Detection and Classification of Leaf Diseases in Tomato Plants and

Recommendations for Controlling the Spread

Sarakesh R.

, Jenila Livingston L. M.

, Rajkumar S.

and Agnel Livingston L. G. X.

School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, TN, India

St. Xavier’s Catholic College of Engineering, Nagercoil, TN, India

Keywords: Leaf Disease Detection, Deep Learning Techniques, Disease Classification, Convolutional Neural Network,

Ensemble Model, Agricultural Practices, Crop Yield.

Abstract: Leaf disease in tomatoes is the most important factor influencing crop output quantity and quality, hence

proper diagnosis and classification are essential. Different diseases affect tomato production. This study

focuses on employing deep learning techniques for the disease’s detection in tomato plants. Despite tomatoes

being a versatile ingredient highly sought after year-round, the significant annual loss in crop yields due to

diseases poses a substantial challenge in cultivation. The objective is to create a system capable of precisely

identifying various tomato diseases by analysing images. The dataset utilized in this study encompasses

different types and stages of tomato diseases, including Bacterial Spot, Early or Late Bright, Leaf Mold,

Spider Mites, Target Spot, Septoria Leaf Spot, Mosaic and Yellow Leaf Curl Viruses. Upon disease

identification, the study presents information on methods to control its spread. The models employed in this

study include Convolutional Neural Network (CNN), DenseNet169, and an ensemble model combining pre-

trained CNN and DenseNet169. The classification results of the study demonstrated an accuracy of 95% for

the ensemble model, surpassing the accuracy of individual models. This success in recognizing diseases in

tomato plants holds promise for enhancing agricultural practices.

1 INTRODUCTION

The intersection of deep learning and agriculture has

ushered in a transformative era for the detection and

identification of plant diseases. As the global

population burgeons, the imperative to ensure food

security has never been more critical, underscoring

the need to safeguard the health of crops, the bedrock

of our sustenance. Deep learning, a sophisticated

facet of artificial intelligence, emerges as a

formidable solution, employing intricate neural

networks to scrutinize images and plant data with

unprecedented precision. This technological

innovation facilitates the early identification of

diseases, empowering farmers to implement timely

preventive measures and thwart potential threats.

This project is aimed to harness the capabilities of

deep learning algorithms, offering a nuanced and

accurate discernment of specific plant issues. By

https://orcid.org/0000-0002-6333-5751

https://orcid.org/0000-0001-5860-7161

https://orcid.org/0000-0002-2222-1643

mitigating the need for extensive chemical

interventions, the initiative embraces sustainable

farming practices, fostering an agricultural landscape

that is both productive and environmentally

responsible. This pioneering approach not only

augments agricultural efficiency but also champions

sustainability, contributing to a robust and resilient

global food supply chain. The profound implications

of these advancements resonate far beyond fields and

farms, promising a future where crops are shielded,

and environmental impacts are curtailed, heralding a

new chapter in agriculture that harmonizes

productivity with ecological well-being.

2 RELATED WORK

Muammer Türkoğlu et. al. (2022) developed a CNN

Ensemble to detect plant diseases and pests. The

R., S., L. M., J. L., S., R. and L. G. X., A. L.

Detection and Classiﬁcation of Leaf Diseases in Tomato Plants and Recommendations for Controlling the Spread.

DOI: 10.5220/0012884700004519

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Emerging Innovations for Sustainable Agriculture (ICEISA 2024), pages 151-160

ISBN: 978-989-758-714-6

151

findings for deep feature extraction outperformed the

traditional classifiers. The work based on deep feature

extraction and classification with fine-tuned CNN

including fc6 layer of the AlexNet, Loss3 layer for

GoogleNet and fc1000 layer for ResNet50,

ResNet101 and DenseNet20. The majority voting

ensemble model attained the highest level of accuracy

(97.56%), next to the early fusion ensemble model

(96.83%).

Norhalina Senan et. al., (2020) proposed a model

that can reliably recognize the affected and healthy

paddy leaves, which is useful in automated paddy

categorization applications. The findings show that

the proposed CNN model outperformed (83%

accuracy) traditional classification techniques in

paddy leaf disease detection and classification.

Yong et. al. (2020) created the Inception-ResNet-

v2 model for early identification of pests.

Experiments demonstrated the recognition accuracy

of 86.1% and the results reveal that this hybrid

network model has a greater recognition accuracy

than the classic model and may be used to

successfully detect and classify the plant diseases and

insect pests.

Morteza Khanramaki et. al. (2021) developed an

ensemble technique for identifying citrus pests that

outperformed competing methods. Data

augmentation increases the quantity of pictures in the

dataset, which enhances classifier generalizability.

For the experimental analysis, a 10-fold cross

validation was performed to determine accuracy, and

it obtained 99.04%.

Lucas et al. (2021) implemented an integrated

CNN architecture that combines instance

segmentation with a Mask R-CNN and semantic

segmentation with UNet and PSPNet to detect

diseases and pests in coffee leaves. The MIoU for the

UNet and PSPNet networks was 94.25% and 93.54%,

respectively. The two networks produced very similar

results, with the UNet slightly outperforming the

PSPNet. However, PSPNet can be selected since its

lesion marker extends somewhat beyond its edge,

which can assist in lesion categorization, as the

intersection of the lesion and the healthy portion of

the leaf is not always immediately identifiable.

Several studies used neural networks to identify

and classify diseases. Earlier research employs shape,

color, and texture feature extraction approaches, as

well as typical machine learning classifiers. In more

recent investigations, CNN-based models have

shown significant success in the automated detection

of plant diseases and pests in leaves (Lu 2017, Liu

2017, Wallelign 2018, Picon 2019, Zhang 2019,

Rahman 2020, Wang 2020).

3 PROPOSED WORK

The existing manual methods for predicting disease

in plants are often crucial, labour-intensive, time-

consuming, lack of accuracy, not scale effectively to

meet the demands of large-scale agriculture. So the

objective of this research is to overcome these

challenges by utilizing the potential benefits of deep

learning techniques. The primary objective is to

design and develop an advanced deep learning-based

system capable of automatically detecting and

identifying plant diseases from images of plants.

This system will utilize CNNs and other deep

learning architectures to analyze visual data,

providing farmers with rapid, precise, and scalable

solutions for monitoring crop health. Ultimately, this

research will contribute to reducing yield losses,

promoting sustainable agriculture, and enhancing

global food production.

3.1 Research Challenges

Finding effective data augmentation and

preprocessing strategies to enhance image quality,

remove noise, and improve model robustness.

Designing models and algorithms that can scale to

handle large volumes of agricultural images

efficiently for timely detection of plant diseases.

Deep learning models often struggle with

generalizing their knowledge to new and unseen

conditions. For plant disease and pest detection,

models need to perform well across different seasons,

regions, and plant species. Achieving this level of

generalization while maintaining high accuracy is a

significant research challenge.

Enhance the interpretability of deep learning

models for plant disease detection. Understanding

how models make decisions is crucial for gaining

trust in their recommendations, especially in

agricultural decision-making.

3.2 Scope of the Project

In this research, we develop a comprehensive system

for the detection and identification of diseases in

plants through deep learning techniques. The scope

encompasses the collection of diverse plant data, the

implementation of advanced models, emphasizing

scalability and ethical considerations. It also involves

exploring novel methods.

ICEISA 2024 - International Conference on ‘Emerging Innovations for Sustainable Agriculture: Leveraging the potential of Digital

Innovations by the Farmers, Agri-tech Startups and Agribusiness Enterprises in Agricu

152

3.3 Data Preprocessing

Rescale: Rescaling involves adjusting the pixel

values of the images to fit within a specific range,

typically [0, 1] to ensure that all pixel values are

proportionate to each other. This process ensures

uniformity in pixel values across different images and

helps in standardizing the data for better processing

by deep learning models.

Shear Range: Shearing is a geometric

transformation that distorts an image by shifting one

part of it in a fixed direction. In this context, a shear

range of 0.2 means that the image can be distorted by

shifting parts of it by a maximum of 20% in a specific

direction. Shearing is useful for introducing

variations in images, which can be beneficial for tasks

such as data augmentation in image classification.

Brightness Range: Adjusting the brightness of

images randomly within a specified range is a

technique used to augment image data. Randomly

adjusting brightness helps in making the model more

robust and versatile to variations in diverse lighting

conditions during inference.

4 METHODOLOGY

Initially, preprocessing and augmentation were

performed, and the dataset was divided into training,

testing, and validation subsets. The evaluation

metrics for each model were gathered through

training, testing, and validation methods, allowing for

a full assessment of model performance. In this study,

CNN, Densenet169, and Ensemble techniques were

used to detect diseases in plants (Figure 1).

4.1 Convolutional Neural Network

The adeptness of CNNs in capturing intricate spatial

relationships within images makes them

indispensable tools for tasks requiring nuanced visual

understanding. CNNs unparalleled effectiveness is

underscored by their remarkable performance in a

spectrum of computer vision tasks, ranging from

accurate image classification to precise object

detection and nuanced image segmentation. Through

the seamless integration of convolutional, pooling,

and fully connected layers, CNNs stand as a

cornerstone in the realm of deep learning, offering

robust solutions to complex challenges in image

analysis and interpretation.

Figure 1: Proposed System Architecture

CNNs play a vital role in leveraging deep learning

for disease detection in plants. The use of CNNs in

this context involves the analysis of images of plants

to identify signs and symptoms of diseases. Filters

(also called kernels) are small learnable matrices that

slide over the input data to perform element-wise

multiplications, producing feature maps. The

architecture diagram of CNN model is given in Figure

Figure 2: Architecture of CNN.

Pooling Layers: Pooling layers are used to

minimize the spatial dimensions of the input volume,

Detection and Classiﬁcation of Leaf Diseases in Tomato Plants and Recommendations for Controlling the Spread

153

with typical pooling procedures including max

pooling and average pooling.

Convolutional layers apply convolution

operations to input data using learnable filters or

kernels. These filters slide over the input, capturing

local patterns and produces feature maps that

represent the presence of specific features or patterns

in the input data. in capturing spatial hierarchies and

are crucial for tasks like image recognition, where

local patterns are essential.

Flattening: Before feeding the output of

convolutional and pooling layers into fully connected

layers, the data is usually flattened into a vector.

Fully Connected Layers: After several

convolutional and pooling layers (Figure 3), CNNs

often include one or more fully connected layers for

making predictions based on the learned features.

Dense layers are fully connected layers, perform

weighted sum operations, applying activation

functions to produce non-linear mappings between

inputs and outputs.

Figure 3: Layers of CNN.

Dropout: Dropout is a regularization technique

commonly used in CNNs to prevent overfitting. It

randomly drops a certain percentage of neurons

during training to promote more robust learning.

Batch normalization is another regularization

technique that normalizes the inputs of a layer,

helping to stabilize and accelerate the training

process.

Loss Function: The choice of a loss function

depends on the specific task; for classification tasks,

cross entropy loss is used

4.2 DenseNet169 Model

Densenet169 is a variation of DenseNet with 169

layers intended to create a deeper network than the

original DenseNet. DenseNet169, being a deeper

model within the DenseNet family, may be suitable

for tasks that require capturing more intricate patterns

in the data.

4.2.1 Key Features and Characteristics of

DenseNet169

Dense Blocks: The network is organized into dense

blocks, each containing multiple densely connected

layers. Within each dense block, the output of each

layer is concatenated with the feature maps of all

previous layers, facilitating feature reuse and compact

model representation. Dense Connectivity for Feature

Extraction: Leverage the dense connectivity within

DenseNet169 for effective feature extraction. The

dense blocks allow for the reuse of features from

previous layers, enabling the model to capture

hierarchical representations of disease-related

patterns in plants.

Bottleneck Layers: DenseNet169 includes

bottleneck layers within dense blocks to reduce

computational complexity. These bottleneck layers

consist of a 1x1 convolution layer followed by a 3x3

convolution layer, which helps in efficient feature

extraction. These layers help reduce the number of

parameters and computational load while maintaining

the capacity to capture complex patterns.

Freezing Base Layers: Freezing base layers, uses

a loop to set all layers in the pre-trained DenseNet169

base model to non-trainable. This prevents these

layers from being updated during the subsequent

training process, ensuring that the previously learned

features remain fixed and only the additional layers

on top are fine-tuned for the specific task. This helps

retain valuable pre-trained knowledge while adapting

the model to a new classification objective.

Transition Layers: Transition layers are inserted

between dense blocks to control the spatial

dimensions of feature maps and manage the spatial

resolution, contributing to the overall efficiency of

the model. These transition layers typically consist of

a 1x1 convolution layer followed by average pooling.

Global Average Pooling: DenseNet169, like other

DenseNet models, employs global average pooling at

the end of the network instead of traditional fully

connected layers. This contributes to a fixed-size

feature vector for each image, which is then used for

disease classification.

ICEISA 2024 - International Conference on ‘Emerging Innovations for Sustainable Agriculture: Leveraging the potential of Digital

Innovations by the Farmers, Agri-tech Startups and Agribusiness Enterprises in Agricu

154

Figure 4: Architecture of Densenet169 model.

Figure 5: layers of Densenet169

Parameter Efficiency: DenseNet169 is designed to

be parameter-efficient. The dense connectivity

structure allows the model to achieve competitive

performance with fewer parameters compared to

traditional architectures.

Transfer Learning: Pre-trained versions of

DenseNet169 on large datasets, such as ImageNet, are

available, making it suitable for transfer learning on

tasks with limited labeled data. It allows the model to

leverage knowledge gained from a broader set of

visual features before fine-tuning on the specific plant

disease dataset.

DenseNet169 is a specific variant (Figure 4),

known for its increased depth with 169 layers. It

incorporates several key layers, similar to other

DenseNet models.

These layers collectively

contribute to the unique architecture of DenseNet169.

The dense connectivity, bottleneck layers, and

transition layers (Figure 5) are key features that allow

DenseNet to capture complex patterns in the data

effectively.

4.3 Ensemble Model of CNN and

DenseNet169

Ensemble models in deep learning involve combining

predictions from multiple individual models to

improve overall performance and generalization.

Ensemble models are particularly useful when

dealing with diverse data, reducing overfitting, and

improving model robustness. They are often

employed in situations where individual models may

have different strengths and weaknesses. While

creating ensemble models requires more

computational resources, they can yield better

generalization performance compared to individual

models. The key features of ensemble model are as

follows:

Bagging (Bootstrap Aggregating): Train multiple

deep learning models on different subsets. Bootstrap

aggregating the predictions through voting or

averaging.

Boosting: In deep learning, boosting can be applied

by training shallow networks sequentially, where

each subsequent network focuses on the

misclassifications of the previous ones.

Examples

include AdaBoost and Gradient Boosting.

Stacking: In deep learning, the base models can be

different architectures or variations of the same

architecture. The meta-model, often a simpler model,

learns to weight or combine the outputs of the base

models for predictions.

Weighted Average Ensembles:

Assign different

weights to the predictions of each model based on

their individual performance. Weights can be

determined through cross-validation or performance

Detection and Classiﬁcation of Leaf Diseases in Tomato Plants and Recommendations for Controlling the Spread

155

on a validation set. The final prediction is a weighted

sum of the individual predictions.

Model Distillation: The student model learns not only

from the ground truth labels but also from the soft

labels (probabilities) provided by the teacher model.

Train a larger, more complex model (teacher model)

and then use its predictions to train a smaller model

(student model).

4.3.1 Workings of Ensemble Model of CNN

and DenseNet169

Creating an ensemble model with a combination of

CNN architecture (e.g., ResNet, VGG, Inception) and

DenseNet169 (Figure 6) for plant disease detection

involves leveraging the strengths of each architecture

to enhance overall performance of image

classification.

Figure 6; Architecture diagram of Ensemble model

Individual Model Training: Train the CNN and

DenseNet169 models independently on the training

dataset. Use transfer learning by initializing the

models with pre-trained weights on large-scale

datasets like ImageNet.

Model Outputs: For each input image, both the CNN

and DenseNet169 models produce predictions,

indicating the likelihood of the presence of a disease.

If it's a binary classification task, the models output

probabilities or binary predictions.

Ensemble Aggregation: Combine the predictions

from both models using an ensemble strategy.

Common strategies include:

Voting (Hard or Soft): For binary classification, use a

majority vote for hard voting or average the

probabilities for soft voting.

Weighted Averaging: Assign different weights to the

predictions of each model based on their individual

performance. The final prediction is a weighted sum

of the individual predictions.

Ensemble Model Output: The final output of the

ensemble model is the aggregated prediction from

both the CNN and DenseNet169 models

5 RESULTS AND DISCUSSION

Figure 7 shows the model configuration of a CNN for

plant disease detection and classification. Figures 8

and 10 show the accuracy and test loss of a CNN

model, respectively.

Figure 7: Model layout of CNN model

Figure 8: CNN model Accuracy

The following images (Figure 9 a,b,c) show the

prediction of tomato leaf diseases with true and

predicted labels using CNN, Densenet169, and

Ensemble models.

ICEISA 2024 - International Conference on ‘Emerging Innovations for Sustainable Agriculture: Leveraging the potential of Digital

Innovations by the Farmers, Agri-tech Startups and Agribusiness Enterprises in Agricu

156

(a)

(b)

(c)

Figure 9: Prediction of tomato leaf disease with true labels and predicted labels. a) Prediction using CNN b) Prediction using

Densenet169 c) Ensemble model

Detection and Classiﬁcation of Leaf Diseases in Tomato Plants and Recommendations for Controlling the Spread

157

Figure 10: CNN model test loss

The below image (Figure 11) depicts the architecture

of the Densenet169 model, the total params has

increased compared to the previous model.

Figure 11: Model layout of Densenet169 model

A confusion matrix is a table that defines a

classification model's output based on test data with

known true values. Figures 12, 13 and 14 shows the

confusion matrix of CNN, Densenet169, and

ensemble model respectively.

Figure 12: Confusion matrix of CNN model

Figure 13: Confusion matrix of Densenet169 model

ICEISA 2024 - International Conference on ‘Emerging Innovations for Sustainable Agriculture: Leveraging the potential of Digital

Innovations by the Farmers, Agri-tech Startups and Agribusiness Enterprises in Agricu

158

Figure 14: Confusion matrix of ensemble model

Table 1 represents the accuracy and test loss of the

models. Among the three models, the ensemble

model demonstrated superior accuracy in identifying

plant diseases, achieving an impressive 95.3%

accuracy rate. This demonstrates the usefulness of

combining the strengths of CNN and DenseNet169 in

a collaborative framework to improve performance.

Table 1: Accuracy and Test loss of models

Models Accuracy Test loss

CNN 94% 17.2%

Densenet169 91% 28.1%

Ensemble 95.3% 9.1%

6 CONCLUSION AND

RECOMMENDATION

This study focused on the crucial task of detection and

identification of disease in tomato plant, addressing a

pressing need in agriculture for early and accurate

diagnosis. Through the utilization of advanced

technology such as deep learning, the project

successfully developed a robust system capable of

recognizing and categorizing plant diseases

efficiently. The implemented solution demonstrated

its effectiveness in automating the detection process,

allowing for timely interventions to prevent the

spread of diseases and mitigate potential crop losses.

By leveraging the power of deep learning, the project

not only enhanced the speed and accuracy of disease

identification but also provided a scalable and

adaptable framework that can be extended to various

crops and regions. Furthermore, the project

contributes to sustainable agriculture practices by

promoting precision farming, reducing the reliance on

misuse of chemical treatments, and ultimately

fostering a more resilient and productive food supply

chain.

This study focused solely on investigating few

diseases affecting only one crop species, excluding

others such as brinjal, ladies finger, chili, and their

respective diseases. Hence, the next phase involves

acquiring additional images of crop species and

diseases for research purposes. Despite achieving

commendable recognition accuracy, the models

warrant further exploration and optimization.

Simultaneously, there's a need to develop a network

model capable of classifying crop images with greater

precision.

REFERENCES

Turkoglu, M., Yanikoğlu, B. and Hanbay, D. Plant Disease

Net: Convolutional Neural Network Ensemble for Plant

Disease and Pest Detection. SIVIP 16, 301–309 (2022).

Https://Doi.Org/10.1007/S11760-021-01909-2

Norhalina Senan , Muhammad Aamir , Rosziati Ibrahim ,

N. S. A. M Taujuddin, and W.H.N Wan Muda, An

Efficient Convolutional Neural Network for Paddy

Leaf Disease and Pest Classification, International

Journal of Advanced Computer Science and

Applications(IJACSA), Volume 11 Issue 7, (2020).

Yong Ai, C. Sun, J. Tie and X. Cai, "Research on

Recognition Model of Crop Diseases and Insect Pests

Based on Deep Learning in Harsh Environments," In

IEEE Access, Vol. 8, Pp. 171686-171693, (2020), Doi:

10.1109/Access.2020.3025325.

Morteza Khanramaki, Ezzatollah Askari Asli-Ardeh and

Ehsan Kozegar , Citrus Pests Classification Using an

Ensemble of Deep Learning Models , Computers and

Electronics in Agriculture, Volume 186, (2021),

Https://Doi.Org/10.1016/J.Compag.2021.106192

Lucas M. Tassis, João E. Tozzi De Souza , and Renato A.

Krohling, Deep Learning Approach Combining

Instance and Semantic Segmentation to Identify

Diseases and Pests of Coffee Leaves from In-Field

Images, Computers and Electronics in Agriculture,

Detection and Classiﬁcation of Leaf Diseases in Tomato Plants and Recommendations for Controlling the Spread

159

Volume 186, (2021),

Https://Doi.Org/10.1016/J.Compag.2021.106191

Lu, Y., Yi, S., Zeng, N., Liu, Y., and Zhang, Y.:

Identification Of Rice Diseases Using Deep

Convolutional Neural Networks. Neurocomputing 267,

378–384 (2017)

Liu, B., Zhang, Y., He, D., and Li, Y.: Identification of

Apple Leaf Diseases Based on Deep Convolutional

Neural Networks. Symmetry, 10(1), Art. 11 (2017).

Wallelign, S., Polceanu, M., and Buche, C.: Soybean Plant

Disease Identification Using Convolutional Neural

Network. In K. Brawner and V. Rus (Eds.),

Proceedings Of 31st International Flairs Conference,

Pp. 146–151 (2018).

Picon, A., Alvarez-Gila, A., Seitz, M., Ortiz-Barredo, A.,

Echazarra, J., and Johannes, A.: Deep Convolutional

Neural Networks For Mobile Capture Device-Based

Crop Disease Classification In The Wild. Comput.

Electron. Agric. 161, 280–290 (2019)

Zhang, S., Huang, Z., and Zhang, C.: Three-Channel

Convolutional Neural Networks for Vegetable Leaf

Disease Recognition. Cogn. Syst. Res. 53, 31–41 (2019)

Rahman, C.R., Arko, P.S., Ali, M.E., Khan, M.A.I., Apon,

S.H., Nowrin, F., and Wasif, A.: Identification and

Recognition of Rice Diseases and Pests Using

Convolutional Neural Networks. Biosys. Eng. 194,

112–120 (2020).

Wang, L., Sun, J., Wu, X., Shen, J., Lu, B., and Tan, W.:

Identification of Crop Diseases Using Improved

Convolutional Neural Networks. Iet Comput. Vision

14(7), 538–545 (2020)

ICEISA 2024 - International Conference on ‘Emerging Innovations for Sustainable Agriculture: Leveraging the potential of Digital

Innovations by the Farmers, Agri-tech Startups and Agribusiness Enterprises in Agricu

160