DERIENet: A Deep Ensemble Learning Approach for High-Performance Detection of Jute Leaf Diseases

Abstract

Jute, a vital lignocellulosic fiber crop with substantial industrial and ecological relevance, continues to suffer considerable yield and quality degradation due to pervasive foliar pathologies. Traditional diagnostic modalities reliant on manual field inspections are inherently constrained by subjectivity, diagnostic latency, and inadequate scalability across geographically distributed agrarian systems. To transcend these limitations, we propose DERIENet, a robust and scalable classification approach within a deep ensemble learning framework. It is meticulously engineered by integrating three high-performing convolutional neural networks—ResNet50, InceptionV3, and EfficientNetB0—along with regularization, batch normalization, and dropout strategies, to accurately classify jute leaf diseases such as Cercospora Leaf Spot, Golden Mosaic Virus, and healthy leaves. A key methodological contribution is the design of a novel augmentation pipeline, termed Geometric Localized Occlusion and Adaptive Rescaling (GLOAR), which dynamically modulates photometric and geometric distortions based on image entropy and luminance to synthetically upscale a limited dataset (920 images) into a significantly enriched and diverse dataset of 7800 samples, thereby mitigating overfitting and enhancing domain generalizability. Empirical evaluation, utilizing a comprehensive set of performance metrics—accuracy, precision, recall, F1-score, confusion matrices, and ROC curves—demonstrates that DERIENet achieves a state-of-the-art classification accuracy of 99.89%, with macro-averaged and weighted average precision, recall, and F1-score uniformly at 99.89%, and an AUC of 1.0 across all disease categories. The reliability of the model is validated by the confusion matrix, which shows that 899 out of 900 test images were correctly identified and that there was only one misclassification. Comparative evaluations of the various ensemble baselines, such as DenseNet201, MobileNetV2, and VGG16, and individual base learners demonstrate that DERIENet performs noticeably superior to all baseline models. It provides a highly interpretable, deployment-ready, and computationally efficient architecture that is ideal for integrating into edge or mobile platforms to facilitate in situ, real-time disease diagnostics in precision agriculture.

1. Introduction

Jute is a naturally occurring, environmentally beneficial fiber that finds extensive usage in the textile industry as well as other industrial sectors. The Tiliaceae family includes the plants Corchorus capsularis L. and Corchorus olitorius L., which are the source of this highly economical vegetable fiber, which is second only to cotton [1]. A popular option in a world where environmental preservation is becoming more and more important, jute is highly acclaimed for its affordability as well as its sustainability and biodegradability. Jute is cultivated in Bangladesh, India, Nepal, China, Taiwan, Thailand, Vietnam, and Cambodia, among other countries. At least 87% of the world’s jute production comes from the Indo-Bangladesh region [2]. Despite India being the main producer, Bangladesh remains the world’s leading exporter of jute [3].

Jute is an environmentally degradable polymer that occurs naturally. Polyvinyl chloride, polystyrene, polyethylene, and polypropylene are examples of synthetic polymers that have excellent mechanical strength but are not biodegradable and have a significant negative impact on the environment. Jute, on the other hand, provides an environmentally responsible and sustainable substitute that lowers pollutants and aids in environmental preservation [4,5]. There are other uses for jute leaves outside of just their fiber. They have cultural value and are utilized in many different culinary preparations. Their anti-inflammatory and anti-cancer activities are among their other possible therapeutic qualities [6]. Numerous diseases and pests pose a global danger to jute production. Smallholder farmers, whose livelihoods rely significantly on successful harvests, could suffer serious consequences. The main diseases affecting jute leaves include Cercospora Leaf Spot, Powdery Mildew, and various viral infections such as Yellow Mosaic and Golden Mosaic. These diseases result in brownish lesions, mosaic patterns, yellowing, and defoliation of the leaves. These illnesses impair photosynthesis, lower the quality of fiber, cause growth retardation, and in extreme situations, can kill the plant, all of which have a major impact on yields. The main pests that harm jute are the jute caterpillar, jute weevil, and jute stem borer. These pests lower productivity by harming the plant’s leaves, stems, and fibers. Furthermore, it is crucial to detect and stop these jute pests and leaf diseases as soon as possible because jute is important for both the economy and environmental protection.

The traditional technique of plant disease identification relies on farmers or agricultural experts performing physical inspections; however, this method has major drawbacks. It is time-consuming and resource-intensive, which makes it inappropriate for extensive agricultural operations. A precise diagnosis necessitates expertise, which is sometimes unavailable, especially in rural locations. Additionally, visual assessments are prone to inaccuracies and are subjective, which leads to inconsistent diagnoses [7]. Implementing early disease detection and categorization utilizing the resources and technologies that farmers have access to can significantly reduce all the issues that have come up [8]. Traditional machine-learning techniques for the classification of plant diseases have been used to propose a number of solutions. Furthermore, by eliminating reliance on manually created features [9,10], deep-learning methods in agriculture have opened up new avenues for researchers with remarkable flexibility. Convolutional neural networks (CNNs) have recently featured as a potent tool for any classification problem since they can automatically extract significant characteristics from images without human oversight. Furthermore, a number of CNN architectures [11] have been successfully applied for classifying plant leaf diseases. These include DenseNet [12] for tomato leaf disease detection; AlexNet [13,14], which classified leaf diseases in rice, potato, and tomato; EfficientNet [15,16], demonstrating strong performance in classifying diseases in rice and sugarcane leaves; GoogLeNet [17], employed for rice leaf disease classification; and NASNet [18], used for a variety of plant leaf diseases. In addition, Residual Networks (ResNet) [19], applied to apple leaf diseases; MobileNet [20,21], supported leaf disease classification tasks in apples and beans; SqueezeNet [22], detected maize leaf diseases; Inception [23] was utilized for jute leaf diseases; and Visual Geometric Group (VGG) Networks [24] contributed to multi-crop leaf disease identification.

While significant progress has been made in leaf disease detection for various crops, research on jute leaf diseases remains limited, highlighting a critical gap in addressing jute crop health and productivity [25]. A major obstacle in this discipline is the lack of geographically and environmentally diverse data, which could potentially impact the generalizability of the model to various field conditions. Furthermore, the dependence on simple augmentation approaches and the scarcity of extensive annotated datasets may limit the models’ ability to adjust to the complex variability found in actual agricultural settings. This study aims to address these challenges by developing an effective image-based classification system capable of accurately identifying Cercospora Leaf Spot, Golden Mosaic, and Healthy jute leaves. However, the absence of publicly available datasets makes it difficult to classify jute leaf diseases, largely due to limited research in this field [26]. To mitigate this constraint, we utilized the “Jute Leaf Disease Detection” dataset available on Kaggle (https://www.kaggle.com/datasets/mdsaimunalam/jute-leaf-disease-detection, accessed on 3 March 2025) and a proposed data-augmentation technique, Geometric Localized Occlusion and Adaptive Rescaling (GLOAR), to expand the size of the dataset and improve the model’s adaptability and resilience. To develop a strong classification scheme, we applied transfer learning (TL) and enhanced performance by introducing a Deep Ensemble of ResNet, Inception, and EfficientNet (DERIENet), our proposed ensemble model that integrates various deep-learning architectures. Utilizing criteria including accuracy, precision, recall, F1-score, ROC curve, and a confusion matrix, the model’s efficacy was evaluated. The DERIENet model’s superior performance over baseline models demonstrated its efficacy in identifying jute leaf diseases, according to the data.

According to our research, the DERIENet model enhances crop management and lowers losses by improving the accuracy of jute leaf disease diagnosis and giving agricultural professionals a dependable tool for early identification. Below is a summary of the main conclusions drawn from this study:

• We introduced DERIENet, a collaborative ensemble deep-learning architecture designed to accurately classify jute leaf diseases. DERIENet achieves a classification accuracy of 99.89% by integrating the predictive capabilities of several refined deep-learning models—EfficientNetB0, ResNet50, and InceptionV3—into a unified framework enhanced with regularization and dropout layers, significantly improving resilience and robustness against noise and variability in agricultural images.
• We proposed a novel data-augmentation pipeline, Geometric Localized Occlusion and Adaptive Rescaling (GLOAR), which dynamically perturbs image geometry and photometry based on entropy-aware heuristics. This critical innovation enabled the synthetic expansion of a limited dataset from 920 to 7800 images, dramatically enhancing generalization capability and mitigating overfitting in data-constrained scenarios.
• We designed a custom ensemble feature fusion strategy using global average pooling and dense layers with L2 regularization and batch normalization, which not only preserved discriminative features from individual CNN backbones but also facilitated a compact, low-variance representation, enabling higher classification fidelity with minimal computational overhead.
• We conducted a comprehensive comparative evaluation against multiple state-of-the-art baseline models, including MobileNetV2, DenseNet201, InceptionV3, VGG16, ResNet50, and EfficientNetB0, conclusively demonstrating DERIENet’s superior performance across all standard classification metrics and its statistical significance through ablation and ensemble analysis. Moreover, we demonstrated the model’s practical feasibility for real-time deployment in precision agriculture, validating its efficacy through precision, recall, and F1-score metrics, supported by interpretability tools such as ROC curves and confusion matrices, underscoring its potential for deployment on mobile or edge computing devices for in-field jute disease detection.

This document’s structure is as follows: A comprehensive overview of recent studies on jute leaf disease prediction is provided in Section 2, which also presents the pertinent papers. In Section 3, the methodology is described, along with the modeling techniques used in our study and image preprocessing steps. The results and discussion are provided in Section 4, where we compare our proposed model with baseline methods and assess its efficacy. In Section 5, the results and their consequences are interpreted in detail. Section 6 concludes the work by summarizing our findings and offering suggestions for more research.

2. Literature Review

Due to the wide range of symptoms and the need for personal inspection, jute leaf disease detection is extremely difficult. To address these issues, however, recent advancements in machine-learning (ML) and deep-learning (DL) approaches have shown great promise. With an emphasis on the use of ML, DL, and federated learning (FL) techniques, this section examines a number of important works on jute disease and pest detection.

Rahman et al. [27] suggested using the Self-Generated Mask (SGM) technique to identify diseases by identifying the Region of Interest (ROI), to extract an image’s Region of Interest, in jute leaves and stems. A 1772-image, 10-class dataset was used for SGM-enabled CNN training, which reduced annotation time while producing results similar to human annotation. A popular method for improving model accuracy is ensemble learning. Using a real-time dataset, Daphal et al. [28] produced five types of sugarcane leaves: rust, red rot, foliar, yellow leaf, and healthy, with a total of 2569 images. Transfer-learning and ensemble approaches were used in a piece of comparative research to categorize the dataset for illnesses of the sugarcane leaves. With a spatially tuned deep CNN and a sequential CNN, the ensemble model achieved an accuracy of 86.53%. Patil et al. [29] presented "Rice Fusion," an inventive multisensory data integration architecture that diagnoses rice diseases by combining a CNN with a multilayer perceptron (MLP). The model was trained and evaluated using 3200 manually collected rice samples from four categories, and it demonstrated excellent outcomes with a test accuracy of 95.31%. Yigit et al. [30] indicated an SVM classifier for the straightforward binary categorization of sugarcane leaves into healthy and damaged categories. By analyzing the color and texture properties of the photos, the researchers achieved an accuracy of 92.91%. A CNN-based technique for identifying yellow mosaic and chlorosis in jute leaves was presented by Karim et al. [31]. It achieved 96% accuracy on a small dataset of 600 images without feature extraction or preprocessing, underscoring the need for more study and bigger datasets to improve model performance. YOLO-JD, a deep-learning model that incorporates advanced modules such as SPPM and DSCFEM, was proposed by Li et al. [32]. It demonstrated the effectiveness of advanced deep-learning architectures for precise disease identification by achieving a state-of-the-art mean average precision (mAP) of 96.63% when trained on a large dataset covering ten disease types. JuLeDI, a four-layer CNN model trained on 4740 images across three disease categories, was also developed by Uddin et al. [7]. It outperformed GPDCNN and SVM and achieved 96% accuracy, highlighting the usefulness of lightweight CNN models for plant disease diagnosis and their potential for practical applications. A deep-learning approach for jute leaf disease identification was developed by Akhand et al. [33], using a dataset of 1820 images. Using SVM, Random Forest, XGBoost, and custom CNN classifiers with DenseNet121, ResNet152, Xception, and VGG19 extractors, the model’s accuracy of 98.41% and AUC of 0.9988 showed how effective deep learning is in diagnosing diseases. RH Hridoy et al. [34] proposed JuteNet, a hybrid deep-learning model utilizing multi-scale feature fusion of Xception, InceptionResNetV2, and InceptionV3, to detect jute diseases and pests. Trained on jute leaf and stem images, JuteNet achieved 99.47% accuracy on a six-class test set of 2803 images, outperforming the individual models, which achieved 91.83%, 96.11%, and 98.86%, respectively. Kaushik et al. [35] used a dataset of 1820 annotated images from Kaggle to diagnose jute leaf diseases using a deep-learning technique based on the ResNet50 architecture. The dataset underwent preprocessing procedures like as rescaling, normalization, and data augmentation in order to reduce overfitting and improve model resilience. After completing these procedures, the improved model showed minimal misclassification rates, balanced precision and recall metrics, and an accuracy of 94%.

The detection of jute pests has been transformed by developments in DL and ML, which allow for accurate and timely identification to improve pest control. Pests have a major influence on the production of jute. Therefore, DL and ML-powered image-based detection systems assist farmers in promptly identifying infestations, minimizing the need for manual inspections, and increasing productivity. Using the VGG19 CNN architecture with transfer learning on the ImageNet database, Sourav et al. [36] detected four common jute pests with 95.86% accuracy and integrated the model into iOS and Android apps for real-world applications. On a dataset of 2200 pictures, Wang et al. [37] created PestDetector, a deep CNN model that achieved 99.18% training accuracy and 99.00% validation accuracy, and 98% test accuracy; nevertheless, the small sample size hampered its generalizability. A transfer-learning-based model called JutePestDetect was presented by Talukder et al. [38]. It was tested on 17 pest classes, each with about 380 pictures, and DenseNet201 had the best accuracy of 99%. The research, however, only examined early-stage detection and did not analyze potential dataset biases. A lightweight CNN model that exceeded current techniques with 99.43% accuracy was proposed by Rana et al. [39]. It was a workable solution for farmers, utilizing SHAP visualizations to understand predictions and Gradio for real-time accessibility. Ali et al. [40] developed transfer-learning-based models using a VCI-validated dataset of 7235 pictures from 17 pest classes; DenseNet201 had the highest accuracy of 97%.

Federated learning (FL), a solution to data privacy problems, has become a viable method for identifying jute diseases. Using FedAvg, Vats et al. [26] created a FL-based CNN model that achieved a macro-average precision of 73.41–90.72%; nevertheless, real-world deployment requires improvements. FL’s potential for decentralized learning was demonstrated by Bansal et al. [41], who used data from numerous users to categorize five jute leaf diseases. They achieved macro-average precision between 79.92% and 83.67%. Rajput et al. [42] improved this method by combining FL with CNN in several ecological zones, demonstrating its suitability for extensive crop health monitoring with a 98% accuracy rate on client OB5. In a similar study, [43] achieved 98% accuracy using federated averaging with CNNs in decentralized data sets, while improving precision, recall, and F1-score, confirming the efficacy of FL in the classification of diseases while preserving data security and scalability. Conventional methods and resistance breeding have been investigated as alternatives to ML and DL procedures in the fight against jute diseases. For resistance to stem rot (Macrophomina phaseolina), a significant jute disease, Rijal et al. [44] assessed six Corchorus olitorius jute germplasm lines. According to disease rating scales of 0.73, 0.68, and 0.63, JRO-524 has continuously shown strong resistance over a three-year period (2019–2021), making it a viable option for resistance breeding. Likewise, Reza et al. [45] created an Android-based system for jute stem disease identification that uses hue-based segmentation and a Multi-SVM classifier. Although it produces findings in less than three seconds, the accuracy of the system is dependent on a pre-defined database, which could reduce its usefulness if the database is not diverse or indicative of real-world circumstances.

Table 1. Overview of previous work on leaf disease detection methods.

Ref.	Model(s) Used	Dataset	#Images	#Classes	Transfer Learning	Ensemble Learning	Data Augmentation	Accuracy
[7]	Four-layer CNN	Jute leaf(own)	4740	3	No	No	Yes	96%
[28]	CNN, Xception, ResNet50, MobileNetV2, VGG19, and EfficientNetB7	Sugarcane (own)	2569	5	Yes	Yes	No	86.53%
[29]	CNN, MLP	Rice(own)	3200	4	No	Yes	No	95.31%
[30]	SVM	Multi-plant (Folio)	637	32	No	No	No	92.91%
[31]	CNN	Jute leaf	600	2	No	No	No	96%
[32]	YOLO-JD	Jute leaf and pest	4418	10	No	No	No	96.63%
[33]	ResNet152 + Custom CNN	Jute leaf(Kaggle)	1820	3	Yes	No	No	98.41%
[35]	ResNet50	Jute leaf(Kaggle)	1820	-	Yes	No	Yes	94%
[36]	DCNN	Jute pest	1535	4	Yes	No	No	95.86%
[38]	DenseNet201	Jute pest	380 per class	17	Yes	No	Yes	99%
[43]	FL-based CNN	Jute leaf	-	5	No	No	No	98%

provides a thorough overview of the research findings in the field of plant leaf disease identification, with a focus on various architectural approaches.

Our suggested DERIENet model employs an ensemble approach that combines InceptionV3, ResNet50, and EfficientNetB0 to detect jute leaf diseases. Through the positive traits of each architecture, the models are combined to enhance classification performance. We begin with 920 high-quality images from the Kaggle jute leaf dataset and use an advanced data-augmentation technique (GLOAR) to reach 7800 images and overcome the problem of limited data. Better robustness and flexibility across a range of environmental variables are ensured by the larger dataset size. The designation of jute leaf diseases with a high accuracy of 99.89% is achieved by the DERIENet model, which has been refined by transfer learning.

3. Methodology

The DERIENet architecture utilizes jute leaf images as input to accurately predict their respective class labels. The procedure begins with collecting a dataset of jute plant images, then proceeds to preprocessing stages such as data augmentation and scaling to a uniform size. These actions improve the size, diversity, and quality of the dataset. After processing, the dataset is split into subsets for testing, validation, and training purposes. A number of baseline models, including MobileNetV2, DenseNet201, InceptionV3, VGG16, ResNet50, and EfficientNetB0, are refined through the integration of transfer learning. To enhance classification performance, the DERIENet model is introduced, incorporating models such as InceptionV3, ResNet50, and EfficientNetB0. According to a comparative study, the suggested ensemble technique is superior to the baseline models and achieves the highest classification accuracy, demonstrating its efficacy in detecting jute leaf diseases. Figure 1 has an illustration of the general structure of the suggested method.

3.1. Data Description

The dataset employed in this research, titled “Jute Leaf Disease Detection,” can be accessed at https://www.kaggle.com/datasets/mdsaimunalam/jute-leaf-disease-detection, accessed on 5 March 2025. It comprises 920 high-resolution images that are divided into three classes (depicted in Figure 2): Cercospora Leaf Spot, Golden Mosaic, and Healthy Leaf. With the full consent of local agricultural officials, these images were first collected during an extensive field trip in Dinajpur and Brahmanbaria, Bangladesh. Using the proposed data-augmentation technique (GLOAR), the total number of images was increased to 7800 to improve model transferability and enhance the dataset’s diversity.

In particular, the dataset includes 2604 images (33.385%) for the Golden Mosaic and Healthy leaf classes, and 2592 images (33.230%) for Cercospora Leaf Spot. The dataset’s distribution of categories after augmentation is presented in Figure 3.

3.2. Data Preprocessing

A crucial phase in obtaining raw data for ML application is data preparation; ensuring that the dataset is organized, consistent, and free from noise enhances the model’s capability for learning and generalizing it [46,47]. Data augmentation and resizing were two important preprocessing methods used in this work to enhance the dataset’s usability.

3.2.1. Data Resizing

To make sure that all input images have consistent dimensions and are compatible with the neural network model, image resizing is a crucial preprocessing step. Every image included in this study was reduced to 224 × 224 pixels, which was the fixed resolution. By ensuring uniformity throughout the dataset, this scaling procedure helps the model handle photos efficiently for both training and testing. The complete instructions for resizing images are provided by Algorithm 1.

Algorithm 1 Image resizing algorithm

Require: List of image files I, target dimensions $w\times h$ (e.g., 224 × 224 pixels)

Ensure: Resized images saved in appropriate directories

1: for each image $img\in I$ do

2:        Open the image file

3:         if image is not in RGB format then

4:             Convert the image to RGB

5:         end if

6:        Resize the image to $w\times h$

7:        Save the resized image to the output directory

8:end for

3.2.2. Data Augmentation

An innovative augmentation technique called Geometric Localized Occlusion and Adaptive Rescaling (GLOAR), described in Algorithm 2, is deployed to enhance the efficacy of the model under a variety of input situations. By creating a controlled cascade of transformations, this technique greatly increases the dataset’s size and produces a huge amount of visual variation. Context-aware transformation is made possible by the GLOAR algorithm, which dynamically modifies the augmentation strength according to the global brightness and entropy of each image. Among the geometric augmentations are translations (up to $15\%$), perspective warping, and random affine transformations (rotations up to ${30}^{°}$, scaling between $0.8$ and $1.2$, shear up to ${15}^{°}$). Furthermore, by cropping small patches, modifying them (for example, by pixel shuffling, rotation, or blurring), and then re-inserting them, a micro-patch jittering step creates local perturbations that resemble localized distortion or occlusion. Adaptive brightness adjustment within $(1-0.2\alpha ,1+0.2\alpha )$ and contrast variation inside $(1-0.3\alpha ,1+0.3\alpha )$ are examples of entropy-aware photometric transformations. $\alpha$ is an intensity factor that is obtained from image statistics. To improve spectral diversity, hue and saturation are also slightly altered in the HSV color space. All images have been cropped to $224\times 224$ pixels after augmentation, and their pixel values are scaled to the $[0,1]$ range for normalization. Any missing pixels from transformations are filled in using nearest-neighbor filling. Figure 4 displays examples of the augmented images.

Algorithm 2 Geometric Localized Occlusion and Adaptive Rescaling (GLOAR)

Require: Raw input image I

Ensure: Augmented image $I^{\prime }$

1: Compute global brightness B and entropy H of I

2: Set augmentation intensity parameter $\alpha \leftarrow f(B,H)$, where f is a learnt or heuristic function

3: Step 1: Geometric Transformations (weighted by $\alpha$)

4: Apply random affine transformation: rotation ($\pm {30}^{\circ }$), scaling ($0.8$–$1.2$), shear ($\pm {15}^{\circ }$)

5: Apply random translation ($\pm 15\%$ of image height/width)

6: Apply perspective warping with probability $p=\alpha$

7: Step 2: Localized Micro-Patch Jittering

8: for $i=1$ to n random patches do

9:        Crop a patch $P_i$ of size $8\%$–$15\%$ at a random location

10:        Apply perturbation: small rotation, pixel shuffle, or blur on $P_i$

11:        Reinsert $P_i$ into its original location in I

12: end for

13: Step 3: Photometric Transformations (entropy-aware)

14: Adjust brightness in range $(1-0.2\alpha ,1+0.2\alpha )$

15: Adjust contrast in range $(1-0.3\alpha ,1+0.3\alpha )$

16: Convert to HSV color space

17: Perturb hue by $\pm {10}^{\circ }$

18: Scale saturation in range $(0.9,1.1)$

19: return Augmented image $I^{\prime }$

3.3. Train, Validation and Test Split

To improve model generalization and robustness, the original dataset, which included 920 images of excellent quality taken in Dinajpur and Brahmanbaria, Bangladesh, was augmented to 7800 images. A total of 6000 images were assigned for training, with 900 images allotted for both testing and validation. The three classes in the dataset, namely Golden Mosaic, Cercospora Leaf Spot, and Healthy, are evenly distributed among all splits. With the training set aiding in feature learning, the validation set aiding in hyperparameter tuning to avoid overfitting, and the test set offering an objective performance evaluation, this hierarchical partitioning guarantees efficient model training.

3.4. Proposed DERIENet Model

For the purpose of detecting jute leaf diseases, our proposed DERIENet model integrates EfficientNetB0, InceptionV3, and ResNet50. As previously mentioned, the Jute Leaf Disease dataset was initially obtained from Kaggle. In order to improve the quality of the dataset and its ability to adapt to new data, a data-augmentation technique (GLOAR) was applied, ensuring variability in image features. As described in Section 3, each image was then resized to a consistent $224\times 224$ pixel dimension. To enhance feature extraction and mitigate overfitting, three TL models were applied by incorporating additional dense layers with L2 regularization, batch normalization, ReLU activation, and dropout layers. These models were trained using the Adam optimizer with the categorical cross-entropy loss function, ensuring efficient learning and convergence. After evaluating their individual performances, the three top-performing TL models were selected and integrated into a custom ensemble framework. This ensemble approach leverages the advantages of each model, resulting in enhanced classification precision and adaptability in detecting jute leaf diseases.

DERIENet Architecture

The DERIENet architecture is specifically designed to handle RGB images of size $224\times 224\times 3$, ensuring compatibility with most deep-learning frameworks. Pre-trained models are utilized as feature extractors, with each model initialized using ImageNet weights. The configuration of the suggested model for identifying jute leaf diseases is depicted in Figure 5.

Let X represent the input image, where $X\in {\mathbb{R}}^{224\times 224\times 3}$. To extract feature maps, the models $M_{\mathrm{EfficientNetB}0},M_{\mathrm{InceptionV}3},M_{\mathrm{ResNet}50}$ independently process the input image X. Compact feature vectors are generated by passing these feature maps through global average pooling (GAP) layers. Specifically, for each model:

$${\mathbf{E}}_{\mathrm{pool}}=\mathrm{GAP}\left(M_{\mathrm{EfficientNetB}0}\left(\mathbf{X}\right)\right)$$

(1)

$${\mathbf{I}}_{\mathrm{pool}}=\mathrm{GAP}\left(M_{\mathrm{InceptionV}3}\left(\mathbf{X}\right)\right)$$

(2)

$${\mathbf{R}}_{\mathrm{pool}}=\mathrm{GAP}\left(M_{\mathrm{ResNet}50}\left(\mathbf{X}\right)\right)$$

(3)

Subsequently, these pooled feature vectors ${\mathbf{E}}_{\mathrm{pool}},{\mathbf{I}}_{\mathrm{pool}},{\mathbf{R}}_{\mathrm{pool}}$ are concatenated to form a single feature vector $\mathbf{C}$:

$$\mathbf{C}=\mathrm{concat}({\mathbf{E}}_{\mathrm{pool}},{\mathbf{I}}_{\mathrm{pool}},{\mathbf{R}}_{\mathrm{pool}})$$

(4)

The concatenated vector $\mathbf{C}$ is processed further by passing it through a sequence of dense, fully linked layers. The final classification decision and the refinement of the characteristics retrieved by the basic models depend heavily on these layers. Figure 6 illustrates that the complex set structure of these fully interconnected layers is displayed. The ReLU (Rectified Linear Unit) function activates each of the 512 neurons in the first fully linked layer to add non-linearity:

$${\mathbf{D}}_{\mathbf{1}}=\mathrm{ReLU}(W_1·\mathbf{C}+b_1)$$

(5)

This layer’s weights are subjected to L2 regularization in order to avoid overfitting. The first dense layer’s L2 penalty term is as follows:

$$\mathrm{L}2\mathrm{Regularization}\mathrm{Term}\mathrm{for}{\mathbf{D}}_{\mathbf{1}}:\lambda \sum _i{W}_{1i}^2$$

(6)

The strength of regularization is denoted by $\lambda$. The model is better able to generalize and prevent overfitting when the value of $\lambda$ applies a stronger penalty to the weights. Following L2 regularization, the output of the dense layer is subject to batch normalization, which stabilizes training by lowering internal covariate shift:

$${\mathbf{D}}_{\mathbf{1}\mathrm{norm}}=\mathrm{BatchNorm}\left({\mathbf{D}}_{\mathbf{1}}\right)$$

(7)

After batch normalization, a 50% dropout layer is also added. By randomly deactivating half of the neurons during training, dropout helps avoid overfitting and strengthens the network by requiring it to acquire more generic properties that are independent of particular neurons.

$${\mathbf{D}}_{\mathbf{1}\mathrm{drop}}=\mathrm{Dropout}\left(0.5\right)\left({\mathbf{D}}_{\mathbf{1}\mathrm{norm}}\right)$$

(8)

The second fully connected layer consists of 256 neurons and follows a similar structure. Additionally, the dropout layer mitigates overfitting by introducing noise during training. Similar to the previous layer, the weights in this layer undergo L2 regularization, followed by batch normalization and another dropout layer.

$${\mathbf{D}}_{\mathbf{2}}=\mathrm{ReLU}(W_2·\left({\mathbf{D}}_{\mathbf{1}\mathrm{drop}}\right)+b_2)$$

(9)

$${\mathbf{D}}_{\mathbf{2}\mathrm{drop}}=\mathrm{Dropout}\left(0.5\right)\left(\mathrm{BatchNorm}\left({\mathbf{D}}_{\mathbf{2}}\right)\right)$$

(10)

The model’s output layer, which comprises three neurons, represents each class of jute leaf disease. The model’s predictions are probabilistic and comprehensible as confidence scores for each class when the SoftMax activation function converts the dense layers’ output into a probability distribution across the classes. The SoftMax activation function is used by the output layer to give each class a probability score:

$$\mathbf{O}=\mathrm{SoftMax}(W_3·{\mathbf{D}}_{\mathbf{2}\mathrm{drop}}+b_3)$$

(11)

The Adam optimizer is employed to train the model. It has a learning rate of $1\times {10}^{-4}$ and uses categorical cross-entropy as the loss function. Early stopping is used to avoid needless overfitting by terminating training after 10 consecutive epochs without improving validation loss. The ReduceLROnPlateau technique reduces the learning rate by 50% if the validation loss fails to improve after five epochs. Later on in the training process, this technique aids in fine-tuning the model’s weights, particularly when achieving further gains at a faster learning rate becomes challenging. At last, the completely trained model, represented by ${\mathbf{M}}_{\mathbf{c}}$, is prepared for comparison and implementation.

3.5. Baseline Models

We utilized a variety of TL models for comparison that have demonstrated strong performance in picture classification. These models, including EfficientNetB0, ResNet50, InceptionV3, MobileNetV2, DenseNet201, and VGG16, are useful for a range of image identification applications because they each have unique architectural benefits.

3.5.1. EfficientNetB0

EfficientNetB0 [48] adaptively recalibrates channel-wise feature responses using Mobile Inverted Bottleneck Convolution (MBConv) blocks with swish activation and Squeeze-and-Excitation (SE) blocks. These components are characterized as follows:

$$\mathrm{Output}=\mathrm{BN}\left({\mathrm{Conv}}_{1\times 1}\left(\mathrm{BN}\left({\mathrm{DWConv}}_{3\times 3}\left(\mathrm{BN}\left({\mathrm{Conv}}_{1\times 1}\left(x\right)\right)\right)\right)\right)\right)$$

(12)

$$\mathrm{Swish}\left(x\right)=x·\sigma \left(x\right)={\displaystyle \frac{x}{1+e^{-x}}}$$

(13)

$$\mathrm{SE}\left(x\right)=x·\sigma (W_2·\delta (W_1·\mathrm{GlobalAvgPool}\left(x\right)))$$

(14)

where ${\mathrm{Conv}}_{1\times 1}$: 1 × 1 pointwise convolution, ${\mathrm{DWConv}}_{3\times 3}$: 3 × 3 depthwise convolution, BN: Batch Normalization, $W_1,W_2$: weight matrices for fully connected layers, $\delta$: ReLU activation and $\sigma$: Sigmoid activation. After the MBConv blocks, a final pooling and dense layer is applied:

$$\mathrm{Output}=\mathrm{SoftMax}(W·\mathrm{GlobalAvgPool}(x)+b)$$

(15)

where W: weight matrix, b: bias, $\mathrm{GlobalAvgPool}\left(x\right)$: Averages each channel before classification.

3.5.2. ResNet50

The ResNet-50 model presents a residual learning architecture in which the residual block serves as the fundamental building block [49]. The residual block’s output, $\mathbf{y}$, can be written as follows:

$$\mathbf{y}=\mathcal{F}(\mathbf{x},\left\{W_i\right\})+\mathbf{x}$$

(16)

where the input feature map is $\mathbf{x}$ and the residual function, $\mathcal{F}(\mathbf{x},\left\{W_i\right\})$, represents the learnt transformation of the input $\mathbf{x}$. Three layers make up the bottleneck design of the residual block in ResNet-50:

$$\mathcal{F}\left(\mathbf{x}\right)=W_3·\sigma (W_2·\sigma (W_1·\mathbf{x}))$$

(17)

where $W_1$ represents a $1\times 1$ convolution used for channel reduction, $W_2$ is a $3\times 3$ convolution responsible for spatial processing, and $W_3$ is another $1\times 1$ convolution that restores the channels. Additionally, $\sigma$ denotes the ReLU activation function.

To match the dimensions if the input and output are different, a projection shortcut is utilized:

$$\mathbf{y}=\mathcal{F}\left(\mathbf{x}\right)+W_s\mathbf{x}$$

(18)

where $W_s$ is a $1\times 1$ convolutional layer.

3.5.3. InceptionV3

Inception-v3 uses methods such as factorized convolutions, Inception modules, label smoothing regularization, and grid size reduction to increase computational efficiency and model performance [50]. The factorized convolutions approach substitutes smaller convolutions for larger ones, such as $5\times 5$). For instance, two $3\times 3$ convolutions are produced by factorizing a $5\times 5$ convolution.

$$\mathbf{y}=W_{3\times 3}^{\left(2\right)}·\sigma (W_{3\times 3}^{\left(1\right)}·\mathbf{x})$$

(19)

The Inception module utilizes parallel convolutional paths to process the input at several scales. These pathways’ results are concatenated.

$$\mathbf{y}=\mathrm{Concat}[W_{1\times 1}·\mathbf{x},W_{3\times 3}·\sigma (W_{1\times 1}·\mathbf{x}),W_{3\times 1}·\sigma (W_{1\times 3}·\sigma (W_{1\times 1}·\mathbf{x})),\mathrm{MaxPool}\left(\mathbf{x}\right)]$$

(20)

Regularization of label smoothing softens the target labels, preventing overfitting. Regularization of label smoothing softens the target labels, preventing overfitting. The below formula is employed to estimate the adjusted probabilities:

$$q^{\prime }\left(k\right|x)=(1-ϵ){\delta }_{k,y}+{\displaystyle \frac{ϵ}{K}}$$

(21)

where $ϵ=0.1$ is the smoothing factor, K is the number of classes, and ${\delta }_{k,y}$ is the Dirac delta function, which equals 1 for the true label y and 0 otherwise. The spatial dimensions of the feature maps are essentially decreased using strided convolutions and max-pooling for grid size reduction.

$$\mathbf{y}=\mathrm{Concat}\left[{\mathrm{Conv}}_{3\times 3,\mathrm{stride}=2}\left(\mathbf{x}\right),{\mathrm{MaxPool}}_{3\times 3,\mathrm{stride}=2}\left(\mathbf{x}\right)\right]$$

(22)

3.5.4. MobileNetV2

MobileNetV2 reduces memory utilization and computational complexity using bottleneck blocks and depthwise separable convolutions [51]. The computational cost of a conventional convolution is:

$${\mathcal{C}}_{\mathrm{std}}=h_i·w_i·d_i·d_j·k^2$$

(23)

The cost of a depthwise separable convolution is:

$${\mathcal{C}}_{\mathrm{dw}}=h_i·w_i·d_i\left(k^2+d_j\right)$$

(24)

The cost for a bottleneck block with expansion factor t, input channels $d^{\prime }$, output channels $d^{″}$, and kernel size k is:

$${\mathcal{C}}_{\mathrm{bottleneck}}=h·w·d^{\prime }·t\left(d^{\prime }+k^2+d^{″}\right)$$

(25)

The peak memory required for a computational graph G is bounded by:

$$M\left(G\right)=\underset{op\in G}{max}\left[\sum _{A\in {\mathrm{op}}_{\mathrm{in}}}\left|A\right|+\sum _{B\in {\mathrm{op}}_{\mathrm{out}}}\left|B\right|+\left|\mathrm{op}\right|\right]$$

(26)

Let $\mathcal{F}\left(x\right)=[A\circ \mathcal{N}\circ B]x$ be a bottleneck block, where A and B are linear transformations, and $\mathcal{N}=\mathrm{ReLU}6\circ \mathrm{dwise}\circ \mathrm{ReLU}6$.

$$\mathcal{F}\left(x\right)=[A\circ \mathcal{N}\circ B]x$$

(27)

3.5.5. DenseNet201

Let L represent the total count of layers in the network. The outcome of the ℓ-th layer ($\ell \in \{1,2,...,L\}$) is determined by applying a function to the concatenated feature maps from all preceding layers [52].

$${\mathbf{x}}_{\ell }=H_{\ell }\left([{\mathbf{x}}_0,{\mathbf{x}}_1,\dots ,{\mathbf{x}}_{\ell -1}]\right)$$

(28)

where ${\mathbf{x}}_{\ell }$ represents the resulting feature maps of the ℓ-th layer, $[{\mathbf{x}}_0,{\mathbf{x}}_1,\dots ,{\mathbf{x}}_{\ell -1}]$ indicates the combination of maps of features from all previous layers along the channel dimension, and $H_{\ell }$ refers to the composite function applied at the ℓ-th layer, which includes the required transformations.

3.5.6. VGG16

The VGG-16 model comprises multiple convolutional layers succeeded by a fully connected layer [53]. The operations in each layer are as follows: The output of the ℓ-th convolutional layer ($\ell \in \{1,2,\dots ,L\}$) is calculated by applying a convolution operation with a $3\times 3$ filter and a ReLU activation:

$${\mathbf{x}}_{\ell }=\mathrm{ReLU}\left({\mathrm{Conv}}_{3\times 3}\left({\mathbf{x}}_{\ell -1}\right)\right)$$

(29)

where ${\mathbf{x}}_{\ell -1}$ is the input map of features from the preceding layer. Max-pooling utilizing a $2\times 2$ window and an interval of 2 is used subsequent to convolutional layers, diminishing the spatial dimensions of the feature maps:

$${\mathbf{x}}_{\mathrm{pool}}={\mathrm{MaxPool}}_{2\times 2}\left({\mathbf{x}}_{\ell }\right)$$

(30)

The result of the fully connected layers is computed as:

$$\mathbf{y}=\mathrm{SoftMax}(W_3·\mathrm{ReLU}(W_2·\mathrm{ReLU}(W_1·{\mathbf{x}}_{\mathrm{flattened}}+b_1)+b_2)+b_3)$$

(31)

where ${\mathbf{x}}_{\mathrm{flattened}}$ denotes the flattened output from the final pooling layer, whereas $W_1,W_2,W_3$ represent the weight matrices for the fully connected layers, accompanied with the associated bias terms $b_1,b_2,b_3$.

3.6. Model Performance Metrics

Model assessment is a crucial stage in the machine-learning process, which establishes the model’s capacity to generalize to unseen data. This study used many methodologies, including the confusion matrix, receiver operating characteristic (ROC) curve, and several key performance measures, to assess the model’s efficacy. Accuracy, recall, F1-score, and precision are metrics that provide important information about the overall efficacy of the model [54].

3.6.1. Accuracy

The proportion of properly estimated instances to all predictions is known as accuracy [55], and it is a key evaluation parameter in machine learning. It offers an overall assessment of a model’s performance on unseen data. Accuracy is computed using samples from the test set to assess the generalizability of a model.

The accuracy is defined as:

$$\mathrm{Accuracy}={\displaystyle \frac{M}{N}}\times 100\%$$

(32)

where N is the overall number of instances in the test set, and M denotes the total quantity of samples for which the model accurately predicted the class labels.

3.6.2. Precision

The accuracy of positive forecasts is measured by a parameter called precision. It calculates how many accurately identified positive instances there were out of all the instances that were projected to be positive [56]. In a multiclass problem, the one-vs-all technique is used to calculate precision for each class. Additional names for precision include Positive Predictive Value (PPV).

The precision for each class c is given by:

$${\mathrm{Precision}}_c={\displaystyle \frac{T{P}_c}{T{P}_c+F{P}_c}}$$

(33)

where $T{P}_c$ denotes the quantity of true positive samples accurately categorized as class c, and $F{P}_c$ signifies the quantity of false positive samples inaccurately classified as class c.

3.6.3. Recall

The completeness of positive predictions is assessed using a statistic called recall [57], sometimes referred to as sensitivity. The percentage of real positive cases that the model accurately detected is measured. When dealing with a multiclass situation, the one-vs-all technique is used to determine recall for each class.

The recall for each class c is given by:

$${\mathrm{Recall}}_c={\displaystyle \frac{T{P}_c}{T{P}_c+F{N}_c}}$$

(34)

where $T{P}_c$ is the number of true positive samples correctly classified as class c, and $F{N}_c$ is the number of false negative samples that should have been classified as class c but were misclassified as other classes.

3.6.4. F1-Score

The F1-score offers a singular metric that equilibrates accuracy and recall by calculating their harmonic mean. [58]. This approach is highly beneficial for addressing unbalanced datasets as it considers both False Positives and False Negatives. An elevated F1-score indicates that the model is proficient in identifying affirmative examples, with recall and precision in equilibrium.

The F1-score for each class c is given by:

$${\mathrm{F}1\text{-score}}_c=2\times {\displaystyle \frac{{\mathrm{Precision}}_c\times {\mathrm{Recall}}_c}{{\mathrm{Precision}}_c+{\mathrm{Recall}}_c}}$$

(35)

where ${\mathrm{Precision}}_c$ and ${\mathrm{Recall}}_c$ are the precision and recall values for class c, respectively.

4. Results and Discussion

Several baseline models, such as EfficientNetB0, ResNet50, InceptionV3, VGG-16, DenseNet201, and MobileNetV2 were studied for jute leaf disease identification. The best-performing models were chosen for additional development. Our suggested DERIENet model achieved an accuracy of 99.89%, outperforming all baseline models. Ablation research validated the contributions of each component, and an analysis of per-class precision, recall, and F1-scores successfully solved the problem of class imbalance. Our method’s superiority was further illustrated by a comparison with the most advanced models. With its exceptional accuracy, robustness, and usefulness in actual agricultural settings, the suggested ensemble model sets a new benchmark for jute leaf disease detection.

4.1. Results of Baseline Models

This section utilizes ROC curve analysis and confusion matrix analysis to analyze the effectiveness of six baseline models. These evaluation methods aid in comprehending each model’s classification accuracy as well as how well it can differentiate between various categories.

4.1.1. Confusion Matrix Analysis

The jute leaf disease classification results of the baseline models, shown in Figure 7, provide useful information based on confusion matrices generated from 900 test cases. As the image shows, the models’ performance varied, with some demonstrating strengths in particular areas. While EfficientNetB0 performed well in distinguishing between Healthy Leaf and Cercospora Leaf Spot, it misclassified three instances of Healthy Leaf as Cercospora Leaf Spot and three instances of Cercospora Leaf Spot as Golden Mosaic. ResNet50 effectively classified both Golden Mosaic and Cercospora Leaf Spot. However, it misclassified two cases of Cercospora Leaf Spot as Healthy Leaf and 13 cases of Golden Mosaic. InceptionV3 correctly detected Healthy Leaf and Cercospora Leaf Spot. However, it misclassified five Golden Mosaic cases as Cercospora Leaf Spot and nine Cercospora Leaf Spot cases as Golden Mosaic, with a few minor misclassifications. MobileNetV2 exhibited noticeable misclassifications, particularly in Golden Mosaic, where 4 cases were incorrectly classified as Healthy Leaf and 9 as Cercospora Leaf Spot. One instance of Cercospora Leaf Spot was also mistakenly identified as Healthy Leaf, and 13 cases as Golden Mosaic. DenseNet201 successfully recognized all 300 Golden Mosaic samples. Ten cases of Cercospora Leaf Spot, however, were mistakenly identified as Golden Mosaic. Healthy Leaf had 298 accurate predictions and only two incorrect classifications: Cercospora Leaf Spot and Golden Mosaic. VGG16 effectively identified 267 Cercospora Leaf Spot cases; however, it incorrectly categorized 14 as Golden Mosaic and 19 as Healthy Leaf. Furthermore, it misidentified three Healthy Leaf cases as Cercospora Leaf Spot and 49 Golden Mosaic cases as such.

4.1.2. ROC Curve Analysis

AUC scores for the evaluation of ROC curves for six develops of the baseline model ranged from 93.38% to an almost flawless 99.99%, indicating a remarkable classification ability in the diagnosis of jute leaf disease (Figure 8).

The most accurate model was DenseNet201, which obtained remarkable AUC scores of 99.97% for Golden Mosaic detection and 99.99% for Cercospora Leaf Spot and Healthy Leaf categories. Across all disease categories, the ResNet50 design consistently produced excellent results (AUC: 99.48–99.82%). With AUC values above 99.87% for every disease class, InceptionV3 and EfficientNetB0 also continued to perform at a high level. Despite being lightweight, MobileNetV2 outperformed some more complicated architectures and produced impressive results (AUC: 99.58–99.97%). Visual examination of the ROC plots revealed that the best-performing models had curves that were primarily centered in the upper-left quadrant, indicating a low percentage of false positives and a high proportion of real positives. However, VGG16 was an anomaly, showing comparatively poorer performance, especially when it came to identifying the Cercospora Leaf Spot (AUC = 93.38%). This ROC-based evaluation ensures that, prior to deployment, the models meet the exacting accuracy requirements needed for precision agriculture.

4.2. Results of Proposed DERIENet Model

The DERIENet model produced exceptional results with a 99.89% accuracy, precision, recall, and F1-score. This outstanding performance made DERIENet the top model in this study, exceeding all baseline models. The training and validation graphs (Figure 9) further illustrate the model’s exceptional performance.

Over 50 epochs, the accuracy and loss curves showed a consistent convergence, with training and validation accuracy hitting 99.89% and the loss dropping off drastically to about 0.1. Strong learning and generalization are indicated by these patterns, indicating that the model successfully avoids overfitting. Figure 10, which displays the training and test confusion matrices, provides additional evidence of the DERIENet model’s performance. As demonstrated, the model correctly identified 899 out of 900 test-set images of jute leaf disease, with just one misclassification. The model’s excellent accuracy and low errors are demonstrated by the thorough confusion matrix analysis, which also shows how well it detects and categorizes jute leaf diseases.

Figure 11 displays the receiver operating characteristic (ROC) curve for the DERIENet model, illustrating its remarkable performance. Each of the three categories has an area under the curve (AUC) of 1, which illustrates the model’s outstanding class distinctiveness and flawless classification performance.

Accuracy alone is not a trustworthy way to evaluate a model’s performance since high accuracy might conceal poor performance on underrepresented classes, which is a prevalent problem in publicly available datasets with class imbalance [55]. We tackled this by concentrating on macro-averaged and weighted-averaged precision, recall, and F1-scores, which guarantee that all classes receive the same treatment regardless of sample size. Regarding F1-score, recall, and precision, our suggested model produced remarkable results of 99.89%. While correctly recognizing True Positives, the model successfully reduces False Positives and False Negatives, based on the elevated recall and precision metrics.

Furthermore, the precision, recall, and F1-scores for every class are shown in

Table 2. Classification report of the proposed model (DERIENet).

Class	Precision (%)	Recall (%)	F1-Score (%)	Support
Cercospora Leaf Spot	99.67	100	99.83	300
Golden Mosaic	100	99.67	99.83	300
Healthy Leaf	100	100	100	300
Accuracy			99.89	900
Macro Average	99.89	99.89	99.89	900
Weighted Average	99.89	99.89	99.89	900

. The classification report of the suggested model makes it abundantly evident how well it can identify the states of jute leaves in both diseased and healthy groups. The model’s 100% recall, 99.67% precision, and 99.83% F1-score for the Cercospora Leaf Spot class demonstrate its exceptional ability to differentiate Cercospora from other classes with almost no false negatives. The Golden Mosaic model yielded an F1-score of 99.83% with 100% precision and 99.67% recall overall. These findings highlight how the model has a very low risk of misclassification and is highly reliable in detecting even minor disease instances. The model’s total accuracy in detecting disease-free samples is demonstrated by its perfect classification for the Healthy Leaf class, which attained 100% precision, recall, and F1-score. This steady and excellent performance in every class confirms the model’s resilience and applicability for practical use in automated agricultural diagnostic systems.

4.3. Model Performance Comparison

The suggested DERIENet model performs noticeably better in testing and training than separate baseline models. Its test accuracy of 99.89% is excellent, and it consistently earns superior categorization results. DERIENet performs better in classification than EfficientNetB0 (98.56%), ResNet50 (96.78%), and InceptionV3 (97.89%). The accuracy of the ensemble overcomes even DenseNet201, which achieves a strong test accuracy of 98.67%. Compared to the suggested ensemble, conventional models like MobileNetV2, which attains 96.44%, and VGG16, which only reaches 90.22%, are less accurate and resilient. Furthermore, DERIENet maintains consistently high precision (99.89%), recall (99.89%), and F1-score (99.89%) throughout testing. The test findings show that it performs better than baseline models on all important performance parameters, highlighting its remarkable versatility. Regarding concerns about potential overfitting, several techniques beyond early stopping were employed during training to mitigate overfitting and improve generalization. These include the use of both L1 and L2 regularization (with L2 providing the maximum accuracy), dropout layers with a rate of 0.5, and significant data augmentation (GLOAR). Furthermore, early stopping was implemented according to validation loss, and the ReduceLROnPlateau callback was used to adaptively modify the learning rate. All these techniques improved the model’s capacity to generalize within the dataset and successfully decreased overfitting.

Table 3. Comparative evaluation of the models’ performance.

Model Name	Train Results			Test Results			Train Accuracy	Test Accuracy
	Precision	Recall	F1-Score	Precision	Recall	F1-Score
EfficientNetB0	99.80%	99.80%	99.80%	98.56%	98.56%	98.56%	99.83%	98.56%
ResNet50	96.54%	96.47%	96.44%	96.81%	96.78%	96.78%	97.17%	96.78%
InceptionV3	98.91%	98.90%	98.90%	97.89%	97.89%	97.88%	99.02%	97.89%
MobileNetV2	98.01%	97.97%	97.97%	96.44%	96.44%	96.44%	98.25%	96.44%
DenseNet201	99.07%	99.05%	99.05%	98.71%	98.67%	98.67%	99.27%	98.67%
VGG16	92.47%	92.45%	92.41%	90.49%	90.22%	90.16%	92.70%	90.22%
DERIENet (Proposed)	99.98%	99.98%	99.98%	99.89%	99.89%	99.89%	99.95%	99.89%

compiles the performance of several models, with particular attention to test accuracy, precision, recall, and F1-score. Using the same measures, Figure 12 compares the performance of these models and offers a visual representation of the data.

4.4. Ablation Study

The notable performance gains at each step of the model’s construction and augmentation process are demonstrated by the ablation research in

Table 4. Ablation analysis of DERIENet’s components in jute leaf disease detection.

Model	Test Classification Accuracy
EfficientNetB0 + InceptionV3	99.44%
EfficientNetB0 + ResNet50	99.67%
ResNet50 + InceptionV3	98.89%
EfficientNetB0 + ResNet50 + InceptionV3	85.71%
(Without GLOAR)
EfficientNetB0 + ResNet50 + InceptionV3	99.22%
(With L1 Regularization)
DERIENet (Proposed)	99.89%

. Two models combined resulted in more improvements, with EfficientNetB0 and InceptionV3 attaining 99.44% and EfficientNetB0 and ResNet50 reaching 99.67%. The combination of ResNet50 and InceptionV3 achieved an accuracy of 98.89%, which, while commendable, was lower in comparison to the other two-model combinations. However, evaluating the combined model of EfficientNetB0, ResNet50, and InceptionV3 without data augmentation (GLOAR) resulted in a significant drop in accuracy to 85.71% due to overfitting, highlighting the crucial importance of data augmentation.The accuracy increased to 99.22% when L1 regularization was used on the same three-model ensemble with augmentation (GLOAR). Finally, the proposed DERIENet model, which integrates all three architectures with a proposed data-augmentation technique (GLOAR) and L2 regularization, attained the maximum accuracy of 99.89%. This illustrates the strong benefits of model ensemble and augmentation in boosting performance and minimizing overfitting.

5. Discussion

One of the key variables in this work was the original dataset’s small size, which necessitated the use of calculated techniques to guarantee dependable model performance. A computer vision-based jute leaf disease classification system was developed using the dataset used in this work, which included high-resolution photos from three categories: Healthy Leaf, Golden Mosaic, and Cercospora Leaf Spot. Nevertheless, there are significant obstacles in developing strong deep-learning models on comparatively limited datasets. In order to boost dataset diversity and reduce the possibility of overfitting, we used GLOAR, a proposed data-augmentation technique. The augmentation methods decrease the possibility of overfitting while increasing the dataset [59]. A number of baseline models were first assessed for the classification of jute leaf disease, including ResNet50, InceptionV3, VGG16, DenseNet201, MobileNetV2, and EfficientNetB0. The three designs that performed the best were chosen for additional improvement using regularization approaches to lessen overfitting, batch normalization, and dense layers. The suggested DERIENet model was then created by combining these improved models using an ensemble technique.

This study evaluated the proposed DERIENet model through comprehensive experiments, where it consistently outperformed baseline models across various key performance indicators. As illustrated in Figure 7, the confusion matrices of every baseline model were examined in order to determine which models were most suited for the final ensemble. EfficientNetB0 properly classified 297 cases, demonstrating the best accuracy in detecting Cercospora Leaf Spot. With no misclassifications, InceptionV3 successfully classified Healthy Leaf images. Although DenseNet201 accurately classified all 300 Golden Mosaic samples, its significantly higher runtime made it less suitable for inclusion. ResNet50 was therefore selected for the ensemble because of its effectiveness and well-rounded performance. EfficientNetB0, InceptionV3, and ResNet50 were ultimately chosen for the ensemble due to their accuracy and lightweight design, which made them ideal for the DERIENet framework.

DERIENet’s performance evaluation further demonstrates how effective the selected ensemble architecture is. It outscored all baseline models in classification across all important evaluation metrics. As shown in Figure 10, DERIENet successfully classified 299 out of 300 test samples, with only one misclassification. Table 3 emphasizes the model’s impressive precision, recall, F1-score, and accuracy, all of which were recorded at 99.89%. Furthermore, as seen in Figure 13, the model showed a high degree of accuracy in classifying it.

As shown in

Table 5. Performance comparison between the model suggested in this paper and earlier research.

No.	Model	Ref.	Accuracy
1	four-layer CNN	Uddin et al. [7]	96.21%
2	CNN	Karim et al. [31]	95.45%
3	ResNet152 + Custom CNN	Akand et al. [33]	98.98%
4	ResNet50	Kaushik et al. [35]	96.50%
5	DCNN	Sourav et al. [36]	96.33%
6	DenseNet201	Talukder et al. [38]	98.21%
7	Lightweight CNN	Rana et al. [39]	95.43%
8	FL + CNN	Rajput et al. [42]	97.67%
9	DERIENet	Our Proposed Method	99.89%

, all models reported in prior research were evaluated using the Jute Leaf Disease dataset from this study. The results show that the suggested DERIENet model performs noticeably better than these current methods in terms of classification accuracy. In particular, when tested on the Jute Leaf Disease dataset, traditional architectures such as four-layer CNN [7], CNN [31], DCNN [36], and Lightweight CNN [39] obtained accuracies ranging from 95.43% to 96.33%. Conversely, transfer-learning models such as ResNet50 [35] and DenseNet201 [38] showed improved performance, 96.50% and 98.21%, respectively, but still failed to surpass our proposed model. More advanced methods, such as federated learning (FL) + CNN [42] and ResNet152 + Custom CNN [33], obtained accuracies of 97.67% and 98.98%, respectively. With a classification accuracy of 99.89%, our suggested DERIENet model outperformed all other approaches noticeably.

This study exclusively focused on three categories in jute leaf disease classification: Cercospora Leaf Spot, Golden Mosaic, and Healthy Leaf. The model was developed based on these diseases, but it has yet to be tested on additional jute diseases in other contexts. Incorporating a wider variety of jute diseases into the model could enhance its relevance in practical situations. The model would be more flexible and perform better in a range of agricultural settings if the dataset were expanded to include more varied disease kinds.

6. Conclusions

This study introduces DERIENet, a high-precision ensemble deep neural network model for automatically classifying diseases of jute leaves. The model incorporates InceptionV3, ResNet50, and EfficientNetB0 to use their combined capabilities and enhance classification performance. In order to provide resilience across a variety of disease patterns, the suggested data-augmentation technique (GLOAR) was used to enhance inference and avoid overfitting. In order to preserve model dependability while better representing disease variability, the dataset was enlarged. After evaluating six transfer-learning models in comparison, the top three models were chosen and fused using an ensemble technique to maximize performance. This led to DERIENet exceeding all baseline models in precision, recall, and F1-score, achieving an outstanding classification accuracy of 99.89%. The scalable and effective early jute disease detection solution offered by this high-performance model allows for prompt intervention and helps farmers reduce crop losses. A more sustainable and effective agricultural system could result from DERIENet’s ability to modernize jute farming, increase crop yields, and improve fiber quality by enabling quick and accurate disease identification.

Future studies should focus on enhancing DERIENet’s capacity to identify a greater variety of jute leaf diseases and evaluating how well it detects different jute plant infections to expand the model’s reach and applicability. Additionally, DERIENet’s incorporation of lightweight neural networks enhances its processing efficiency, making it appropriate for use in mobile apps. In instances with limited resources, this would enable real-time disease diagnosis, giving farmers a diagnostic option that is both accessible and reasonably priced. A mobile or web-based application would also make it simpler to engage with the model, enabling farmers and agricultural specialists to use DERIENet for effective disease management.