Available online at www.sciencedirect.

com

ScienceDirect
Procedia Computer Science 252 (2025) 341–354

4th International Conference on Evolutionary Computing and Mobile Sustainable Networks

A Hybrid Learning Model for Tomato Plant Disease

Detection using Deep Reinforcement Learning with Transfer
Learning
Kadambari Raghurama* , Malaya Dutta Borahb
aDepartment of Computer Science and Engineering, National Institute of Technology Silchar, Silchar,
Assam, India., kadambari21_rs@cse.nits.ac.in
bDepartment of Computer Science and Engineering, National Institute of Technology Silchar, Silchar,

Assam, India., malayaduttaborah@cse.nits.ac.in

Abstract

Plant diseases play a significant role in damaging crop production and food security. Detecting and diagnosing plant
diseases in the early stages obtains better management of diseases. Many plants are affected by these diseases,
which are very dangerous for crop yield. This paper introduces advanced plant disease detection using an advanced
preprocessing technique and a Hybrid Learning Model (HLM). The advanced preprocessing method uses a digital
camera to capture high-resolution images of plant leaves from multiple angles. These images are then processed
using an enhancement algorithm to improve the visual quality and clarity. The preprocessed images are subjected to
a HLM model, which utilizes Deep Reinforcement Learning with Transfer Learning (DRL-TL). The DRL-TL
architecture is designed to extract features from the preprocessed images in a three-dimensional manner, considering
the spatial information of the leaves. It enables the model to capture precise patterns and variations indicative of
disease symptoms. The pre-trained model MobileNetV2 trained on a tomato disease dataset belongs to labeled
images; it consists of standard and affected plant leaves to learn the discriminative features associated with different
diseases. Results obtained the effectiveness of the hybrid learning model. The preprocessing technique significantly
increases the input images' quality, enhancing the subsequent HLM's performance. The model accurately identifies
and classifies various plant diseases, outperforming existing methods. Furthermore, the hybrid learning model shows
robustness and abstraction, successfully detecting diseases across plant species and environmental conditions.
© 2025 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the 4th International Conference on Evolutionary Computing and
Mobile Sustainable Networks
Keywords: Deep Learning; preprocessing; Reinforcement Learning; Transfer Learning; Hybrid Learning Model; MobileNetV2;

* Kadambari Raghuram.
E-mail address: kadambari21_rs@cse.nits.ac.in

1877-0509 © 2025 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the 4th International Conference on Evolutionary
Computing and Mobile Sustainable Networks
10.1016/j.procs.2024.12.036
342 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354

1. Introduction

Plant diseases affect global food supply and sustainable farming. Early and accurate diagnosis of these diseases is
critical for effective crop preservation and avoiding loss [1] [2]. In recent years, deep learning methods have become
very effective tools for automated and effective plant disease identification. Artificial neural networks (ANN) are
the foundation of deep learning (DL) is a subset of machine learning (ML) that seeks to emulate the architecture and
functions of the human brain. These networks are ideal for image analysis tasks because they can learn and extract
relevant characteristics from complex datasets in real time. Deep learning for plant disease detection involves
training deep neural networks on large datasets of plant images, with each image labeled with its corresponding
disease or healthy status [3] [4]. The trained models can then analyze new images and accurately classify them into
healthy or diseased categories based on learned patterns and features. One of the primary benefits of DL in plant
disease diagnosis is its capacity to deal with multidimensional and diverse data, such as images captured under
different lighting conditions, scales, and perspectives. Deep learning models can learn complex patterns and subtle
variations in plant images that may be indicative of diseases, even when human experts find it challenging to
differentiate. Moreover, deep learning-based approaches can provide fast and automated disease detection, enabling
early intervention and appropriate management strategies. This can significantly reduce crop losses, minimize the
use of chemical treatments, and promote sustainable agriculture practices [5] [6].
Plant infections are a major challenge in farming, resulting in significant quality and yield losses. Conventional
detection of plant illnesses methods, such as skilled observation, are time-consuming and labor-intensive—deep
learning models trained using large healthy and diseased plant image datasets. These models can correctly identify
new photos of plants as healthy or unhealthy once they have been trained. Without the assistance of human
specialists, this can be completed swiftly. Plant disease detection has been applied to a variety of deep-learning
architectures. The requirement for massive image databases of healthy and sick plants presents one difficulty. These
datasets can be expensive and time-consuming to collect. Another challenge is the need to develop DL models that
are robust to variations in lighting, background, and other factors. This is because plant diseases can appear
differently in different conditions [7-10]. Despite these challenges, deep learning has more strength to detect plant
diseases. As the model continued to develop by adding a new approach, more accurate and reliable DL models were
introduced.
Deep Reinforcement Learning with Transfer Learning (DRL-TL) is a machine learning techniques combined to
detect tomato plant diseases. DRL-TL begins by training a deep reinforcement learning agent on images of healthy
and diseased tomato plants. The agent learns to recognize the characteristics that differentiate healthy plants from
diseased plants. Once trained, the agent can detect diseases in new images of tomato plants. Transfer learning used
to boost DRL-TL performance [11] [12]. Transfer learning entails using a previously trained deep learning model as
the foundation for training a new model. It can be beneficial as it can save both energy and time while also helping
to improve the new model's performance [13] [14]. Transfer learning can be used to train a DRL-TL agent that can
identify a wider range of diseases in tomato plants. It is because the pre-trained model will have already learned to
recognize features common to many different types of diseases.

Figure 1: Types of Tomato Plant Diseases

 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354 343

1.1. Contribution

This work mainly focused on detecting the causes and detection of tomato plant diseases. The proposed approach
is the integrated model that contains the Augmentation technique with advanced preprocessing filtering techniques
such as Adaptive Median Filter (AMF). MobileNetV2, a pre-trained model, is employed with transfer learning. To
extract significant features from the input samples, use an ensemble feature extraction technique such as Color
Histogram Computation (CHT). Finally, Deep Reinforcement Learning with Transfer Learning (DRL-TL) is
employed to categories tomato disease leaf images.

2. Literature Survey

Gao et al. [15] discussed various deep learning (DL) models that can detect crop leaf disease from the given crop
leaf images. The proposed approach mainly focused on plant diseases detection and insect pests. Also discussed
were various advantages and issues that required to be solved. Azimi et al. [16] introduced a hybrid model, such as
CNN-LSTM, to find the water stress classification between the chickpea plant dataset. It is very significant to
identify the plant health and status of diseases by using existing models. These models mainly failed to detect the
water stress level at the plants. To overcome this, the proposed model CNN-LSTM focused on water stress among
the two datasets, such as chickpea models JG-62 and Pusa-372. Finally, the classification model achieved an
accuracy of 98.12% on Pusa-372 and 98.34% on JG-62.

B. Liu et al. [17] developed Leaf GAN, a novel approach for creating four different types of grape leaf diseases. The
training model deep regret gradient approach is used to find accurate dataset features to identify genuine and fake
disease images. Additionally, it takes elements out of the photos of grape leaves, which aids in precisely identifying
the diseases. Ultimately, the accuracy of the suggested model was 98.67%, which is high compared to other
currently used models, such as WGAN and DCGAN. Citrus leaf pictures gathered from the plant village and crowd
were found to be impacted by Huanglongbing (HLB) using a hybrid model established by Q. Zeng et al. [18] in
conjunction with Inceptionv3 and DCGANs. The main aim of this model is to find the seriousness of the HLB
diseases among the given leaf images. The training model Inceptionv3 was mainly used to train the citrus images,
which obtained high accuracy. The Inceptionv3 combined with DCGANs gives an accuracy of 93.23%, which is
very high compared with existing models.

E. Ozbılge et al. [19] presented a model that finds the tomato disease from the given dataset. The training model
ImageNet has used a transfer learning approach that helps transfer the deep network models. The accuracy is
99.78% for the testing set from the plant village dataset. K. Roy et al. [20] proposed the PCA-based DNN model that
finds the diseases among the tomato leaves. It also adopted the GAN model to get better results. It achieves the
accuracy of 99.45% and a precision of 98.45%. Y. Wu et al. [21] proposed a fine-tuned classification model that
mainly focused on solving the issues in classification. Finally, the classification model was utilized to improve the
finding potentiality of disease detection. Two real-time datasets, such as peach and tomato leaf diseases, are used for
the performance evaluation. The accuracy improved based on the features such as the redesigned model, and
Discrimination Model helps find the conditions.

N. Ullah et al. [22] proposed the unique end-to-end DeepPestNet system that finds the pests and classification. It
contains 11 layers with eight conv and three FC layers. The size of the image dataset achieved better augmentations
to get better results. The classification focused on classifying several pests among the given "Pest dataset." The
proposed approach gives the accuracy of 98.34%. S. Ahmed et al. [23] proposed a LTBA diagnosis the tomato plant
leaf diseases. An efficient preprocessing method removes the noise and lighting modification in the given leaf
images. An effective and efficient feature extracted by using the pre-trained MobileNetV2 architecture that classifies
the diseased and non-diseased plant images. Existing models obtained several issues, such as data leakage, and
found the imbalanced problem. The accuracy is 99.32% for the proposed model, which is high. E. Elfatimi et al.
[24] proposed the classification model based on bean leaf disease. MobileNetV2 gives the training for this dataset to
train very fastly. The Mobile-Net model was also used for the bean leaf dataset for testing. The classification mainly
focused on three classes: standard, angular spot disease and bean rust disease. The obtained training accuracy is
97.67% and a testing accuracy is 92.34%. An upgraded disease detection model for apple leaves was proposed by
Jiang, P., and others [25] discussed various pre-trained models such as the Google-NET Inception framework and
Rainbow merging. The suggested model can distinguish between the five primary apple leaf diseases at a rate of
23.13 FPS. Compared to earlier models, the results demonstrate that the presented model offers high and quick
disease detection.
344 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354

C. Zhou et al. [26] proposed using a redesigned residual dense network (RDN) to detect tomato leaves. The
proposed model combines DRN and RDN, which reduces the training metrics to increase the measurement accuracy
with advanced data. Generally, the RDN mainly focuses on the resolution of images, which helps redesign the
proposed architecture with advanced classification to get better results. The DCGAN model was first introduced by
Q. Wu et al. [27] to identify tomato plant diseases. The accuracy of 95.12% is determined by using the suggested
model trained using GoogLe-Net and data augmentation. The Visual Turing Test and the performance of t-
distributed Stochastic Neighbour embedding (t-SNE) were combined to address the mismatched problems with the
current models.

3. Dataset Description

The tomato plant disease dataset collected from the PlantVillage database contains training, testing, and validation
folders. It includes nine disease types and one health type sample. 14k images are ready for experimentation.
Finally, unnecessary images are removed from the dataset, and the image size is reduced from 256 * 256 to 227 *
227.
Table 1: The Training and Testing of Tomato Plant Diseases Dataset
Dataset Samples
Training 7k
Testing 7k
Total 14k

Figure 2: Architecture of MobileNetV2

3.1 MobileNetV2
Tomato plants are susceptible and are affected by several diseases, gradually reducing yield and quality. Prediction
of tomato plant diseases in the early stages may prevent crop loss and improve production. This paper uses a pre-
trained model MobileNetV2 to extract the features from every tomato image, which helps in classification. It is also
called a lightweight CNN model that can be used for several cases. It also provides better accuracy and efficiency
and is most widely used in real-time applications. The aim of MobileNetV2 model used for tomato plant disease
detection. By utilizing a large dataset of labeled images, the model will learn to differentiate between healthy tomato
plants and various disease symptoms. These symptoms may include leaf spots, yellowing, wilting, fungal infections,
and other common diseases affecting tomato plants.
The architecture of MobileNetV2 consists of several building blocks, including depth-wise separable convolutions,
inverted residual blocks, and linear bottlenecks. The architecture steps are as follows:
Input: It is the color input image with the size of 224x224 pixels.
 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354 345

Initial Convolution: The input tomato image was sent to CNN with a tiny kernel size and a stride of 2. This layer
performs basic feature extraction and reduces the spatial dimensions.
Inverted Residual Blocks: MobileNetV2 extensively uses inverted residual blocks, which are designed to balance
model size and computational efficiency. Each block consists of the following steps:
a. Depth-wise Separable Convolution: It splits the default CNN into a depth-wise convolution and a point-by-point
convolution. The number of factors and computing expenses decreases as a result.
b. Expansion: Next, a linear bottleneck layer expands the number of channels, allowing the network to capture
richer representations.
c. Squeeze and Excitation: In some variants of MobileNetV2, each inverted residual block incorporates a squeeze-
and-excitation module. This module adaptively scales the channels based on their importance, improving the
model's expressive power.
d. Skip Connections: Skip connections are added in some blocks to enable the gradient flow and facilitate training.
Final Layers: After several stacked inverted residual blocks, the network applies a few additional layers to further
refine the features.
The training process involves several steps. First, a comprehensive dataset of tomato plant images, comprising both
healthy and diseased samples, is collected and annotated. The dataset should be diverse, including images from
different perspectives, lighting conditions, and stages of disease progression. Three folders—training, validation,
and testing sets—are included in the dataset. Using back-propagation, the model learns to minimize the gap between
the original value and the estimated value based on illness labels, which helps to enhance the training process. The
model's performance was tracked through training and adjusting measures to validate it.
The model's generalization improved by adopting the data augmentation techniques like rotation and scaling, which
is applied to the training set. It prevents various overfitting and enhances the model's strength to tackle the variations
in the real-time datasets. At the time of training, the MobileNetV2 processes several epochs, with every epoch
consisting of front and back passes through the network. It is evaluated on validation set after each epoch to monitor
its progress.
Once the training is complete, the final trained MobileNetV2 model can be deployed to detect diseases in real-time
tomato plant images. It takes an input image and generates predictions, indicating whether the plant is healthy or
infected with a specific disease. The model's accuracy can be assessed by evaluating its performance on the testing
set, which contains images not seen during training or validation.
The following data augmentation techniques that are used in this paper
One common data augmentation technique for image classification tasks, including the detection of tomato plant
diseases, is rotation. The image is rotated to certain angle that creates new training samples. The rotation can be
performed clockwise or counterclockwise.
The basic equation for rotating an image around its center is as follows:

𝑥𝑥 ′ = 𝑐𝑐𝑐𝑐𝑐𝑐(𝜃𝜃) ∗ (𝜒𝜒 − 𝜒𝜒𝑐𝑐 ) − 𝑠𝑠𝑠𝑠𝑠𝑠(𝜃𝜃) ∗ (𝛶𝛶 − 𝛶𝛶𝑐𝑐 ) + 𝜒𝜒𝑐𝑐 (1)

𝑦𝑦 ′ = 𝑠𝑠𝑠𝑠𝑠𝑠(𝜃𝜃) ∗ (𝜒𝜒 − 𝜒𝜒𝑐𝑐 ) − 𝑐𝑐𝑐𝑐𝑐𝑐(𝜃𝜃) ∗ (𝛶𝛶 − 𝛶𝛶𝑐𝑐 ) + 𝛶𝛶𝑐𝑐 (2)
In this equation:
(𝑥𝑥, 𝑦𝑦) Coordinates of original image.
(x', y') coordinates of the rotated image.
𝑥𝑥𝑐𝑐 , 𝑦𝑦𝑐𝑐 coordinates of center point among rotation is performed.
θ represents the rotation angle.
By applying this rotation equation to each pixel in the original image, a new image that is rotated by a specific angle
can be obtained.
For instance, if the image needs to be rotated at an angle of 45 degrees counterclockwise, then θ = -45 degrees is
used in the equation. If it is clockwise, θ = 45 degrees has to be used.
3.2 Uniform Scaling:
This equation scales the image uniformly in both dimensions by a factor of "s."
𝑁𝑁𝑁𝑁𝑁𝑁 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ = 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ ∗ 𝑥𝑥 (3)
𝑁𝑁𝑁𝑁𝑁𝑁 ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡 = 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡 ∗ 𝑦𝑦 (4)
3.3 Non-Uniform Scaling:
346 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354

This equation scales the image by different factors for width and height. Use "sx " to scale the width and "sy " to scale
the height.
𝑁𝑁𝑁𝑁𝑁𝑁 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ = 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ ∗ 𝑠𝑠𝑥𝑥 (5)
𝑁𝑁𝑁𝑁𝑁𝑁 ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡 = 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡 ∗ 𝑠𝑠𝑦𝑦 (6)

3.4 Aspect Ratio Preservation:

This equation scales the image while preserving the aspect ratio. The target width (tw) or target height (th) are
defined and used to calculate the scaling factors (sx and sy) accordingly.
𝑡𝑡𝑤𝑤 𝑡𝑡ℎ
𝑠𝑠𝑥𝑥 − 𝑠𝑠𝑦𝑦 = 𝑚𝑚𝑚𝑚𝑚𝑚 ( , ) (7)
𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡

3.5 Random Scaling:

Random scaling introduces a random factor to the scaling equation to add variability. It is defined as a range of
scaling factors 𝑚𝑚𝑚𝑚𝑚𝑚𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑎𝑎𝑎𝑎𝑎𝑎 𝑚𝑚𝑚𝑚𝑚𝑚𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 and randomly select a scaling factor for each image.
A scaling factor "s" selected between 𝑚𝑚𝑚𝑚𝑚𝑚𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑎𝑎𝑎𝑎𝑎𝑎 𝑚𝑚𝑚𝑚𝑚𝑚𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠
3.6 Adaptive Median Filter (AMF):
The AMF is a nonlinear filter that effectively removes impulse noise, also known as salt-and-pepper noise, which
appears as isolated white and black pixels. Figure 3 shows the step-by-step process proposed approach and included
in this figure. This filter examines local neighborhoods of pixels and replaces the central pixel with the median value
from the neighborhood if it is considered an outlier. The AMF is useful for removing sporadic noise present in plant
disease images, which may result from factors like image acquisition, sensor noise, or other environmental factors.
By adaptively adjusting the size of the neighborhood based on the noise level, this filter can effectively suppress
impulse noise while preserving image details.
Start with the anisotropic diffusion filter to reduce overall noise while preserving edge information and important
image structures. Adjust the parameters of the filter, such as the number of iterations and diffusion coefficients,
based on the particular factors based on plant disease images.
The algorithm for the AMF is as follows:
For each pixel (x, y) in the image:
a. Define a window centered at (x, y) with an initial size s (s = 3, 5, 7, etc.).
b. Compute the median M of pixel values within the window.
c. Compute the minimum and maximum pixel values within the window.
d. If I(x, y) < min(M) or I(x, y) > max(M), increase the window size s by 2 and repeat steps b-d.
e. Set the filtered pixel value as median(M).
By iteratively increasing the window size until the noisy pixel falls within the valid range, the Adaptive Median
Filter adapts to the local image characteristics, ensuring better noise removal.
To combine these filters, first apply the Anisotropic Diffusion Filter to the plant disease image to reduce the overall
noise while preserving important edges. Then, apply the Adaptive Median Filter to further remove any remaining
noise and outliers.
3.7 Color Histogram Technique (CHT)
The color histogram technique most widely used method for preprocessing technique. It involves computing the
color distribution of an image using histograms and analyzing the variations in color to identify disease-related
changes. Here's an explanation of the color histogram technique with formulas and equations:

3.8 Image Preprocessing:

It is usual to practice preprocessing an image by converting it from the RGB color space to a different color system,
such as the HSV (Hue, Saturation, and Value) color space, which better simulates human perception of color before
generating the color histogram. The conversion can be performed using the following equations:
Hue (H):
1
( )((𝑅𝑅−𝐺𝐺)+(𝑅𝑅−𝐵𝐵))
𝐻𝐻 = 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 [ 2
, 𝑖𝑖𝑖𝑖 𝐵𝐵 <= 𝐺𝐺 (8)
𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠((𝑅𝑅−𝐺𝐺)2 +(𝑅𝑅−𝐵𝐵)(𝐺𝐺−𝐵𝐵))]
1
( )((𝑅𝑅−𝐺𝐺)+(𝑅𝑅−𝐵𝐵))
𝐻𝐻 = 360 − 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 [ 2
, 𝑖𝑖𝑖𝑖 𝐵𝐵 <= 𝐺𝐺 (9)
𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠((𝑅𝑅−𝐺𝐺)2 +(𝑅𝑅−𝐵𝐵)(𝐺𝐺−𝐵𝐵))]
 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354 347

Saturation (S):
3∗𝑚𝑚𝑚𝑚𝑚𝑚(𝑅𝑅,𝐺𝐺,𝐵𝐵)
𝑆𝑆 = 1 − (10)
𝑅𝑅+𝐺𝐺+𝐵𝐵
Value (V):
(𝑅𝑅+𝐺𝐺+𝐵𝐵)
𝑉𝑉 = (11)
3
These equations convert the RGB values of each pixel into corresponding HSV values, which are more suitable for
color analysis.
3.9 Deep Reinforcement Learning with Transfer Learning (DRL-TL)
A. Fine-tuning:
Add a new fully connected layer on top of the feature extractor with a softmax activation function, representing the
disease classes.
➢ Initialize the weights of the new layer randomly.
➢ Unfreeze the weights of the feature extractor layers and the newly added layer.
➢ The following shows the network training on labeled dataset
a. Convert the input images into feature vectors using the feature extractor.
b. Pass the feature vectors through the fully connected layer to obtain class probabilities.
c. Calculate the loss using a suitable loss function, such as categorical cross-entropy, comparing the predicted
probabilities with the true labels.
d. Update the weights of the network using back propagation and gradient descent, minimizing the loss.

B. Reinforcement Learning:
➢ Use the fine-tuned network as a policy network in a reinforcement learning framework.
➢ Define the state representation, action space, reward function, and environment dynamics suitable for the
plant leaf disease classification problem.
➢ Initialize the DRL agent's parameters and hyperparameters
➢ for episode in range(total_episodes)
➢ Initialize the environment and get the initial stat

for time_step in range(max_time_steps):

Select an action using the policy network (fine-tuned network)
The action should be performed in the platform and analyze the next state and reward
The tuple is stored (state, action, reward, next_state) in the response barrier
From the reply barrier, the mini-batch of experiences
Update the policy network using the sampled experiences and the DRL optimization
algorithm (e.g., DQN, A2C, PPO)
If the episode ends, break the inner loop
348 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354

Figure 3: System Architecture

4. Performance Metrics

The performance metrics mainly focused on estimating the labels over the original labels by showing the attributes.
The model's performance in each class and identifying any specific misclassification patterns. When evaluating the
performance of Deep Reinforcement Learning for Transfer Learning (DRL-TL) classification models, several
metrics can be used to assess their effectiveness. All the experiments conducted by using python programming
language by using several libraries such as keras, pandas etc.
In the context of binary classification, the terms "true positive," "true negative," "false positive," and "false
negative" are used to describe the outcomes of a prediction or classification.
True Positive (TP): Labeled value correct estimated value also correct.
True Negative (TN): Labeled value correct and estimated value wrong.
False Positive (FP): Labeled value wrong and estimated value right.
False Negative (FN): Labeled value right and estimated value wrong.
The following are some commonly used performance metrics:
𝑇𝑇𝑇𝑇+𝑇𝑇𝑇𝑇
𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 (𝐴𝐴𝐴𝐴𝐴𝐴) = (12)
𝑇𝑇𝑇𝑇+𝑇𝑇𝑇𝑇+𝐹𝐹𝐹𝐹+𝐹𝐹𝐹𝐹
 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354 349

𝑁𝑁𝑁𝑁 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇

𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 (𝑆𝑆𝑆𝑆𝑆𝑆) = (13)
𝑁𝑁𝑁𝑁 𝑜𝑜𝑜𝑜 𝑇𝑇𝑇𝑇+ 𝑁𝑁𝑁𝑁 𝑜𝑜𝑜𝑜 𝐹𝐹𝐹𝐹

𝑇𝑇𝑇𝑇
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 (𝑅𝑅𝑅𝑅) = (14)
𝑇𝑇𝑇𝑇+ 𝐹𝐹𝐹𝐹

(𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃∗𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅)
𝐹𝐹1 − 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 (𝐹𝐹1𝑆𝑆) = 2 ∗ (15)
(𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃+𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅)

𝑇𝑇𝑇𝑇
𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 (𝑃𝑃𝑃𝑃𝑃𝑃) = (16)
𝑇𝑇𝑇𝑇+𝐹𝐹𝐹𝐹
4.1 Training and Testing Loss
MobileNet is a family of lightweight deep-learning models for mobile and edge devices. These models are known
for their efficiency and low computational requirements, making them suitable for deployment on resource-
constrained devices. When training and evaluating a MobileNetV2 model, monitoring training loss and testing loss
is expected to assess the model's performance. Training loss is a measure of how well the model performs on
training data. It denotes the error associated with the model's recommendations and the target values. The purpose of
training is to reduce training loss. A low training loss indicates that the model is learning to generate better
predictions from the training data. The testing loss indicates how well the model generalizes to offered image data. It
is computed on a separate dataset that the model has not seen during the training phase. The testing loss helps to
assess whether the model is over fitting (performing well on training data). The proposed pre-trained model
MobileNetV2 obtained a low training loss of 12.12% and a low testing loss of 13.56%.

Training Loss Testing Loss

0.5
Percentage(%)

0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8 9 10
No of Epochs

Figure 4: Training and Testing Loss of MobileNetV2

4.2 Training and Testing Loss

Training accuracy measures a model's performance on the training dataset. It is calculated by dividing the total
number of training examples by the number of occurrences correctly predicted by the proposed model. It provides
information on how well the model identified the relationships and patterns in the training set. The MobileNetV2
obtained a good training and testing accuracy. The proposed model then learns the patterns in the training data and is
capable of successfully generalizing to fresh data.
350 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354

Training Accuracy Testing Accuracy

0.5

Accuracy(%) 0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8 9 10 11
No of Epochs
Figure 5: Training and Testing Accuracy of MobileNetV2

Table 2: List of Existing and Proposed Models to Detect the Tomato Plant Diseases

ACC Spc Re F1S Pre

VGG19 95.56 96.12 94.23 95.45 95.23
DoubleGAN Expansion 98.98 97.78 97.23 98.11 98.02
DRL-TL 99.23 98.89 98.98 99.12 99.23

Figure 6: Existing and Proposed Models to Detect the Tomato Plant Diseases

Table 3: Classification performance of VGG19 based on each type of disease

ACC Spc Re F1S Pre

BS 95.67 95.34 95.23 95.45 91.23

EB 93.45 94.34 94.45 94.37 94.38
 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354 351

Healthy 94.12 94.87 94.34 92.12 94.21

LB 94.78 94.23 94.54 93.23 94.23
LM 94.12 94.87 94.78 94.65 94.87
SLS 94.89 95.44 95.22 95.21 95.44
TS 95.11 93.56 93.11 93.8 93.56
TMV 94.12 95.34 95.21 95.66 95.67
TYLCV 95.23 94.89 94.31 94.89 94.46

TSSM 94.23 95.34 95.38 95.41 95.43

Figure 7: Classification performance of VGG19 based on each type of disease

Table 4: Classification performance of DoubleGAN expansion based on each type of disease

ACC Spc Re F1S Pre

BS 97.67 96.12 97.13 97.12 95.67

EB 97.45 97.23 97.45 97.37 97.38
Healthy 96.12 97.87 97.34 96.42 97.21
LB 95.78 97.23 97.54 96.23 97.23
LM 97.12 96.87 97.78 97.65 97.87
SLS 96.89 97.44 97.22 96.21 96.44
TS 95.98 96.56 97.11 97.18 97.56
TMV 95.23 97.34 97.21 97.66 97.67
TYLCV 96.67 96.89 97.31 96.9 96.46

TSSM 96.34 96.34 96.38 96.51 97.43

352 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354

Figure 8: Classification performance of DoubleGAN expansion based on each type of disease

Table 5 Classification performance of DRL-TL based on each type of disease

ACC Spc Re F1S Pre

BS 99.67 99.12 99.23 99.35 99.23

EB 99.45 99.54 99.45 99.39 99.38
Healthy 100 100 100 100 100
LB 99.78 99.23 98.54 98.23 99.23
LM 99.23 99.87 99.78 99.65 99.87
SLS 99.56 99.44 99.22 99.21 99.44
TS 99.34 99.56 93.41 98.8 99.56
TMV 99.87 99.34 95.31 99.66 99.67
TYLCV 99.67 99.89 94.37 99.89 99.46

TSSM 99.43 99.34 99.28 99.41 99.43

Figure 9: Classification performance of DRL-TL based on each type of disease

 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354 353

5. Conclusion and Future Work

In this paper, the Deep Reinforcement Learning with Transfer Learning (DRL-TL) for tomato plant disease
detection shows promising results. Through the integration of transfer learning methods with the capacity of deep
reinforcement learning algorithms, this strategy provides multiple benefits for precisely diagnosing and categorizing
tomato plant illnesses. By utilizing deep reinforcement learning models' ability to learn from interactions with the
environment, DRL-TL enables the system to enhance its performance gradually. This dynamic learning process
allows the model to adapt and adjust its detection capabilities based on the specific characteristics of tomato plant
diseases. Furthermore, transfer learning plays a crucial role in enhancing the efficiency and effectiveness of the
DRL-TL approach. The combination of DRL and transfer learning in tomato plant disease detection has
demonstrated promising results in terms of accuracy and efficiency. The model shows high accuracy in correctly
identifying and classifying various diseases affecting tomato plants, including early detection of symptoms that may
be challenging to detect with the naked eye. The implementation of DRL-TL in tomato plant disease detection also
offers practical advantages. It provides a non-invasive and cost-effective solution for farmers and plant pathologists,
allowing for early intervention and preventive measures to be taken, which can significantly reduce crop losses and
improve overall agricultural productivity. However, it is important to note that while DRL-TL shows promise, there
are still some challenges that need to be addressed. In future, a multi-model learning approach is to be developed by
using adaptive deep learning algorithms.

References

[1] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A. A. Bharath, “Generative adversarial networks: An overview,”
IEEE Signal Process. Mag., vol. 35, no. 1, pp. 53–65, Jan. 2018.
[2] L. C. Ngugi, M. Abdelwahab and M. Abo-Zahhad, "Tomato leaf segmentation algorithms for mobile phone applications using deep learning",
Comput. Electron. Agricult., vol. 178, Nov. 2020.
[3] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov and L.-C. Chen, "MobileNetV2: Inverted residuals and linear bottlenecks", Proc. IEEE/CVF
Conf. Comput. Vis. Pattern Recognit., pp. 4510-4520, Jun. 2018.
[4] A. Khattak, M. U. Asghar, U. Batool, M. Z. Asghar, H. Ullah, M. Al-Rakhami, et al., "Automatic detection of citrus fruit and leaves diseases
using deep neural network model", IEEE Access, vol. 9, pp. 112942-112954, 2021.
[5] F. Alvaro, Y. Sook, K. Sang and P. Dong, "A robust deep-learning-based detector for real-time tomato plant diseases and pests recognition",
Sensors, vol. 17, no. 9, pp. 2022, 2017.
[6] M. Arsenovic, M. Karanovic, S. Sladojevic, A. Anderla and D. Stefanovic, "Solving current limitations of deep learning based approaches for
plant disease detection", Symmetry, vol. 11, no. 7, pp. 939, Jul. 2019.
[7] Arunabha M. Roy et al., "A Deep Learning Enabled Multi-Class Plant Disease Detection Model Based on Computer Vision", 2021 AI 2, no.
3, pp. 413-428.
[8] Sk Mahmudul Hassan et al., "Identification of Plant Leaf Diseases Using CNN and Transfer Learning Approach 2021", Electronics, pp. 1-19,
2021.
[9] Shiva Azimi, Taranjit Kaur, Tapan K. Gandhi, "A deep learning approach to measure stress level in plants due to Nitrogen deficiency,"
Measurement, Volume 173, 2021.
[10] J. G. M. Esgario, R. A. Krohling, and J. A. Ventura, "Deep learning for classification and severity estimation of coffee leaf biotic stress,"
Comput. Electron. Agricult., vol. 169, Feb. 2020.
[11] Brahimi, M., Boukhalfa, K., and Moussaoui, A.: Deep Learning for Tomato Diseases: Classification and Symptoms Visualization. Appl.
Artif. Intell., 31 (2017) 299–315.
[12] I. Ahmad, M. Hamid, S. Yousaf, S. T. Shah, and M. O. Ahmad, "Optimizing pretrained convolutional neural networks for tomato leaf
disease detection," Complexity, vol. 2020, pp. 1–6, Sep. 2020, doi: 10.1155/2020/8812019.
[13] J.-H. Xu, M.-Y. Shao, Y.-C. Wang, and W.-T. Han, "Recognition of corn leaf spot and rust based on transfer learning with convolutional
neural network," Trans. Chin. Soc. Agricult. Mach., vol. 51, no. 2, pp. 230–236, Feb. 2020.
[14] X. Q. Guo, T. J. Fan, and X. Shu, "Tomato leaf diseases recognition based on improved multi-scale AlexNet," Trans. Chin. Soc. Agricult.
Eng., vol. 35, no. 13, pp. 162–169, Jul. 2019.
[15] Gao, Z.; Luo, Z.; Zhang, W.; Lv, Z.; Xu, Y. Deep Learning Application in Plant Stress Imaging: A Review. AgriEngineering 2020, 2, 430-
446. https://doi.org/10.3390/agriengineering2030029.
[16] Azimi, Shiva & Wadhawan, Rohan & Gandhi, Tapan. (2021). "Intelligent Monitoring of Stress Induced by Water Deficiency in Plants Using
Deep Learning." IEEE Transactions on Instrumentation and Measurement. PP. 1-1. 10.1109/TIM.2021.3111994.
[17] B. Liu, C. Tan, S. Li, J. He, and H. Wang, "A data augmentation method based on generative adversarial networks for grape leaf disease
identification," IEEE Access, vol. 8, pp. 102188–102198, 2020, doi: 10.1109/ACCESS.2020.2998839.
[18] Q. Zeng, X. Ma, B. Cheng, E. Zhou, and W. Pang, "GANs-based data augmentation for citrus disease severity detection using deep
learning," IEEE Access, vol. 8, pp. 172882–172891, 2020, doi: 10.1109/ACCESS.2020.3025196.
[19] E. Ozbılge, M. K. Ulukok, O. Toygar and E. Ozbılge, "Tomato Disease Recognition Using a Compact Convolutional Neural Network," in
IEEE Access, vol. 10, pp. 77213-77224, 2022, doi: 10.1109/ACCESS.2022.3192428.
354 Kadambari Raghuram et al. / Procedia Computer Science 252 (2025) 341–354

[20] K. Roy et al., "Detection of Tomato Leaf Diseases for Agro-Based Industries Using Novel PCA DeepNet," in IEEE Access, vol. 11, pp.
14983-15001, 2023, doi: 10.1109/ACCESS.2023.3244499.
[21] Y. Wu, X. Feng, and G. Chen, "Plant leaf diseases fine-grained categorization using convolutional neural networks," IEEE Access, vol. 10,
pp. 41087–41096, 2022.
[22] N. Ullah, J. A. Khan, L. A. Alharbi, A. Raza, W. Khan, and I. Ahmad, "An efficient approach for crops pests recognition and classification
based on novel DeepPestNet deep learning model," IEEE Access, vol. 10, pp. 73019–73032, 2022.
[23] S. Ahmed, M. B. Hasan, T. Ahmed, M. R. K. Sony, and M. H. Kabir, "Less is more: Lighter and faster deep neural architecture for tomato
leaf disease classification," IEEE Access, vol. 10, pp. 68868–68884, 2022.
[24] E. Elfatimi, R. Eryigit, and L. Elfatimi, "Beans leaf diseases classification using MobileNet models," IEEE Access, vol. 10, pp. 9471–9482,
2022.
[25] P. Jiang, Y. Chen, B. Liu, D. He, and C. Liang, "Real-time detection of apple leaf diseases using deep learning approach based on improved
convolutional neural networks," IEEE Access, vol. 7, pp. 59069–59080, 2019.
[26] C. Zhou, S. Zhou, J. Xing, and J. Song, "Tomato leaf disease identification by restructured deep residual dense network," IEEE Access, vol.
9, pp. 28822–28831, 2021.
[27] Q. Wu, Y. Chen, and J. Meng, "DCGAN-based data augmentation for tomato leaf disease identification," IEEE Access, vol. 8, pp. 9716–
9828, 2020.