2024 27th International Conference on Computer and Information Technology (ICCIT), 20-22 December 2024, Cox’s Bazar, Bangladesh

HTL-TLD: An Improved Hybrid Transfer Learning

Based Tomato Leaf Diseases Detection Model
Katura Khushbu Raihan Ul Islam Ahmed Wasif Reza
Computer Science and Engineering Computer Science and Engineering Computer Science and Engineering
East West University East West University East Wet\st University
Dhaka, Bangladesh Dhaka, Bangladesh Dhaka, Bangladesh
[email protected] [email protected] [email protected]

Shamim Ripon*
Computer Science and Engineering
East West University
Dhaka, Bangladesh
[email protected]

Abstract— Plants are crucial to society, the environment, and detection; however, limitations remain. Extensive feature
the economy but are susceptible to diseases that threaten engineering is required, and generalization across diverse
agricultural productivity and food security. Early detection of datasets may be problematic [5]. Pre-trained deep models, like
plant diseases is essential to minimize crop losses. Traditional EfficientNet and DenseNet, are used to extract quality features
methods are often time-consuming and error-prone, particularly from images, taking advantage of their capabilities of
in the early stages. Recently, Artificial Intelligence (AI) capturing complex patterns even on a limited amount of data.
techniques, particularly machine learning and transfer learning, The ML models then use these features to enhance the
have become pivotal for early leaf disease diagnosis. However, performance of classification. This decoupling lets HTL avoid
these techniques require further improvement. This paper
the issues related to overfitting while reducing the
proposes a Hybrid Transfer Learning (HTL) model that combines
Transfer Learning (TL) for feature extraction with Machine
computational complexity of the end-to-end deep learning
Learning (ML) for classification to enhance accuracy and models at the same time [6]. Furthermore, machine learning
efficiency. Using 18,160 images from the PlantVillage dataset, classifiers have improved interpretability and generalization,
covering nine tomato disease classes, various image enhancement especially in cases where deep learning models could struggle
techniques were applied to improve image quality and disease with imbalanced classes.
detection performance. The experimental results show that the This study will not only improve the accuracy of the
proposed method achieves a significant accuracy of 99.56%,
detection of diseases in tomato leaves but also reduce the
surpassing that of the current state-of-the-art methods. This
computational demands and improve the generalization on
approach offers a practical solution for farmers, enabling faster
and more accurate disease identification and better crop
imbalanced datasets. We propose the DenseNet201 + SVM
management. model, which leverages the advantages of DenseNet201 in
hierarchical feature reuse together with the high-dimensional
Keywords—plant disease, transfer learning, machine learning generalization capabilities of SVM. This model increases the
agricultural efficiency, tomato leaf classification, leaf health accuracy while optimizing computational efficiency. The
assurance, farming yield, food security design meets the requirements of variability in environmental
conditions and resource limitations, thus being appropriate for
I. INTRODUCTION agricultural applications. We used the PlantVillage dataset,
Agriculture is a vital part of the global economy, which containing ten classes of diseases of tomato leaves. The
contributes to rural development and food security. Tomatoes methodology commences with image preprocessing
are of critical importance worldwide, with a market value of techniques, including resizing, contrast adjustment, histogram
$230 billion and annual production of approximately 182 equalization, median filtering, and data augmentation
million metric tons [1]. They are rich in essential vitamins, are encompassing rotations, shifts, and flips to enhance image
extensively utilized in sauces, salads, soups, and juices, and quality and diversity. Pre-trained models, namely
play a crucial role in global nutrition. Tomato cultivation faces EfficientNetB3, XceptionNet, InceptionResNetV2, and
challenges, particularly leaf diseases, which significantly DenseNet201, were employed for feature extraction, and the
affect crop yields [2], production costs, and fruit quality. extracted features were subsequently input into an SVM
These diseases result in substantial economic losses, classifier for robust classification. The proposed hybrid
especially in major producing countries, such as Bangladesh, models demonstrated superior classification accuracy and
China, the United States, and Spain. Farmers, who typically computational efficiency compared to the other models
earn between $2,000 and $4,000 annually, often experience evaluated. This study makes several key contributions.
reduced earnings owing to disease management costs [3]. The • Developing an HTL framework that enhances the
prevalence of these diseases will always threaten production precision and generalization of classification, even
rates and, hence, the need for early detection to contain their when dealing with imbalanced datasets.
spread for crop health.
• Implementation of optimized image preprocessing and
Conventional disease detection is carried out through augmentation techniques to improve robustness and
manual diagnosis by experts and chemical testing. This is feature extraction under real-world conditions.
time-consuming, reliant on manual input, and prone to error
[4]. ML models have improved the performance of early

979-8-3315-1909-4/24/$31.00 ©2024 IEEE

 Fig. 1. Proposed system architecture

• Conveying the advantages of the suggested model in III. METHODOLOGY

relation to earlier research.
We began by preprocessing a diverse set of tomato leaf
The paper is structured as follows: Section 2 reviews samples, followed by the application of various Transfer
related work, Section 3 outlines methods, datasets, and Learning (TL) and Hybrid Transfer Learning (HTL)
architectural blueprints, Section 4 presents results including algorithms. Finally, we used different evaluation metrics to
model performance, Section 5 discusses broader implications, compare the proposed HTL model with conventional TL
and Section 6 concludes with reflections on limitations and models. An architectural diagram of the proposed model is
suggestions for future research. shown in Fig. 1. This shows the grayscale conversion used
experimentally to assess the texture-based features. However,
II. LITERATURE REVIEW the final model employs full-color (RGB) images to preserve
The PlantVillage dataset has been used extensively to the crucial color information necessary for detecting diseases
explore the CNN capabilities for detecting plant diseases. such as discoloration and pigmentation changes.
Agarwal et al. [7], for example, attained 91.2% accuracy in Preprocessing includes contrast adjustments and histogram
classifying nine tomato leaf diseases with a 3-layer CNN. equalization to highlight disease-related color features,
Nevertheless, the simplistic architecture of this model fails to ensuring the model effectively uses both texture and color
capture complex features accurately. cues for precise classification.
David et al. [8] proposed a hybrid CNN-RNN model to A. Data Description
identify ten tomato diseases by combining the sequential The publicly available PlantVillage [15] dataset from
learning of RNN with CNN's feature extraction, improving Kaggle contains 18,160 images of 10 tomato leaf disease
classification performance but facing challenges with classes, divided into training, validation, and test sets at 80%,
processing time due to increased computational overhead. Das 10%, and 10% ratios. Figure 2 shows sample images for each
et al. [9] used VGGNet's deep architecture to classify bell class. Bacterial Spot (BS) shows small black lesions with
pepper leaf disease images, which excels in extracting yellow halos from Xanthomonas spp., while Early Blight (EB)
intricate features. Abdullah et al. [10] selected YOLOV8s for presents concentric dark rings. Healthy Leaf (HL) is
real-time detection of 10 disease classes due to its superior uniformly green without distress. Late Blight (LB) has water-
speed and accuracy over YOLOV5 and Faster R-CNN. Billah soaked lesions due to Phytophthora infestans. Leaf Mold
et al. [11] utilized DenseNet121, achieving 98.62% accuracy (LM) appears as olive-brown mold underneath and yellow
on 6,000 disease data across 6 classes. patches on top. Septoria Lycopersici (SL) causes circular
Hybrid modelling has gained popularity in recent research. lesions with grey centers. Spider Mites (SM) produce white or
[12] developed a hybrid model combining EfficientNetB3 and yellow stippling. The target Spot (TS) had lesions with
MobileNet for tomato leaf disease identification using concentric rings from Corynespora cassiicola. The Mosaic
regularization, dropout, batch normalization, and Virus (MV) creates light-green and yellow mosaic patches,
hyperparameter tuning to enhance performance and prevent and the Yellow Leaf Curl (YLC) virus results in leaf
overfitting, achieving 99.92% accuracy, but may be unsuitable yellowing and curling.
for resource-limited environments because of its The distribution of the images for each class is shown in
computational demands. [13] proposed seven Bayesian- Fig. 3. YLC was the most common, representing 29% of the
optimized deep hybrid models for automated tomato leaf images, followed by BS at 12%, and LB at 11%. SLS
disease classification, employing CNN for feature extraction accounted for 10%, while HL and SM accounted for 9%. TS
and classifiers such as Random Forest and SVM, with a constitutes 8% of the dataset. EB and LM each accounted for
Boruta feature filtering layer, with the CNN-SVM model 5%, and MV was the least represented at only 2%.
achieving 98.268% accuracy. [14] used EfficientNet B7 with
Logistic Regression to classify grape leaf diseases, balancing
the model size and accuracy across four classes, although its
real-world agricultural performance remains untested.
Previous studies indicate high computational costs, long
processing times, and insufficient tuning in most models.
 manages fine-grained visual features, and detects localized
textural patterns in diseased leaves by independently handling
channel-wise spatial correlation.
We chose the InceptionResNetV2 model for its hybrid
architecture, integrating Inception's multiscale image
processing with ResNet's residual connections, enabling
deeper network training. This model excels in multiscale
feature extraction and avoids vanishing gradients, making it
ideal for detecting both macroscopic and microscopic disease
signs on tomato leaves. DenseNet201, with dense connections
Fig. 2. Sample images from each class where each layer receives inputs from all previous layers and
passes its feature maps forward, shortens pathways between
layers, improves information, and gradient flow. This feature
reuse aids in the detection of complex patterns in plant images.
D. Proposed Model: HTL models
This study combines baseline TL models with SVM to
efficiently meet the demands for advanced feature extraction
and classification. The architecture of each model is detailed
below, emphasizing the operational mechanics and integration
points.
(a) EfficientNetB3 + SVM:
The input image is preprocessed and resized to fit
EfficientNet-B3's dimensions. In Fig. 4, a 224 × 224 × 3 image
is passed through an initial convolution layer to extract basic
features, such as edges and textures. The image data then
moves through multiple Mobile Inverted Bottleneck
Convolution (MBConv) layers with 3x3 and 5x5 kernels,
capturing deeper, complex hierarchical features. These
varying kernel sizes enabled the model to capture fine and
coarse details. After the MBConv layers, a Conv (1x1) layer
reduces the spatial dimensions while preserving the depth
information and refining the features. The output is a flattened
Fig. 3. Class wise image distribution
feature vector that represents the essential features of the
image. This vector is fed into an SVM classifier, which is
B. Image Preprocessing known for its generalization ability, to classify the image.
We optimized the PlantVillage tomato disease dataset, Combining EfficientNet-B3 for feature extraction and SVM
comprising 18,060 images, for our experimental models for classification leverages deep learning's feature extraction
through several pre-processing steps. All the images were power and SVM's simplicity, enhancing the effectiveness.
resized to 224 × 224 pixels to ensure uniformity and model
compatibility. Techniques such as contrast correction and
histogram equalization have been employed to improve the
image quality and highlight disease features, which are crucial
for accurate classification. Additionally, median filtering was
applied to reduce noise and enhance image clarity while
preserving the edges and minimizing random noise [16].
To enhance the dataset diversity, we utilized various data
augmentation methods, including random rotations (±20 °),
shifts (up to 20%), shear transformations, zoom adjustments,
Fig. 4. EfficientNetB3 + SVM Architecture
and horizontal flipping (50% probability). These
augmentations were applied dynamically during training to (b) Xception + SVM:
increase the variability and prevent overfitting. Image pixel Xception's depthwise separable convolutional architecture
values were standardized and rescaled between 0 and 1 to minimizes computational redundancy while effectively
maintain consistent brightness and contrast, thus facilitating extracting features through SeparableConv2D layers, which
faster convergence during training. decompose convolution into spatial and depthwise convolutions
C. Baseline Transfer Learning Models for enhanced efficiency and abstraction (Fig. 5). The extracted
Four tailored transfer learning models, EfficientNetB3, features are flattened and input into an SVM classifier, enabling
XceptionNet, InceptionResNetV2, and DenseNet201, served efficient classification in high-dimensional spaces.
as baselines. EfficientNetB3, chosen for its scalable
architecture to balance model depth, width, and resolution,
was trained using augmentation techniques to prevent
overfitting. This efficient model excels with small datasets,
 Fig. 5. Xception + SVM Architecture
Fig. 7. DenseNet201 + SVM Architecture
(c) InceptionResNetV2+SVM:
This model integrates the ability of the inception E. Training Parameters
architecture to process multiple visual feature sizes with The HTL and TL models were implemented on an
ResNet residual connections to prevent gradient vanishing. NVIDIA RTX 3060 Twin Edge GPU with 3584 CUDA cores,
The design is effective in classifying images of different 12GB GDDR6 VRAM, and 64GB system RAM. Data
diseases while ensuring comprehensive feature extraction at all augmentation techniques have been used to enhance class
levels (Fig. 6). The model uses inception modules to capture diversity owing to varying image counts per class. The
multi-scale features and residual connections to preserve hyperparameters listed in Table III were optimized for
feature integrity through different layers: the Stem, Inception- accuracy. Each model featured eight convolutional layers and
ResNet-A, Reduction-A, Inception-ResNet-B, Reduction-B, a max-pooling layer for feature extraction and dimensionality
and Inception-ResNet-C blocks. These layers enhance the reduction. Dropout rates of 0.5 and 0.2, helped prevent
features and reduce the spatial dimensions, producing a feature overfitting while preserving key features. The ReLU
vector that represents the core attributes of the image. The activation function supports complex pattern recognition.
features were then combined by an SVM classifier with most Categorical cross-entropy is a loss function for multiclass
of the improvements of regularization, enhancing the ability to classification. The Adam optimizer, with a learning rate of
generalize to new, unseen data. 0.0001, was used for stable training and consistent
performance.

TABLE I. PERFORMANCE OF INDIVIDUAL AND HYBRID TL

CLASSIFICATION ALGORITHMS.
Hyperparameter Description
No. of Convolution 8
No. of Max Pooling 1
Dropout Rate Dropout rates of 0.5 and 0.2 are used
Activation Function ReLU
Fig. 6. InceptionResNetV2 + SVM Architecture
Learning Rate 0.0001
(d) DenseNet201 + SVM:
F. Evaluation
The DenseNet201 + SVM architecture integrates
DenseNet201's hierarchical learning and SVM's high- Several criteria are used to assess a multi-class model for
dimensional classification (Fig. 7) to ensure superior disease tomato leaf disease classification. Precision (Equation 2)
detection under various conditions. Robust preprocessing, measures the accuracy within each disease class, while
including noise reduction and augmentation, enhances accuracy (Equation 1) indicates the model's overall
adaptability to real-world agricultural scenarios and reduces categorization ability. Recall (Equation 3) evaluates the
the computational overhead. High connectivity ensures that capacity of the model to identify cases in specific disease
each layer directly contributes to the final classification, categories. The F1-score (Equation 4), which combines
improves information flow, and reduces loss. The design Precision and Recall, provides a comprehensive view of the
includes a convolutional layer followed by max pooling to performance of the model across multiple tomato leaf
reduce the spatial dimensions while preserving critical diseases.
features (Fig. 7). The architecture comprises several Dense (1)
Blocks and Transition Layers arranged in sequences of 6, 12, Accuracy =
+ +
and 48 blocks, thereby enhancing the network's ability to learn (2)
Precision =
complex features through depth concatenation and effective +
feature propagation. An additional Dense Block further (3)
Recall =
refines features for elevated abstraction. An SVM classifier +
2 × (Precision × Recall) (4)
ensures an accurate and reliable disease classification and F1 − score =
Precision + Recall
detailed feature interactions.
IV. RESULT ANALYSIS
A. Performance of the Classifiers
Table II shows the performance metrics of the TL and
hybrid models, with the latter performing better across all
individual TL models. The DenseNet201+SVM model
achieved an accuracy of 0.9955, higher than the performance
achieved by the single DenseNet201, which was 0.9344. In
addition, this model performed very well in terms of precision,
recall, and F1 score, each at 0.9955, indicating almost perfect
classification. Also, in the Xception + SVM case, accuracy
increased from 0.8249 to 0.9945. InceptionResNetV2+SVM
has an increase in accuracy from 0.8392 to 0.9834 and
EfficientNetB3+SVM hybrid increased it from 0.9361 to
0.9801. Besides that, all these models also present increased
precision, recall, and F1 score, making the proposed hybrids
Fig. 9. Learning curve of highest performing model.
even more effective than their counterparts

TABLE II. RESULTS OF EXPERIMENTAL CLASSIFIERS.

To mitigate overfitting and ensure high classification
accuracy, we employed robust model validation
Model Accuracy Precision Recall
F1 techniques. The closely aligned training and validation
score accuracy curves in Fig. 9 suggest effective learning without
EfficientNetB3 0.9361 0.9317 0.9361 0.9351 overfitting. Over 20 epochs, accuracy increased while loss
DenseNet201 0.9344 0.93 0.9345 0.9338
TL

decreased, with validation loss plateauing at a low value,

InceptionResNetV2 0.8392 0.8814 0.8392 0.8369
Xception 0.8249 0.8428 0.8249 0.8068
indicating good generalization. Data augmentation
DenseNet201 + SVM 0.9955 0.9942 0.9955 0.9955 enhanced training set diversity, reducing overfitting risk.
Xception+SVM 0.9945 0.9944 0.9945 0.9944 Ablation studies (Table III) confirmed individual
HTL

InceptionResNetV2 + component contributions: removing SVM and excluding

0.9834 0.9845 0.9834 0.9836 augmentation reduced accuracy from 99.56% to 93.44%
SVM
EfficientNetB3+SVM 0.9801 0.9815 0.9802 0.98 and 91.72%, respectively, affirming the model's robustness,
generalization ability, and reliability..
B. Result Validation
The confusion matrix of DenseNet201+SVM in Fig. 8 TABLE III. SAMPLE ABLATION STUDY RESULT
indicates high accuracy for most categories. The BS had 223 Model Accurace
correct predictions with no misclassifications, showcasing the
model accuracy for this disease. However, some diseases DenseNet201 + SVM 0.9955
show notable misclassification. The EB and LB were DenseNet201 Only 0.9344
HTL

misclassified twice. Similarly, Tomato Healthy had 36

Without augmentation 0.9172
instances misclassified as LCV, indicating a significant
misclassification rate in the healthy category. Grayscale Images only 0.9801

C. State-of-the-art Comparison
The proposed DenseNet201 + SVM model achieved an
impressive accuracy of 99.56%, outperforming MobileNetV2
(92.5%) and VGGNet (98%). Its hybrid architecture strikes a
balance between computational efficiency and accuracy,
setting it apart from existing approaches. Advanced
preprocessing techniques bolster their robustness against
noise and environmental variations, making them highly
suitable for practical agricultural applications. Accuracy
compared with existing methods, as shown in Table IV. Our
approach achieved an impressive accuracy of 99.56% across
10 disease classes, utilizing the DenseNet201+SVM model
and a dataset comprising 18,160 samples. Our model exhibits
strong alignment with large datasets, as in [7],[10], and [14].
Fig. 8. Confusion matrix of the highest-performing classifier
The model demonstrated the highest recorded in the
comparison, despite the substantial volume of data involved.
The learning curve in Fig. 9 illustrates the progress of the The proposed model outperformed other models such as
model over 20 epochs, with consistent accuracy MobileNetV2 [10] with an accuracy of 92.5%, and VGGNet
improvements and a steady reduction in loss. Both the training [9] with an accuracy of 98% and was limited to two classes,
and validation loss curves decreased uniformly, with the along with the CNN-RNN [8] model, with only 2000 data. In
validation loss stabilizing, indicating effective learning and addition, it surpassed DenseNet121 in [11], which collected
generalization. The accuracy curves for training and six classes of disease, with an accuracy of 98.62%. The
validation rose sharply, converging near maximum values by enhanced performance of our model was attributed to the
the end, all exceeding 99%, signifying strong model systematic combination of DenseNet201 and SVM. All these
performance. The close convergence of these curves suggests studies utilized the Plant Village tomato disease dataset. Our
good generalization without overfitting. approach not only surpasses these methods in terms of
accuracy, but also demonstrates superior performance with a
larger dataset, emphasizing its effectiveness in classifying and conditions. The HTL model, improved further, thus is
tomato leaf diseases. another leap in AI-driven agrarian solutions that support
sustainable farming for global food security. Efforts will be in
T ABL E I V. COMPARISON WITH THE STATE-OF-THE-ART improving computational efficiency and its application to a
Data
greater number of crops and conditions.
Reference Model Classes Accuracy Year
instances
REFERENCES
[7] 17,500 CNN 10 91.2% 2000
CNN-RNN [1] C. C. Patel, N. M. Gohel, C. B. Dhobi, A. K. Maru, R. G. Parmar, and
[8] 2000 10 98.25% 2022 D. M. Korat, “Plant Health Management”.
model
[2] K. B. Vegesana, “A Framework and System for a Multi-Model
EfficientNet Decision Aid for Sustainable Farming Practices,” Computational
[14] 9027 4 98.7% 2022
B7 +LR Modeling & Simulation Engineering Theses & Dissertations, Jan.
[9] 2,475 VGGNet 2 98% 2023 2015, doi: 10.25777/n7kx-f612.
[10] 17,500 MobileNetV2 10 92.5% 2024 [3] E. C. D. Todd, “Economic Loss from Foodborne Disease and Non-
Illness Related Recalls Because of Mishandling by Food Processors,”
[11] 6.000 DenseNet121 6 98.62% 2024 J Food Prot, vol. 48, no. 7, pp. 621–633, Jul. 1985, doi: 10.4315/0362-
Proposed DenseNet201 028X-48.7.621.
18,160 10 99.56% 2024
Method + SVM [4] K. Weiss, T. M. Khoshgoftaar, and D. D. Wang, “A survey of transfer
learning,” J Big Data, vol. 3, no. 1, pp. 1–40, Dec. 2016, doi:
V. DISCUSSION 10.1186/S40537-016-0043-6/TABLES/6.
[5] V. Bolón-Canedo, N. Sánchez-Maroño, and A. Alonso-Betanzos,
Our study shows that combining transfer learning (TL) “Recent advances and emerging challenges of feature selection in the
with machine learning (ML) outperforms TL models alone. context of big data,” Knowl Based Syst, vol. 86, pp. 33–45, Sep. 2015,
doi: 10.1016/J.KNOSYS.2015.05.014.
The SVM approach effectively addresses overfitting in high- [6] M. R. Ahmed et al., “Towards Automated Detection of Tomato Leaf
dimensional spaces, thereby enhancing the accuracy and Diseases,” Proceedings - 6th International Conference on Electrical
generalization of various diseases. The DenseNet201 + SVM Engineering and Information and Communication Technology,
model distinguishes overlapping disease symptoms with ICEEICT 2024, pp. 387–392, 2024, doi:
significant implications for agriculture by enabling timely 10.1109/ICEEICT62016.2024.10534559.
[7] M. Agarwal, A. Singh, S. Arjaria, A. Sinha, and S. Gupta, “ToLeD:
actions to reduce crop losses and costs, thereby enhancing Tomato Leaf Disease Detection using Convolution Neural Network,”
food security and sustainability. Their versatility across Procedia Comput Sci, vol. 167, pp. 293–301, 2020, doi:
different crops and diseases makes them valuable for smart 10.1016/J.PROCS.2020.03.225.
farming and automated disease control. However, this study [8] H. E. David, K. Ramalakshmi, R. Venkatesan, and G. Hemalatha,
had several limitations. Relying on the PlantVillage dataset “Tomato leaf disease detection using hybrid CNN-RNN model,”
Advances in Parallel Computing, vol. 38, pp. 593–597, 2021, doi:
may not fully represent the diverse range of global tomato leaf 10.3233/APC210108.
diseases, considering environmental and genetic variations. [9] P. Kumar Das, “Leaf Disease Classification in Bell Pepper Plant using
The controlled conditions of the dataset might not accurately VGGNet,” Journal of Innovative Image Processing, vol. 5, no. 1, pp.
reflect real-life factors such as noise, lighting, and background 36–46, Mar. 2023, doi: 10.36548/JIIP.2023.1.003.
[10] A. Abdullah, G. A. Amran, S. M. A. Tahmid, A. Alabrah, A. A. AL-
differences, potentially impacting real-world effectiveness. Bakhrani, and A. Ali, “A Deep-Learning-Based Model for the
Additionally, we did not assess the adaptability of the model Detection of Diseased Tomato Leaves,” Agronomy 2024, Vol. 14,
to various weather conditions, soil types, or farming methods. Page 1593, vol. 14, no. 7, p. 1593, Jul. 2024, doi:
Using DenseNet201 with an SVM could be resource- 10.3390/AGRONOMY14071593.
intensive, limiting its widespread adoption. Future work could [11] M. M. Billah, A. Sultana, R. Sad Aftab, M. M. Ahmed, and M. Shorif
Uddin, “Leaf disease detection using convolutional neural networks:
involve incorporating more crops and diseases with hybrid a proposed model using tomato plant leaves,” Neural Comput Appl,
transfer learning, exploring enhancements, such as model pp. 1–11, Aug. 2024, doi: 10.1007/S00521-024-10283-2/METRICS.
quantization and low-rank matrix decomposition, and using [12] Z. Ullah, N. Alsubaie, M. Jamjoom, S. H. Alajmani, and F. Saleem,
edge processing or cloud computing for real-time disease “EffiMob-Net: A Deep Learning-Based Hybrid Model for Detection
diagnosis in limited-capacity environments. and Identification of Tomato Diseases Using Leaf Images,”
Agriculture (Switzerland), vol. 13, no. 3, p. 737, Mar. 2023, doi:
10.3390/AGRICULTURE13030737/S1.
VI. CONCLUSIONS [13] B. Khan et al., “Bayesian optimized multimodal deep hybrid learning
This work presents a hybrid transfer learning approach for approach for tomato leaf disease classification,” Scientific Reports
enhancing the accuracy in detecting diseases of tomato leaves. 2024 14:1, vol. 14, no. 1, pp. 1–30, Sep. 2024, doi: 10.1038/s41598-
024-72237-x.
The DenseNet201 + Support Vector Machine model showed [14] P. Kaur et al., “Recognition of Leaf Disease Using Hybrid
the best performance with an accuracy of 99.56%, which is Convolutional Neural Network by Applying Feature Reduction,”
scalable for practical applications. It outperforms other Sensors 2022, Vol. 22, Page 575, vol. 22, no. 2, p. 575, Jan. 2022, doi:
traditional transfer-learning models on efficient handling of 10.3390/S22020575.
[15] “PlantVillage-Dataset/raw/color at master · spMohanty/PlantVillage-
high-dimensional feature space and reducing overfitting in Dataset · GitHub.” Accessed: Sep. 01, 2024. [Online]. Available:
imbalanced datasets. The model generalizes very well across https://github.com/spMohanty/PlantVillage-
a wide range of disease categories with the help of advanced Dataset/tree/master/raw/color
image preprocessing and data augmentation techniques, hence [16] B. R. Jana, H. Thotakura, A. Baliyan, M. Sankararao, R. G.
making it feasible for early agricultural disease diagnosis. The Deshmukh, and S. R. Karanam, “Pixel density based trimmed median
filter for removal of noise from surface image,” Applied Nanoscience
PlantVillage dataset limits its applicability in diverse (Switzerland), vol. 13, no. 2, pp. 1017–1028, Feb. 2023, doi:
environments. Future research should use diverse datasets, 10.1007/S13204-021-01950-0/METRICS
optimize model efficiency, and extend its use to other crops