irrelevant. (H6 x W6 x D3) x network (CLTN) model for Lemon Citrus Canker Disease Multi-
N determines the number of parameters. This table 3 provides classification,” in in International conference on decision aid
crucial details for reproducing and analyzing your study sciences and applications (DASA), pp.326–329, 2022.
results, serving as a useful reference for comprehending the [4] V.Bhardwaj and V. Kukreja, “Effect of pitch enhancement in Punjabi
children’s speech recognition system under disparate acoustic
layer setups and their associated variables in your CNN conditions,” Appl. Acoust., Vol.177, pp.107918–107931, 2021.
design. [5] Sathiya, M. S. J. V., and V. Jeyabalaraja, “Plant Disease
Table 3. Systematic Review of Layer Parameters Classification of Basil and Mint Leaves using Convolutional Neural
Layer Out Kern Stri Pad Acti Num Number Networks,” Int. J. Intell. Syst. Appl. Eng. 11, Vol.2, pp.153–163,
Type put el de ding vati ber of of 2023.
Sha Size on Filter Paramet [6] A. Hassan, Md, M. S. Islam, S. B. S. Md Mehedi Hasan, M. T.
pe s ers Habib, and M. S. Uddin, “Medicinal plant recognition from leaf
Convolut (H1 (K1, S1 P1 ReL F1 (K1 x images using deep learning,” Comput. Vis. Mach. Learn. Agric.,
ional W1 K1) U K1 x Vol.2, pp.137–154, 2022.
Layer1 D1) D1) x F1 [7] Lochan, R. Naga, A. S. Tomar, and R. Srinivasan, “Plant detection
Max (H2 (P1, S2 0 - - 0
and classification using fast region-based convolution neural
Pooling W2 P1)
networks,” Artif. Intell. Evol. Comput. Eng. Syst., pp.623–634,
Layer 1 D1)
2020.
Convolut (H3 (K2, S3 P2 ReL F2 (K2 x
ional W3 K2) U K2 x [8] Sathiya,V.,Josephine and D. Jeyabalaraja, “Identification And
Layer 2 D2) D2) x F2 Classification Of Diseases In Basil And Mint Plants Using
Max (H4 (P2, S4 0 - - 0 Psorbfnn,” J. Theory. Appl. Inf. Technol. 100, Vol.21, 2022.
Pooling W4 P2) [9] Borman, R. Indra, F. Rossi, D. Alamsyah, R. Nuraini, and Y.
Layer 2 D2) Jusman, “Classification of Medicinal Wild Plants Using Radial Basis
Convolut (H5 (K3, S5 P3 ReL F3 (K3 x Function Neural Network with Least Mean Square,” in In 2022 2nd
ional W5 K3) U K3 x International Conference on Electronic and Electrical Engineering
Layer 3 D3) D3) x F3 and Intelligent System (ICE3IS), pp.141–146, 2022.
Max (H6 (P3, S6 0 - - 0 [10] Hsieh, Yu-Lun, Y.-C. Chang and W.-Y. Chang, “Hsieh, Yu-Lun,
Pooling W6 P3) Yung-Chun Chang, and Wei-Yu Chang,” Nutr. 13, Vol.6, 2021.
Layer 3 D3) [11] Oliveira, Americo, S. Pereira, and C. A. Silva, “Augmenting data
Fully (N,) - - - ReL - (H6 x when training a CNN for retinal vessel segmentation: How to
Connecte U W6 x warp?,” 2017 IEEE 5th Port. Meet. Bioeng., pp.1–4, 2017.
d Layer D3) x N

© 2023, IJSRMS All Rights Reserved 11

Int. J. Sci. Res. in Multidisciplinary Studies Vol.9, Issue.9, Sept 2023

[12] Atha, A. Mutiara, and E. Zuliarso, “Deteksi Tanaman Herbal

Khusus Untuk Penyakit Kulit Dan Penyakit Rambut Menggunakan
Convolutional Neural Network (CNN) Dan Tensor flow,” JUPITER
(Journal Penelit. Ilmu dan Tek. Computer) 14, Vol.2-a, pp.1–10,
2022.
[13] Bakkali Yedri, Oussama, M. B. Mohamed Ben Ahmed, and L. El
Achaak, “A smart agricultural system to classify agricultural plants
and fungus diseases using deep learning,” in In Emerging Trends in
ICT for Sustainable Development: The Proceedings of NICE2020
International Conference, pp.229–239, 2021.
[14] Sunil, C. K., C. D. Jaidhar, and N. Patil, “Cardamom plant disease
detection approach using EfficientNetV2,” IEEE Access 10, pp.789–
804, 2021.

AUTHORS PROFILE
Dr.Umesha K. earned his B.E IN
Electrical Engineering, M.E in Power and
Energy Systems, and Ph.D. in 1989,
2004, and 2010, respectively. He is
currently working as Professor in
Department of Electronics &
Communication Engineering from
Jawaharlal College of Engineering and
Technology since 2012. He is a life member of ISTE since
2003, member of Professional Trainer for IMRT consortium -
USA since 2022,President for MoE IIC and a life member
for International Association of Academic plus
Corporate(IAAC), Nodal officer for KSITIL sponsored S D P
K, ARIIA, Coordinator NISP, UBA,Technical
member LITE. Technical Mentor for TBI,SPOC for
PMKVY-4.0 under NSDC-Skill hub.
His main research work focuses on Power and Energy
systems ,Power electronics and drives ,Bio medical .He has
33 years of teaching experience and 13 years of research
experience.

Nandhiniumesh earned her B. Tech.,

VTU Karnataka in 2020. She is currently
pursuing M. Tech in cyber security from
California Sate University, Dominguez
Hills Carson, USA .

© 2023, IJSRMS All Rights Reserved 12