IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 4, December 2024, pp. 4757~4765

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i4.pp4757-4765  4757

Enhanced detection of tomato leaf diseases using ensemble deep

learning: INCVX-NET model

Shruthi Kikkeri Subramanya, Naveen Bettahalli

Department of Electronics and Communication, BGS Institute of Technology, Adichunchanagiri University, Karnataka, India

Article Info ABSTRACT

Article history: Automated leaf disease detection quickly identifies early symptoms, and saves
time on large farms. Traditional methods like visual inspection and laboratory
Received Feb 10, 2024 detection are prevalent despite being labor-intensive, time-consuming, and
Revised May 11, 2024 susceptible to human error. Recently, deep learning (DL) has emerged as a
Accepted Jun 21, 2024 promising alternative for crop disease recognition. However, these models
usually demand extensive training data and face problems in generalization
due to the diverse features among different crop diseases. This complexity
Keywords: makes it difficult to achieve optimal recognition performance across all
scenarios. To solve this issue, a novel ensemble approach INCVX-Net is
Convolutional neural network proposed to integrate the three DL models, ‘Inception, visual geometry group
Hybrid deep learning (VGG)-16, and Xception’ using weighted averaging ensemble for tomato crop
Inception V2 model leaf disease detection. This approach utilizes the strengths of three DL models
Tomato plant disease to recognize a wide range of disease patterns and captures even slight changes
Visual geometry group-16 in leaf characteristics. INCVX-Net achieves an impressive 99.5% accuracy in
model disease detection, outperforming base models such as InceptionV2 (93.4%),
Xception model VGG-16 Net (92.7%), and Xception (95.2%). This significant leap in
accuracy demonstrates the growing power of ensemble DL models in disease
detection compared to standalone DL models. The research paves the
groundwork for future advancements in disease detection, enhancing
precision agriculture through ensemble models.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Shruthi Kikkeri Subramanya
Department of Electronics and Communication, BGS Institute of Technology, Adichunchanagiri University
Bengaluru–NH-75, Nagamangala Taluk, Mandya District, B.G Nagara, Karnataka 571448, India
Email: [email protected]

1. INTRODUCTION
Agriculture is the essence of human survival: it feeds and offers a living for billions of people across
the globe [1], [2]. However, the sector is associated with several threats, including adverse weather, pest
attacks, and the complex spread of plant diseases [3]. Yet, the farmers are more concerned about the potential
impact of diseases affecting their fields [4], particularly due to their limited knowledge regarding the symptoms
of the new diseases and effective preventive measures for any such disease, remains a worrying possibility [5].
A practical approach is to employ early identification of the diseases which can help to prevent substantial
losses.
When the crops are affected by pests or diseases, farmers typically attempt to identify the disease
themselves or seek help from experts [6] and then turn to laboratory methods [7], which can result in delayed
or incorrect disease diagnosis. For example, early blight, a tomato disease, is characterized by circular dark
spots on the leaves. It was often mistaken for other leaf spot diseases. One of the wide spread fungal infections
is septoria leaf spot, which looks like small dark spots with yellow halos surrounding them. It confounds experts
because it mimics the symptoms of other distinct fungal infections. In laboratory methods, sample processing

Journal homepage: http://ijai.iaescore.com

4758  ISSN: 2252-8938

and analysis by the experts results in delayed detection. These challenges become more burdensome if one
looks at the rural and remote areas where resources in terms of expert advice and advanced diagnostic tools are
hardly available. This is where the technological advancements come into the picture to give a ray of hope in
the form of a solution to the most pressing issue.
The development in information technology made the researchers employ machine learning (ML) and
deep learning (DL) techniques for the efficient detection of plant diseases [8], [9]. DL, particularly has gained
popularity and brought new highs to automatic feature extraction from images and new-found efficiency, joined
with a high accuracy in identifying diseases [10]. The various DL models such as AlexNet, visual geometry
group (VGG)Net, residual neural network (ResNet), Inception Net, custom convolutional neural network
(CNN), and MobileNet are proven to provide a better performance provided a large amount of data containing
information about the diseases [11]. However, in practical situations, the appearance of new diseases presents
a significant challenge to the applicability of DL models particularly regarding unknown images. These models
are typically trained on pre-existing datasets of leaf diseases [12], making them familiar with known conditions.
But when encountering new diseases, the DL models may struggle to identify them correctly. This demands
the need for a new approach that can handle evolving diseases and able to provide robust performance in real
world applications.
The ensemble technique [13], [14] can mitigate the challenge associated with the emergence of new
diseases in plant health management. This approach integrates DL models, each trained on various datasets,
focusing on a particular area of characteristics. For example, when a DL model in the ensemble approach
concentrates on fungal diseases whereas another specializes in detecting viral infections. In that case, the
combined knowledge of multiple models can improve the overall disease recognition performance. In the present
work, the INCVX-Net ensemble learning model for disease detection in tomato plants is proposed, which
integrates the three DL models and is able to enhance the generalization ability of the classification method.
Before considering the ensemble approach, it’s necessary to know the outcomes for plant disease detection
based on standalone DL models [15]–[17] and to understand the effectiveness of the ensemble method [18]–[20].
Trivedi et al. [15] suggested CNN to distinguish ten different classes of tomato leaves with a dataset containing
3,000 images. The proposed model with tuned hyperparameters showed 98.49% accuracy. However, achieving
this level of accuracy underscores the challenges of collecting diverse data. Ahmad et al. [16] use both laboratory
and field-based datasets ,employing four CNN architectures-VGG16, VGG19, Inception v3, and ResNet for tomato
leaf disease classification. Despite promising results on the laboratory dataset, with an accuracy variation of 10 to
15%, Inception V3 is best for both datasets. Ferentinos [17] considered VGG, AlexNet, AlexNetOWTBn,
GoogleNet, and Overfeat, to conduct plant disease detection on the open database including 87,848 photos of both
healthy leaves and diseased plants. Among these, VGG produced a success rate of 99.53%.
Arshad et al. [18] propose a PLDPNet, a novel hybrid DL model, for automatic segmentation and
classification of potato (3 classes), tomato (10 classes), and apple (4 classes) leaf diseases. The PLDPNet uses
auto-segmentation and deep feature ensemble fusion modules to increase disease classification accuracy. The
end-to-end performance of the proposed PLDPNet achieves a recorded accuracy of 98.66% with 100 epochs.
Ganaie et al. [19] employs a novel TomConv model, a modified CNN, a classification system among the ten
different kinds of tomato plant leaves. In this study 105 epochs of training were applied to the publicly available
dataset Plant Village, which contains over 16,000 pictures of tomato leaves, and an accuracy of 98.19% was
attained. By utilizing 3,602,506 total parameters, the suggested model performs greater than current state-of-
the-art models. He et al. [20] introduced a novel ensembling approach ensemble learning method for crop leaf
disease recognition (ELCDR), for disease detection in 4 different crops, which assigns weights to models based
on their feature vectors. Experimental results demonstrate that ELCDR outperforms single model recognition
and traditional voting methods, showing 98.13% apple, 95.88% corn, 98.38% grape, and 90.75% rice accuracy
which validates the effectiveness of the ensembling approach in crop leaf disease recognition.
The researchers have used various well-built pre-trained DL models to recognize the diseases and get
noteworthy performance.They must generalize easily to new or unseen data/features [21]. Later, in the mid-2000s,
Research slowly considered integrating the multiple DL model, thus leading to the effective ensemble approach.
So, the present work, emphasizes ensemble learning, by exploring efficient ensemble strategies that perform well
in plant disease detection areas.The contributions of this paper involve the following core components: i) proposed
an ensemble approach for tomato crop disease detection named INCVX-Net; ii) three different DL models are
subjected to transfer learning, wherein the bottom layers are frozen, and each model’s performance is individually
recorded; iii) incorporated weighted averaging ensemble strategy and compared the outcomes with the crop
disease detection methods that are based on a single DL model; iv) the experiments consist of a tomato dataset
comprising five classes, conducted on a Google Colab platform. The results showed the necessity of considering
the ensemble approach. The remainder of the paper is structured as follows: section 2 discusses the methodology,
section 3 presents the results and subsequent discussion, and section 4 encapsulates the concluding remarks.

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4757-4765

Int J Artif Intell ISSN: 2252-8938  4759

2. METHOD
The research aims to improve the prediction and recognition of plant leaf disease efficiently by
concatenating the features of three DL models, namely the Inception model, VGG-16 Net model, and Xception
model. The basic flow of INCVX-Net comprises four stages as listed as follows and depicted in Figure 1. There
are: i) dataset collection, ii) extraction of features from the base models, iii) predictions and weight averaging,
and iv) validation of the model.

Figure 1. The overview of INCVX-Net model

2.1. Dataset collection

To achieve the research aims, tomato leaf images obtained from the plant village dataset is employed.
The plant village dataset contains large scale high-resolution images that have been captured from unmanned
aerial vehicle (UAV) [21]. The present dataset comprises 5,250 images from five different classes such as
bacterial spot, healthy leaf, late blight, spot leafe, and yellow leaf curl virus. The images of size 256×256 pixels
which are split as training, validation, and testing images as displayed in Table 1. The visualization of the
sample images of the tomato leaf is shown in Figure 2.

Table 1. Dataset information of the tomato leaf

S. No. Leaf type Total no. of images Training images Validation images Testing images
1. Bacterial spot 1,050 900 50 100
2. Healthy leaf 1,050 900 50 100
3. Late blight 1,050 900 50 100
4. Spot leafe 1,050 900 50 100
5. Yellow leaf curl virus 1,050 900 50 100

Figure 2. Dataset description of tomato leaf dataset

Later in the process, the tomato leaf images are subjected to pre-processing step. Initially, all images
are resized to a standard format of 224×224×3 pixels using open CV library in Python. Additionally, to enhance
the models performance, the pixel values of the images are normalized to a range of 0 to 1. The resizing
provides uniformity in image dimensions, whereas normalization minimize the effects of varying pixel
intensities.

2.2. Extraction of the features from base models

The feature extraction process begins with selecting the base models from a set of pre-trained DL
models. Inceptionv2, VGG-16, and Xception are chosen as a base model which is initially trained on large
scale datasets such as ImageNet and comes with pre-trained weights. The models are subjected to transfer
Enhanced detection of tomato leaf diseases using ensemble deep learning … (Shruthi Kikkeri Subramanya)
4760  ISSN: 2252-8938

learning, which involves adjusting them by incorporating five output units into the final output layer [22], [23].
The pre-trained layers of the base models are frozen to a certain point by excluding the final classification
layers, allowing only the newly added layers to be trained on the tomato leaf dataset by fine tuning the
hyperparameters. The overview of transfer learning is shown in Figure 3. The hyperparameters settings of all
three models are shown in Table 2. Training base models undergo a feature extraction process tailored to their
architecture. Once the features are extracted, they serve as input to the prediction phase.

Figure 3. The overview of transfer learning

Table 2. Hyperparameters settings

Hyperparameters Value
Epochs 8
Batch_ size 64
Activation_ function Relu
Optimization algorithm Adam
Loss _function Categorical Loss Entropy
Learning rate 0.00001

In Inception V2, features are extracted from intermediate Convolutional layers. These layers capture
both high level and low level visual patterns in the input data, providing a rich representation at various levels
of abstraction. The VGG-16 features are obtained from the second to last fully connected layer or the global
average pooling layer. These layers captures complex patterns and structures while abstracting away specific
pixel level details. On the other hand, in Xception, features are extracted from intermediate depth wise
separable Convolutional layers. These layers enable the model to learn rich representations with various levels
of abstraction. The distribution of the extracted features is analyzed using gradient-weighted class activation
mapping (Grad-Cam) [24] to ensure they are rich in all possible features, if not, the models should be retrained
to capture a more comprehensive set of features. Thus, extracting features from pre-trained DL models
harnesses the power of learned representations to capture essential characteristics of the input data. This allows
us to reduce the need for extensive labeled data and also enables us to leverage the knowledge encoded in these
models from their training on large-scale datasets like ImageNet.

2.3. Predictions and weight averaging

The extracted features are passed through a final classification layer of the models with a softmax
activation function. The final layer generates predictions in the form of class probabilities for each input sample.
Weights are assigned to each model’s prediction, to reflect the importance placed on each models predictions.
For each class C, a weighted average of the predicted probabilities from all models is calculated using (1).

Fc= ∑𝑛𝑖=1 𝑊𝑖*Pi,c (1)

Where, Fc is final probability for each class C, Wi is weight assigned to each model I, Pi,c is predicted probability
of class C by model I. Three DL models may extract different features from the same dataset. The weights are
assigned [25] to the predictions depending on factors such as model performance on different classes of data.
The name INCVX-Net is coined to indicate the combined prediction of the Inception, VGG16, and Xception
models.

2.4. Validation of the model

The performance of the proposed model is evaluated by using various performance metrics such as
precision, recall, accuracy and F1-measure [26]. The confusion matrix is a two dimensional table, as shown in
Figure 4, and is used to calculate the above mentioned metrics. In this matrix actual values are on the column
side and predicted values are on the row side. Let true positive (TP), true negative (TN), false positive (FP), and
false negative (FN) denote the number of TP, TN, FP, and FN respectively. The TP is an outcome where the
models correctly predict the positive class. The TN is an outcome where the models correctly predict the negative
class. The FP is an outcome where the models incorrectly predict the positive class. The FN is an outcome where
the models incorrectly predict the negative class [27]. The metrics are calculated using (2) to (5).

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4757-4765

Int J Artif Intell ISSN: 2252-8938  4761

Accuracy = (TP + TN)/(TP + TN + FP + FN) (2)

Precision = TP/(TP + FP) (3)

Recall = TP/(TP + FN) (4)

F1 − Measure = 2/(1/precision + 1/Recall) (5)

Figure 4. Confusion matrix

3. RESULTS AND DISCUSSION

The work employed three DL models: inception, VGG-16 Net, and Xception model individually, to
recognize various tomato leaf diseases. The experiments were conducted on a Python 3 Google Compute
Engine back end (GPU) using a Jupyter Notebook and various DL libraries including OpenCV, NumPy,
Matplotlib, scikit-learn, Keras, and TensorFlow. Google Colab, running on an Intel (R) Core (TM) i7-8700
CPU @ 3.20 GHz with 8.00 GB of RAM, was utilized for execution. The training accuracy and loss of the
base models Inception, VGG16, and Xception are shown in Figures 5(a) to 5(c) respectively. As the number
of epochs increases, there is a gradual reduction in loss and a corresponding rise in accuracy.

(a) (b) (c)

Figure 5. Classification accuracy and loss of (a) InceptionV2, (b) VGG-16, and (c) Xception models

The confusion matrix was utilized to calculate various performance metrics. Figures 6 and 7 show the
confusion matrix of the base models and the proposed INCVX-NET model. In InceptiontV2, VGG-16, and
Xception models some of the classes are misclassified. For example, in the confusion matrix of the VGG-16
model, late blight is misclassified as healthy leaf and spot leaf. Hence, to improve the tomato leaf disease
prediction further the three models namely InceptionV2, VGG-16, and Xception model are combined at the
decision level by weighted averaging method. After the integration, the resulted model INCVX-Net able to
identify more correct images than the base models as shown in Figure 7. The accuracy of the ensemble approach
reached 99.5%, with 99.1% precision, 99.4% recall, and 99.25% F1-score. Figure 8 shows the performance
comparison of the ensemble model with three base models. The findings of this study are compared with those
of previous works carried out within the same application domain, as detailed in Tables 3 and 4.

Enhanced detection of tomato leaf diseases using ensemble deep learning … (Shruthi Kikkeri Subramanya)
4762  ISSN: 2252-8938

Figure 6. Confusion matrix of InceptionV2, VGG-16 net and xception model

Figure 7. Confusion matrix of INCVX-Net model

Figure 8. Performance comparison of INCVX-Net with the base model

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4757-4765

Int J Artif Intell ISSN: 2252-8938  4763

Table 3. Performance comparison of proposed model with the base models

Model Accuracy (%) Precision (%) Recall (%) F1-Score (%)
InceptionResnetV2 Model 93.4 92.6 93.1 92.75
VGG-16 Net Model 92.7 91.25 92.4 91.7
Xception Model 95.2 94.7 95.1 94.8
Proposed INCVX-Net Model 99.5 99.1 99.4 99.25

Table 4.performance comparison of INCVX-Net with previous works

Ref. Model Dataset Epochs Accuracy (%) Precision (%) Recall (%) F1-Score (%)
[20] EDCLR 8,000, multi- 60 Apple: 98.13 98.14 98.13 98.13
crop Corn: 95.88 96 95.88 95.94
Grape: 98.3 98.39 98.38 98.38
Rice: 90.75 91.86 90.75 91.30
[28] Hybrid CNN 9,027, Grapes 92 98.7 98 99 94
[29] SE-ResNet50 22,925 tomato 150 96.81 96.77 96.81 96.79
[30] Resnet50-CBAM+SVM - 50 97.2 --- --- ---
[31] Attention-dilated CNN and 15,989 100 100 100 100 100
Logistic regression Tomato
[32] ResNet50+MobileNet 5,400 Olive 100 97.08 97.61 97.11 96.86
[33] EfficientNetB3+MobileNet 32,535, 20 99.92 - - -
Tomato
[34] Ensemble hydra model 15,600, 50 95 95 94 93
(CNN+VGG19) Cotton
Proposed INCVX-Net
99.1 99.4 99.25
Model 5,250, Tomato 8 99.5

3.1. Discussion
This study investigated the effectiveness of an ensembling approach on disease recognition of plants.
While most of the previous studies have examined the impact of single DL models on well-established datasets,
they still need to address how these models work when encountering a limited dataset, and their adaptability
to new images which are common in real world applications remains under explored. Stand-alone DL models
operate independently, yet there’s limited investigation into the potential benefits of integrating multiple DL
models.
The proposed INCVX-Net model, is an integration of the three base models Inception V2, VGG-16,
and Xception to classify the diseases in tomato plants with a limited dataset. The study employed the feature
vectors extracted by the base models, since in the context of disease detection in plants using DL models, there
is a strong relation between the features extracted from the images and how the model learns these features to
classify the diseases. Grad-cam is employed to visualize and interpret the areas of the images that contribute
most to the classification decision. An easy and efficient ensembling approach known as weighted average is
used to combine predictions from three models to produce final predictions. The results demonstrate the success
and effectiveness of the proposed ensemble approach, INCVX-Net, which attains a predictive accuracy of
99.5% for recognizing five classes of tomato leaf disease. Unlike the single DL model, the INCVX-Net
achieves significant performance even with limited data, offering a promising avenue for applications with
restricted data availability.
Despite the promising results from the INCVX-Net model, there are a couple of limitations to
consider. Firstly, how well ensemble methods works depends on choosing the right base models and the right
way to put them together with the help of suitable aggregation technique. Secondly, implementing these
methods practically might face challenges due to the computational complexities involved, which could require
additional research to find viable solutions. Hence, it is important to focus on improving ensemble techniques
to enhance predictive accuracy while minimizing computational demands. Additionally, investigating the
applicability of ensemble approaches across diverse domains and datasets could provide valuable insights into
their generalization and scalability. This study underscores the effectiveness of ensemble methods in image
recognition applications which helps to mitigate the challenges like over-fitting, data variability, and model
instability.

4. CONCLUSION
The farmers can minimize crop losses by taking timely preventive actions against the diseases. This
is possible only when they can identify the diseases correctly at their early stage. Adopting a DL model to
classify plant diseases is emerged as a promising solution. State-of-the-art DL models extract features from the
training images, which they learn to classify the diseases. However, it is uncertain that the extracted features
are not consistent among all models. This is where the concept of integrating multiple DL models arises. The
Enhanced detection of tomato leaf diseases using ensemble deep learning … (Shruthi Kikkeri Subramanya)
4764  ISSN: 2252-8938

resultant model will effectively identify the diseases by aggregating different viewpoints that help to find
hidden or small features of diseases more effectively. In the proposed INCVX-Net model the feature vectors
are extracted from the base models, which are validated with GRAD-cam and combined using the weighted
average Ensembling technique. To prevent over-fitting, techniques such as dropout, and batch normalization
were employed. The INCVX-Net model exhibits promising outcomes in recognizing and categorizing five
different tomato plant diseases with a remarkable accuracy rate of 99.5% with minimal classification error. The
results are evaluated by comparing with the outcomes of the base models, that the proposed model shows good
performance, proves the adoption of the ensembling technique. Despite its potential as an automatic detector
for early identification of tomato crop leaf diseases, it is necessary to acknowledge the limitations of the study,
including the need for further validation across different tomato/plant species thereby potentially increasing
production. Furthermore, real-world validation with field-collected data under diverse conditions such as
varying lighting, weather, and image quality is essential to accurately assess the ensemble’s effectiveness. The
INCVX-Net model could be deployed on compact computing platforms, transforming it into a standalone
device. This would facilitate easier integration with various systems, potentially enhancing accessibility and
usability in real-world applications.

REFERENCES
[1] L. D. Rosa, S. C. D. S. Nothaft, and M. A. Busato, “Agriculture, human and environmental health: possible approaches,” Collection
of International Topics in Health Science, 2023, doi: 10.56238/colleinternhealthscienv1-020.
[2] D. K. Mahato et al., “Occurrence, impact on agriculture, human health, and management strategies of zearalenone in food and feed:
A review,” Toxins, vol. 13, no. 2, Jan. 2021, doi: 10.3390/toxins13020092.
[3] A. Devlet, “Modern agriculture and challenges,” Frontiers in Life Sciences and Related Technologies, vol. 2, no. 1, pp. 21–29, Apr.
2021, doi: 10.51753/flsrt.856349.
[4] S. Skendzic, M. Zovko, I. P. Zivkovic, V. Lesic, and D. Lemic, “The impact of climate change on agricultural insect pests,” Insects,
vol. 12, no. 5, May 2021, doi: 10.3390/insects12050440.
[5] R. W. Mwangi, M. Mustafa, K. Charles, I. W. Wagara, and N. Kappel, “Selected emerging and reemerging plant pathogens affecting
the food basket: A threat to food security,” Journal of Agriculture and Food Research, vol. 14, Dec. 2023, doi:
10.1016/j.jafr.2023.100827.
[6] M. M. Ali, N. A. Bachik, N. ‘Atirah Muhadi, T. N. Tuan Yusof, and C. Gomes, “Non-destructive techniques of detecting plant
diseases: A review,” Physiological and Molecular Plant Pathology, vol. 108, Dec. 2019, doi: 10.1016/j.pmpp.2019.101426.
[7] R. Patel et al., “A review of recent advances in plant-pathogen detection systems,” Heliyon, vol. 8, no. 12, Dec. 2022, doi:
10.1016/j.heliyon.2022.e11855.
[8] M. Shoaib et al., “An advanced deep learning models-based plant disease detection: A review of recent research,” Frontiers in Plant
Science, vol. 14, Mar. 2023, doi: 10.3389/fpls.2023.1158933.
[9] I. Ahmed, and P. K. Yadav, “A systematic analysis of machine learning and deep learning based approaches for identifying and
diagnosing plant diseases,” Sustainable Operations and Computers, vol. 4, pp. 96-104, 2023, doi: 10.1016/j.susoc.2023.03.001.
[10] A. Ahmad, D. Saraswat, and A. El Gamal, “A survey on using deep learning techniques for plant disease diagnosis and
recommendations for development of appropriate tools,” Smart Agricultural Technology, vol. 3, 2023, doi:
10.1016/j.atech.2022.100083.
[11] R. I. Hasan, S. M. Yusuf, and L. Alzubaidi, “Review of the state of the art of deep learning for plant diseases: A broad analysis and
discussion,” Plants, vol. 9, no. 10, pp. 1–25, 2020, doi: 10.3390/plants9101302.
[12] L. Alzubaidi et al., “Review of deep learning: concepts, CNN architectures, challenges, applications, future directions,” Journal of
Big Data, vol. 8, no. 1, 2021, doi: 10.1186/s40537-021-00444-8.
[13] K. Zhou, Y. Yang, Y. Qiao, and T. Xiang, “Domain adaptive ensemble learning,” IEEE Transactions on Image Processing, vol.
30, pp. 8008–8018, 2021, doi: 10.1109/TIP.2021.3112012.
[14] A. Mohammed and R. Kora, “A comprehensive review on ensemble deep learning: Opportunities and challenges,” Journal of King
Saud University - Computer and Information Sciences, vol. 35, no. 2, pp. 757–774, Feb. 2023, doi: 10.1016/j.jksuci.2023.01.014.
[15] N. K. Trivedi et al., “Early detection and classification of tomato leaf disease using high-performance deep neural network,”
Sensors, vol. 21, no. 23, Nov. 2021, doi: 10.3390/s21237987.
[16] 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.
[17] K. P. Ferentinos, “Deep learning models for plant disease detection and diagnosis,” Computers and Electronics in Agriculture, vol.
145, pp. 311–318, Feb. 2018, doi: 10.1016/j.compag.2018.01.009.
[18] F. Arshad et al., “PLDPNet: End-to-end hybrid deep learning framework for potato leaf disease prediction,” Alexandria Engineering
Journal, vol. 78, pp. 406–418, 2023, doi: 10.1016/j.aej.2023.07.076.
[19] M. A. Ganaie, M. Hu, A. K. Malik, M. Tanveer, and P. N. Suganthan, “Ensemble deep learning: A review,” Engineering
Applications of Artificial Intelligence, vol. 115, 2022, doi: 10.1016/j.engappai.2022.105151.
[20] Y. He, G. Zhang, and Q. Gao, “A novel ensemble learning method for crop leaf disease recognition,” Frontiers in Plant Science,
vol. 14, 2023, doi: 10.3389/fpls.2023.1280671.
[21] S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using deep learning for image-based plant disease detection,” Frontiers in Plant
Science, vol. 7, Sep. 2016, doi: 10.3389/fpls.2016.01419.
[22] M. Iman, H. R. Arabnia, and K. Rasheed, “A review of deep transfer learning and recent advancements,” Technologies, vol. 11, no.
2, Mar. 2023, doi: 10.3390/technologies11020040.
[23] W. Albattah, M. Nawaz, A. Javed, M. Masood, and S. Albahli, “A novel deep learning method for detection and classification of
plant diseases,” Complex & Intelligent Systems, vol. 8, no. 1, pp. 507–524, Feb. 2022, doi: 10.1007/s40747-021-00536-1.
[24] S. Rathi and V. Nirgude, “Improving the accuracy of real field pomegranate fruit diseases detection and visualization using
convolution neural networks and grad-CAM,” International Journal of Data Analysis Techniques and Strategies, vol. 15, no. 1/2,
2023, doi: 10.1504/IJDATS.2023.10055387.

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4757-4765

Int J Artif Intell ISSN: 2252-8938  4765

[25] K. Bayoudh, “A survey of multimodal hybrid deep learning for computer vision: Architectures, applications, trends, and
challenges,” Information Fusion, vol. 105, May 2024, doi: 10.1016/j.inffus.2023.102217.
[26] S. Rajaraman, S. Jaeger, and S. K. Antani, “Performance evaluation of deep neural ensembles toward malaria parasite detection in
thin-blood smear images,” PeerJ, vol. 7, May 2019, doi: 10.7717/peerj.6977.
[27] Z. Ð. Vujovic, “Classification model evaluation metrics,” International Journal of Advanced Computer Science and Applications,
vol. 12, no. 6, 2021, doi: 10.14569/IJACSA.2021.0120670.
[28] P. Kaur et al., “Recognition of leaf disease using hybrid Convolutional neural network by applying feature reduction,” Sensors, vol.
22, no. 2, Jan. 2022, doi: 10.3390/s22020575.
[29] S. Zhao, Y. Peng, J. Liu, and S. Wu, “Tomato leaf disease diagnosis based on improved convolution neural network by attention
module,” Agriculture, vol. 11, no. 7, 2021, doi: 10.3390/agriculture11070651.
[30] M. Altalak, M. A. Uddin, A. Alajmi, and A. Rizg, “A hybrid approach for the detection and classification of tomato leaf diseases,”
Applied Sciences, vol. 12, no. 16, 2022, doi: 10.3390/app12168182.
[31] M. S. Islam et al., “Multimodal hybrid deep learning approach to detect tomato leaf disease using attention based dilated convolution
feature extractor with logistic regression classification,” Sensors, vol. 22, no. 16, 2022, doi: 10.3390/s22166079.
[32] A. Ksibi, M. Ayadi, B. O. Soufiene, M. M. Jamjoom, and Z. Ullah, “MobiRes-Net: A hybrid deep learning model for detecting and
classifying olive leaf diseases,” Applied Sciences, vol. 12, no. 20, 2022, doi: 10.3390/app122010278.
[33] Z. Ullah, N. Alsubaie, M. Jamjoom, S. H. Alajmani, and F. Saleem, “EffiMob-Net: A deep learning-based hybrid model for
detection and identification of tomato diseases using leaf images,” Agriculture, vol. 13, no. 3, Mar. 2023, doi:
10.3390/agriculture13030737.
[34] S. Pudumalar and S. Muthuramalingam, “Hydra: An ensemble deep learning recognition model for plant diseases,” Journal of
Engineering Research, Sep. 2023, doi: 10.1016/j.jer.2023.09.033.

BIOGRAPHIES OF AUTHORS

Shruthi Kikkeri Subramanya completed her M.Tech. degree in Digital

Electronics and Communication at Malnad College of Engineering, Hassan, under Visveswaraya
Technological University, Belgaum, with a dissertation titled “text and image encryption based
on session”. She had 4.5 years of experience as an assistant professor. Currently, she is a full-
time research scholar at BGSIT Adichunchanagiri University, B.G. Nagara, Mandya, India since
2021, focusing her research on AI, particularly deep learning applied to agriculture, specifically
in plant pathology. She had Published two conference papers which are published in scopus. She
can be contacted at email: [email protected].

Naveen Bettahalli is working as an Associate Professor at Department of

Electronics and Communication Engineering in BGS Institute of Technology,
Adichunchanagiri University, B G Nagara, Mandya. He also worked as a Head of the
Department at Electronics and Communication Engineering of Jnanavikas Institute of
Technology, Bangalore. He has 15 Experience includes 13 years in teaching and 2 years of
industry. He has contributed 25 referred articles, including Scopus indexed journals,
conferences, proceedings, and book chapters published by international and national
publishers. He has applied for 5 patents. He has received sponsored research projects from
VGST, AICTE. He is currently guiding 6 Ph.D. students and more than 15 M.Tech. dissertation
in the area of image processing, VLSI and embedded systems, and artificial intelligence. He
can be contacted at email: [email protected].

Enhanced detection of tomato leaf diseases using ensemble deep learning … (Shruthi Kikkeri Subramanya)