A Robust Lightweight Vision Transformer for Classification of Crop Diseases

Abstract

Rice, wheat, and maize are important food grains consumed by most of the population in Asian countries (like India, Japan, Singapore, Malaysia, China, and Thailand). These crops’ production is affected by biotic and abiotic factors that cause diseases in several parts of the crops (including leaves, stems, roots, nodes, and panicles). A severe infection affects the growth of the plant, thereby undermining the economy of a country, if not detected at an early stage. This may cause extensive damage to crops, resulting in decreased yield and productivity. Early safeguarding methods are overlooked because of farmers’ lack of awareness and the variety of crop diseases. This causes significant crop damage and can consequently lower productivity. In this manuscript, a lightweight vision transformer (MaxViT) with 814.7 K learnable parameters and 85 layers is designed for classifying crop diseases in paddy and wheat. The MaxViT DNN architecture consists of a convolutional block attention module (CBAM), squeeze and excitation (SE), and depth-wise (DW) convolution, followed by a ConvNeXt module. This network architecture enhances feature representation by eliminating redundant information (using CBAM) and aggregating spatial information (using SE), and spatial filtering by the DW layer cumulatively enhances the overall classification performance. The proposed model was tested using a paddy dataset (with 7857 images and eight classes, obtained from local paddy farms in Lalgudi district, Tiruchirappalli) and a wheat dataset (with 5000 images and five classes, downloaded from the Kaggle platform). The model’s classification performance for various diseases has been evaluated based on accuracy, sensitivity, specificity, mean accuracy, precision, F1-score, and MCC. During training and testing, the model’s overall accuracy on the paddy dataset was 99.43% and 98.47%, respectively. Training and testing accuracies were 94% and 92.8%, respectively, for the wheat dataset. Ablation analysis was carried out to study the significant contribution of each module to improving the performance. It was found that the model’s performance was immune to the presence of noise. Additionally, there are a minimal number of parameters involved in the proposed model as compared to pre-trained networks, which ensures that the model trains faster.

1. Introduction

Agriculture holds a historic significance in developing worldwide economic conditions by enabling the accumulation of wealth and resources. However, the quantity and quality of crop yields are greatly influenced by various factors, such as soil degradation, water scarcity, natural disasters, climate change, and plant diseases [1,2]. Among these, plant diseases are the leading cause of losses in crop yield, which can significantly threaten global food security. Rice is the most cultivated crop globally and requires a large amount of attention during its growth phase, as it is easily infected by various plant leaf diseases (such as bacterial leaf blight, brown spot, leaf blast, leaf scald, narrow brown spot, rice hispa, neck blast, and tungro), affecting its optimal growth and compromising its nutritional value [3,4]. This leads to yield losses ranging from 10% to 100%, depending on the growth phase of the crop and the degree of infestation [5]. The diagnosis of ill plants is often performed visually, which is time-consuming, imprecise, and impractical in large-scale applications, further exacerbated by the subjectivity of human assessment based on experience and expertise. This introduces inconsistency and variability during diagnosis, which leads to ineffective management strategies and unreliable results [6,7]. Furthermore, these diseases exert pressure on the environment, as farmers increasingly rely on heavy pesticide use with the aim of controlling disease outbreaks. This excessive dependence on chemical interventions engenders significant environmental pollution, ecosystem disruption, and adverse health consequences for both humans and wildlife. Therefore, early anomaly detection is essential in rice crops, which can help in taking the necessary steps to control the outbreak; this includes proper administration of pesticides and development of more resistant varieties [8,9].

Due to the recent advancements in computer vision and artificial intelligence (AI)-assisted techniques, automated disease detection that is highly accurate in classifying various leaf diseases has been made possible. Machine learning (ML) techniques such as K-Nearest Neighbor (KNN) and support vector machine (SVM) can easily achieve the aforementioned task but suffer from having a hand-crafted feature extraction step, which makes them less adaptive. On the other hand, convolutional neural networks (CNNs) under deep learning (DL) can precisely detect leaf diseases, as the features are extracted automatically without any human intervention [10,11]. However, these models experience difficulty when training on imbalanced datasets and facing environmental variability and are computationally complex in nature, making them nonviable for real-world applications.

Bacterial leaf blight (BLB), rice blast, brown spot, sheath blight, tungro, and bacterial leaf streak are common rice crop diseases affecting its yield. The pathogens responsible for the crop diseases discussed above collectively threaten global rice security, with climate change exacerbating disease dynamics. The descriptions of these diseases are discussed below. Modern solutions rely heavily on CRISPR-engineered resistance, microbiome modulation, and climate-smart surveillance systems to reduce the damage [12]. BLB is caused by Xanthomonas oryzae pv. oryzae. In this case, the pathogen infiltrates rice leaves via hydathodes or wounds, colonizing the xylem and forming water-soaked lesions that subsequently become necrotic, leading to leaf wilting and plant mortality at a later stage. BLB is particularly prevalent in tropical and subtropical regions, where warm, humid conditions facilitate its proliferation, often causing yield reductions of 20–50% [13]. The bacterium utilizes virulence mechanisms such as type III secretion systems (T3SSs) and effector proteins to suppress host immunity, thereby promoting systemic infection. Disease management necessitates an integrated approach, including the use of resistant varieties (e.g., those carrying Xa genes), chemical treatments, and cultural practices such as field sanitation and balanced fertilization. Recent advancements in genome editing (e.g., CRISPR-Cas9) present promising strategies for developing durable and resistant varieties [12]. However, climate change may exacerbate BLB outbreaks, necessitating ongoing research into adaptive control measures to protect global rice production. Rice blast (Magnaporthe oryzae), caused by fungal pathogens, causes severe yield losses in paddy crops. The fungus employs specialized pressure-driven appressoria to breach leaf cuticles, followed by invasive hyphal growth that rapidly colonizes tissues [14]. It is characterized by diamond-shaped spots with gray centers and dark margins, which often coalesce under optimal temperatures (i.e., 25–30 °C) with prolonged leaf wetness [15]. Brown spot (Bipolaris oryzae) is mainly caused by phosphorus-deficient soils and leads to reduced photosynthetic efficiency [16]. The pathogen can survive in crop residues and produces wind-dispersed conidia, which makes it persistent throughout the year. But, it is less severe than leaf blast. Sheath blight (Rhizoctonia solani) is problematic in rice due to its sclerotia-based survival mechanism. Modern high-yielding semi-dwarf varieties with dense canopies create ideal microclimates for disease progression [17]. The pathogen’s wide host range and resistance to common fungicides complicate control efforts and can reduce yields by up to 35% [18]. Tungro disease, caused by the RTBV-RTSV complex, continues to evolve through vector adaptation. Crops affected by recent outbreaks in Southeast Asia show altered symptoms, including more severe stunting and complete panicle sterility [19]. The green leafhopper vector (Nephotettix spp.) has developed resistance to common insecticides, necessitating integrated vector management [20]. Bacterial leaf streak (Xanthomonas oryzae pv. oryzicola) is emerging as a significant concern in mechanized farming systems, where harvesting injuries facilitate bacterial entry. New genomic studies reveal extensive horizontal gene transfer driving virulence evolution [13]. Unlike BLB, which causes systemic wilting, leaf streak remains localized, but it can reduce yields by 15–30% due to impaired photosynthesis.

Similar kinds of crop diseases occur in wheat and reduce productivity. The common diseases affecting wheat crops are black point, wheat blast, fusarium root rot, and leaf blight. Black point affects the quality and germination of wheat and is caused by Bipolaris sorokiniana [21]. It can be recognized by a brown or black tip on the grain embryo. Wheat blast is caused by Magnaporthe oryzae pathotype triticum. This fungus can affect various plant parts, including the leaf, awn, glume, spike, seed, and other ground tissues. Observable symptoms (such as silver spikes, dry spikes, black or dark gray sporulation by fungi, lesions, and wrinkled kernels) of this disease are noted during the reproductive phase of wheat [22]. Fusarium root rot is a soil-borne disease, which is caused by Fusarium pseudograminearum, F. culmorum, and F. graminearum. The fungus releases mycotoxin, which affects the stem and root portions of the crop (having visible symptoms such as increased lodging and impaired function in the vascular tissue). Infected plants display chocolate brown discoloration, which can extend to internodes along the stem [23,24]. Leaf blight is caused by the fungal pathogen Alternaria triticina and is characterized by symptoms such as choriotic lesions (on the leaves), blighted leaves, and a burnt appearance.

Therefore, diseases affecting paddy and wheat crops need to be detected accurately at their early stages to apply the necessary precautions to improve the crop yield. In the Indian subcontinent, the identification of crop diseases presents a significant challenge for farmers due to their lack of awareness and the similarity in symptoms among disease classes. So, development of a scalable, environmentally robust diagnostic system capable of precise detection and classification of crop leaf diseases with demonstrated efficacy across diverse environmental conditions, while maintaining scalability for large-scale deployment and integration into agricultural supply chains, is a top priority for advancement in current agricultural systems. By taking into account the current needs of the agricultural sector, a lightweight model based on a vision transformer (i.e., MaxViT) and a convolutional block attention mechanism (CBAM) is proposed. The method was evaluated on a customized eight-class paddy dataset and a five-class wheat dataset for detecting and classifying crop diseases.

The main contributions of this research are as follows:

• The developed architecture consists of CBAM, depth-wise separable (DW) convolution, squeeze and excitation (SE), and ConvNeXt modules. The CBAM module uses channel and spatial attention mechanisms to focus on minute and essential details from the feature map, while the SE module compresses the feature map in order to retain only the key channel that contributes to the identification of diverse disease attributes. Furthermore, the DW module captures the spatial features from each key channel separately, followed by improvement in channel contrast by using regularization (to enhance intricate details), which is aided by the ConvNeXt module.
• The modules collectively have 85 layers with 814.7K parameters. This model is found to be efficient compared to MobileNet, as well as other light networks reported in the literature. Due to its lightweight architecture, it can be implemented in real-time applications through edge devices.
• The developed model serves as an environmentally robust diagnostic system capable of precise detection and classification of crop leaf diseases, demonstrating efficacy across diverse environmental conditions while maintaining scalability for large-scale agricultural systems.
• To validate its effectiveness, robustness, and generalizability, the method has been tested on a publicly available wheat database and a custom paddy dataset (collected from Lalgudi district of Tiruchirapalli, Tamil Nadu, India) and attained overall testing accuracies of 98.47% (paddy) and 92.8% (wheat). Therefore, this model is highly effective for classifying multi-class diseases in paddy (eight classes) and wheat (five classes) crops.

The rest of the manuscript is organized as follows: Section 2 examines prior studies on plant disease detection using deep learning, identifying gaps that motivate this study. Section 3 elaborates on the model design, dataset characteristics, and evaluation strategies employed to assess performance. Section 4 presents experimental findings, including cross-fold validation, interpretability through Grad-CAM, comparisons with pre-trained models, ablation tests, and statistical validation. Finally, Section 5 concludes with reflections on the findings and suggests future directions for broader applicability and real-time deployment.

2. Related Works

Automatic discrimination of paddy crop diseases has been carried out in previous studies using image processing, ML, and DL techniques. ML techniques have employed feature extraction, followed by classification and performance evaluation for leaf disease categorization. Feature extraction captures only the fine details characterizing the disease symptoms and eliminates redundant data for image classification [25]. Color features (extracted using color moments and color histogram), texture features (extracted through gray level co-occurrence matrix (GLCM), Gabor filter, and local binary patterns (LBPs)), and shape features (extracted using elliptic Fourier and discriminant analysis and geometric calculation with moment invariants) [26] are some of the feature engineering techniques employed for leaf disease classification. Selecting the appropriate feature is very crucial for image classification [27]. Feature extraction can be carried out using hand-crafted or deep visual features. Hand-crafted features include shape, color, texture, etc., whereas deep visual features are extracted automatically using deep layers present in the deep neural network (DNN) [28]. ML models rely heavily on hand-crafted features, whereas DL models can extract them automatically. Automatic feature extraction in DL models is subject to the number of images available in the dataset. The larger the number of images in the dataset, the greater the number of visual features that can be captured by the deep layers [29]. In ML models, the features extracted are classified using ML classifiers for disease categorization. After classification, the performance metrics are evaluated.

A few of the recently available techniques for disease classification in paddy crops are represented below in

Table 1. A summary of methods utilized in the literature for identification of paddy crop diseases.

Sl. No.	Author (Year)	Techniques	Dataset (Images, Class)	Performance (Accuracy %)	Remarks
1	Bui et al. [30] (2025)	Ghost-Attention- YOLOv8	3563, 3	mAP-95.4	Limited classes; More parameters
2	Haque et al. [31] (2025)	ViT	12,322, 11	97.99	No generalizability
3	Pai et al. [32] (2025)	Twin CNN, Feature fusion	6763, 4	96.80	Computationally complex
4	Elakya et al. [33] (2024)	Ensemble model	13,876, 9	97.00	No cross-crop generalization
5	Bhola et al. [34] (2024)	Hybrid (DL and ML)	2788, 4	87.23	Computationally complex
6	Cao et al. [35] (2024)	Pyramid-YOLOv8	2011, 2	84.30	Limited classes; Poor performance
7	Altabaji et al. [36] (2024)	Modified LeafNet	2658, 3	87.76	Low performance; Computationally complex
8	Padhi et al. [37] (2024)	EfficientNet B4	23,916, 10	96.91	No cross-crop generalization
9	Zhang et al. [38] (2024)	ISMSFuse	ImgS-RBB2022 dataset, 2	98.14	Binary classification model
10	Singh et al. [39] (2024)	DL	6300, 2	98.70	Binary classification model
11	Senthil et al. [40] (2024)	DL, MCSVM	450, 4	96.85	Limited dataset; No generalization
12	Lu et al. [41] (2024)	YOLOv8-Rice	1200, 4	86.00	Low performance
13	Rai et al. [10] (2023)	Attention residual U-Net	479, 4	96.26	Limited dataset
14	Chen et al. [42] (2023)	DL, CPAM	2400, 4	95.25	No cross-crop generalization
15	Bharanidharan et al. [43] (2023)	MLOA	636, 5	90.00	Limited dataset

with their corresponding limitations and performance.

Bui et al. [30], Haque et al. [31], Pai et al. [32], Cao et al. [35], Altabaji et al. [36], Padhi et al. [37], etc., have used DL architectures to identify paddy crop diseases. Bui et al. [30] developed a Ghost-Attention-YOLOv8 DL architecture (consisting of CBAM, triplet attention, and efficiency multi-scale attention (EMA)) for the prediction of three paddy diseases from 3563 images collected from the Vietnam National University of Agriculture. The authors used a Ghost module, which minimizes the computational cost. This network has 5.5 Million learnable parameters. The triplet attention mechanism helps extract refined features. This method reported a mean average precision of 95.4%.

Haque et al. [31] developed a DL model with ViT (consisting of triplet multi-head attention) to identify leaf diseases in paddy and apple crops. The authors used the RicApp dataset (which consists of 12,322 images of rice crops (11 classes) and 7331 images of apple leaves (10 classes)). This method attained an overall accuracy of 97.99%.

Pai et al. [32] developed a twin CNN architecture for the classification of four classes (three diseased and one healthy) of paddy crops. The multi-scale features extracted from different pre-trained networks (VGG 16, ResNet50, and InceptionV3) were concatenated, followed by feature selection through principal component analysis (PCA). The features were then fed into the twin CNN architecture (with an identical structure) for classification. An overall accuracy of 96.8% was attained.

Elakya and Manoranjitham [33] developed an ensemble model (comprising ResNet, DenseNet, and EfficientNet) to classify nine paddy diseases. The authors used a dataset with 13,876 images collected from various repositories. The images were pre-processed using a Gaussian filter and adaptive thresholding and segmented by a Gaussian mixture model (GMM). This method resulted in an accuracy of 97% and an F1-score of about 90% for all nine rice diseases.

Bhola et al. [34] proposed a hybrid model with 20.2 M parameters for classification of multi-crop (such as corn, rice, and wheat) leaf diseases. The deep features from the crop images were extracted using the DenseNet201 model and fed into an SVM classifier. The dataset contained 2788 images (four classes each for corn, rice, and wheat) collected from various sources. This approach attained an accuracy of 87.23%.

Cao et al. [35] developed the Pyramid-YOLOv8 network for classification of rice blast disease. The dataset includes 2011 rice blast images captured using a mobile phone. The authors incorporated a detection head to capture tiny and fine target details, a CBAM for extracting multi-attention channel and spatial features, and a C2F-Pyramid module for feature extraction, achieving a low computational cost. This method obtained a mean average precision of 84.3%.

Altabaji et al. [36] proposed a modified LeafNet model for classification of three paddy diseases (brown spot, hispa, and leaf spot). The dataset has 2658 images collected from the Kaggle platform. The pre-processed images were augmented by rotation, flipping, height and width adjustments, and brightness variations. Modifications (consistent kernel size in the convolution layer and a modified max pooling layer) were applied to the original LeafNet model to improve its performance. The model’s performance was compared with the original LeafNet model and pre-trained networks (Xception and MobileNetV2). The proposed model attained accuracies of 97.44% and 87.76% for validation and testing, respectively.

Padhi et al. [37] developed an EfficientNet B4 model for classification of 10 paddy diseases using the Paddy Doctor dataset. The dataset consisted of 19131 and 4785 images for training and testing, respectively. The architectural features of this model included a compound scaling technique, the Swish activation function, and mobile inverted bottleneck convolution (MBConv) for better feature extraction and lower computational cost. The authors reported accuracies of 99.09% and 96.91% for training and testing, respectively.

Zhang et al. [38] proposed the ISMSFuse algorithm for classification of rice bacterial blight disease. The ImgS-RBB2022 dataset was utilized in this work. The features from the image were extracted by the MobileNetV2 model and fused with spectral features using the ISMSFuse algorithm and then classified by a multi-class linear SVM classifier. A prototype was developed and implemented on a Raspberry Pi for real-time detection of rice bacterial blight disease. The explainability of the model was analyzed by SHAP visualization. The authors reported an accuracy of 98.14%.

Singh et al. [39] performed classification of rice blast disease using DL models (such as AlexNet, VGG 16, LeNet, Xception, and InceptionV1). The dataset comprised 6300 images (both healthy and blast) captured using a digital camera from local paddy fields in Punjab, India. Among the pre-trained models used, AlexNet produced an accuracy of 98.7%.

Senthil et al. [40] developed a DNN incorporating a multi-class support vector machine (MCSVM) for four classes of paddy diseases. The dataset consists of 450 images sourced from the plantvillage platform. The images were pre-processed using a Gaussian filter and segmented by K-means clustering. The features were extracted using GLCM and classified using MCSVM. This approach achieved an accuracy of 96.85%.

Lu et al. [41] developed the YOLOv8-Rice model for the detection of four paddy diseases. The database consists of 1200 images collected from paddy fields. Twelve different attention mechanisms were used to focus on the target diseased regions. This method achieved a mean average precision score of 86%.

Chitranjan Kumar Rai and Roop Pahuja [10] developed an attention residual U-Net model for classification of four paddy diseases. The dataset comprised 479 images captured using a digital camera from different paddy fields in Orissa. The pre-processed images were augmented by rotation and contrast adjustment. Attention gates in the residual U-Net model performed semantic segmentation of the diseased regions from the leaf images. This method reported Dice coefficient scores of 0.9626 and 0.8192 for training and validation, respectively.

Chen et al. [42] introduced a novel attention mechanism for the detection of four paddy diseases. The dataset comprised 2400 images captured using a camera as well as collected from internet sources. The channel and position attention mechanism (CPAM) was used for deep feature extraction. These features were integrated with a domain adaption method for enhanced recognition of the diseased regions. This method achieved the highest recognition accuracy of 95.25%.

Bharanidharan et al. [43] developed a modified Lemurs optimization algorithm (MLOA) for categorization of five-class paddy images. The dataset has 636 thermal images obtained from the Kaggle platform. Statistical features (such as Box–Cox transformed features) were extracted and then classified by ML classifiers (such as KNN, RF, linear discriminant analysis (LDA), and histogram gradient boosting (HGB)). The authors reported a balanced accuracy of 90% for the proposed method with the KNN classifier.

From the above discussion, it is found that researchers used DL models, pre-trained networks, ensemble learning, attention mechanisms, vision transformer (ViT), etc., for leaf disease classification in crops and fruits. A few of the researchers used ensemble architectures for classification. These networks fused a unique approach derived from different pre-trained networks with traditional ML classifiers to improve performance. Researchers mainly used a few classes of crop diseases. The highest accuracy was reported for the two-class problem (obtained by Singh et al. [39]). The classification performance was drastically reduced for multiple classes. The introduction of ViT by Haque et al. [31] increased the performance to 97.99% for 11 classes. Various techniques adopted by researchers for detecting paddy crop diseases are summarized and presented in Table 1. The highest accuracy was reported to be 98.7% for two classes and 97.99% for multiple classes. There is a need to develop a light DL architecture for identifying diseases in crops and fruits effectively and efficiently.

Table 2. Techniques used in literature for identification of wheat crop diseases.

Sl. No.	Author, (Year)	Method	Dataset (Images, Class)	Performance (Accuracy (in %))	Remarks
1	Kumar et al. [45] (2025)	CaiT-YOLOv9	40,330, 7	94.85	Inaccurate for some diseases
2	Yasar et al. [44] (2025)	ML, WOA	8354, 5	95.2	No generalizability
3	Ruby et al. [46] (2024)	Modified ResNet50	4500, 4	98.44	Computationally complex
4	Chang et al. [47] (2024)	Improved DenseNet	5251, 4	98.32	No cross-crop generalization
5	Liu et al. [48] (2024)	CNN	3875, 4	mAP-60.8	Low performance; Less sensitive

summarizes the recent techniques used by researchers in the literature for identification of wheat crop diseases. An ML technique [44] and DL techniques [45,46,47,48] are adopted by the authors for identification of wheat crop diseases. Among these techniques, the modified ResNet50 architecture proposed by Ruby et al. [46] achieved the highest accuracy of 98.44% for classifying four wheat diseases.

3. Proposed Methodology

In this work, a multi-axis vision transformer (MaxViT), which is an extension of ViT, is used to identify paddy and wheat crop diseases. The ViT architecture was initially proposed by Dosovitskiy et al. [49] to address the limitations of CNNs in modeling global contextual relationships within visual data. By representing an image as a sequence of fixed-size patches and applying self-attention mechanisms, ViT captures long-range dependencies and holistic spatial interactions. It provides a robust and scalable alternative to CNNs by effectively capturing global features for improved performance in large-scale image recognition tasks.

Traditional ViTs and CNNs often face trade-offs between capturing fine-grained local details and efficiently modeling long-range dependencies. The MaxViT architecture proposed by Tu et al. [50] addresses this drawback by integrating convolutional token embeddings, block-wise grid attention, and dilated axial attention in a multi-axis attention framework. In this manuscript, a ViT architecture was proposed, considering CBAM, SE, DW, and the ConvNeXt modules.

The CBAM was proposed by Woo et al. [51] to enhance feature representation by integrating channel and spatial attention mechanisms into a lightweight, end-to-end trainable form. It sequentially infers attention maps along the channel and spatial dimensions, allowing networks to focus on informative features while suppressing irrelevant ones. The CBAM improves performance across various tasks, such as image classification, object detection, and segmentation, with minimal computational overhead.

The SE network was introduced by Hu et al. [52] to enhance the representational capacity of convolutional networks by modeling inter-channel dependencies. It achieves this through a two-step process, i.e., squeeze (global information embedding) and excitation (adaptive recalibration of channel-wise features). These processes allow the network to capture more informative features. SE modules are lightweight and easily integrable.

The DW block was used in MobileNet by Chollet et al. [53] to reduce the computational complexity of standard convolutions in DNNs. It decomposes a conventional convolution into two operations: depth-wise convolution (which applies a single filter per input channel) and point-wise convolution (which combines these outputs using $1\times 1$ convolutions). This factorization significantly reduces the number of parameters.

ConvNeXt, proposed by Liu et al. [54], enhances ResNet-like architectures using features such as large kernel convolutions, inverted bottlenecks, and layer normalization. This enhances feature learning and accuracy without improving computational efficiency, making it suitable for large-scale visual tasks.

The computational complexity of the original MaxViT means that the model has more parameters and may not be suitable for edge devices due to resource and power constraints. The complexity is reduced in this work by incorporating CBAM, DW, SE, and ConvNeXt modules. The modified architecture used the following strategies:

• Reduction in filter counts: All convolutional layers, including those within the attention and feedforward components, use only eight filters, in contrast to the higher-dimensional feature maps typically used in the standard MaxViT.
• Shallow convolutional stem: The stem block comprises only two $3\times 3$ convolutions with stride 2, minimizing early-stage computation while preserving essential spatial features.
• Simplified CBAM and SE modules: Attention mechanisms inspired by the CBAM and SE utilize low-dimensional fully connected layers and minimal activation operations (e.g., global pooling, rectified linear unit (ReLU), and sigmoid). These are implemented without increasing the number of channels or introducing multi-head attention.
• DW: Spatial convolutions are replaced with depth-wise separable convolutions to minimize both the number of operations and parameters.
• Compact ConvNeXt block: The ConvNeXt component is retained in a minimal form by reducing convolutional depth and maintaining residual connections, thereby ensuring feature learning without excessive computation.
• Reuse and residual connections: Intermediate outputs are reused through skip connections and addition layers to facilitate gradient flow and reduce the need for deeper or wider layers.
• Global average pooling: Instead of fully connected layers for spatial flattening, global average pooling (GAP) is used to reduce feature dimensions before classification, further minimizing the parameter count.

The description of the various modules and their functionalities in the proposed architecture for crop disease classification is presented in

Table 3. Description and functionalities of various modules in the proposed DL model for crop disease classification.

Sl. No.	Module	No. of Layers	Purpose and Strength
1.	CBAM	9	Channel and spatial attention; Concentrates only on disease-relevant regions; Enhances feature refinement.
2.	DW	4	Channel-specific feature learning; Ensures minimal overhead in network scaling; Preserves attention-weighted features from CBAM.
3.	SE	8	Learns inter-channel dependencies; Emphasizes critical feature channels.
4.	ConvNeXt	4	Convolutional feature extraction; Captures hierarchical features; Integrates normalization, activation, and residual connection for deeper representation.
5.	Proposed model	85	Lightweight architecture suitable for real-time detection; Robust performance in multi-class crop disease classification.

.

The proposed DNN architecture for crop (paddy and wheat) disease classification consists of a stem block with convolutional layers, a MaxViT module (two stages), and a classification head. The proposed DNN has a lightweight architecture (i.e., 814.7 K learnable parameters and 85 layers) and incurs low computational cost compared to existing pre-trained networks and ViTs. A block diagram illustrating the sequential stages of the proposed DNN model is depicted in Figure 1a. The significance and functionality of each block used in this model are described below. The proposed DL network consists of two stages of the MaxViT module, as presented in Figure 1b. The MaxViT module comprises convolution blocks, a CBAM module, a DW block, another CBAM module, max pooling, a resize layer, and an SE module arranged sequentially.

The CBAM module enhances feature representation by preserving only the important features across channels while eliminating redundant/unnecessary features. It ensures that attention is focused on features that contribute to disease classification. The convolution process, followed by a skip connection, extracts the local and global features and feeds them into the GAP layer to compress the spatial features. Sigmoid activation generates the channel-wise attention weights. The attention vector is up-scaled using a resize layer, thus aligning it with the original feature map size. The CBAM module focuses only on salient features through its attention mechanism and aids the DNN model in achieving better discrimination of crop diseases. The CBAM module is represented in Figure 1c. Considering $X\in {\mathbb{R}}^{C\times H\times W}$ as the input feature map and $X^{″}$ as the final refined output, the overall CBAM can be represented as

$$X^{\prime }=M_c\left(X\right)\odot X$$

(1)

$$X^{″}=M_s\left(X^{\prime }\right)\odot X^{\prime }$$

(2)

where $X\in {\mathbb{R}}^{C\times H\times W}$ is the input feature map, $X^{\prime }\in {\mathbb{R}}^{C\times H\times W}$ is the feature map after channel attention, $X^{″}\in {\mathbb{R}}^{C\times H\times W}$ is the final feature map after channel and spatial attention, $M_c\left(X\right)\in {\mathbb{R}}^{C\times 1\times 1}$ denotes the channel attention map, $M_s\left(X^{\prime }\right)\in {\mathbb{R}}^{1\times H\times W}$ denotes the spatial attention map, and ⊙ denotes element-wise multiplication.

The channel attention map $M_c$ can be expressed as

$$M_c\left(X\right)=\sigma \left(\mathrm{MLP}\left(X_{\mathrm{avg}}^c\right)+\mathrm{MLP}\left(X_{\max }^c\right)\right)$$

(3)

where $X_{\mathrm{avg}}^c$ is the average pooled feature map along the spatial dimensions of X, $X_{\max }^c$ is the max-pooled feature map along the spatial dimensions of X, and $\sigma$ denotes sigmoid activation.

The spatial attention output is formulated as

$$M_s\left(X^{\prime }\right)=\sigma \left(\mathrm{Conv}\left([X_{\mathrm{avg}}^s;X_{\max }^s]\right)\right)$$

(4)

where $X_{\mathrm{avg}}^s\in {\mathbb{R}}^{1\times H\times W}$ and $X_{\max }^s\in {\mathbb{R}}^{1\times H\times W}$ are the 2D maps produced by channel attention, ‘Conv’ denotes a convolutional layer, and $[·;·]$ represents channel-wise concatenation. Equations (1)–(4) are adapted from the CBAM formulation by Woo et al. [51], with notational adjustments to align with our implementation. In this work, the CBAM is modified, and the final output of this block is represented by $X+F\left(X^{″}\right)$, where $F\left(X^{″}\right)$ is the transformed output obtained by passing the signal through GAP, FC, ReLU, sigmoid, and resize blocks. It is represented by $F\left(X^{″}\right)=\Gamma \left(X+F\left(X^{\prime }\right)\right)$, where $\Gamma$ is the transformation function. $F\left(X^{\prime }\right)$ is the output obtained by passing the signal via two-stage convolution, i.e., $F\left(X^{\prime }\right)=\Gamma \left(X\right)$. The two-convolution block performs channel attention. Spatial attention is performed by GAP, ReLU, sigmoid, and resize blocks.

The SE module (depicted in Figure 1d) aggregates spatial information and learns channel-wise dependencies. Sigmoid activation generates per-channel gating weights in [0, 1], which is applied to the feature maps to recalibrate channel importance. For any input feature map $U\in {\mathbb{R}}^{C\times H\times W}$, the operations in the SE network are represented as follows:

$$V_1=\delta \left({\mathrm{BN}}_1\left({\mathrm{Conv}}_1\left(U\right)\right)\right)$$

(5)

$$V_2=\mathrm{MaxPool}\left(V_1\right)$$

(6)

$$V_3={\mathrm{BN}}_2\left({\mathrm{Conv}}_2\left(V_2\right)\right)$$

(7)

$$S=\sigma \left(V_3\right)$$

(8)

$$X_{\mathrm{SE}}=S\oplus U$$

(9)

where $U\in {\mathbb{R}}^{C\times H\times W}$ is the input feature map; $V_1$ is the output after the first convolution, batch normalization, and ReLU activation; $V_2$ is the downsampled feature map after max pooling; $V_3$ is the output after the second convolution and batch normalization; $S\in {\mathbb{R}}^{C\times 1\times 1}$ is the channel-wise attention vector obtained after sigmoid activation; $X_{\mathrm{SE}}\in {\mathbb{R}}^{C\times H\times W}$ is the output feature map of the SE block; ${\mathrm{Conv}}_1,{\mathrm{Conv}}_2$ denote convolutional layers (typically $1\times 1$ convolutions), ${\mathrm{BN}}_1,{\mathrm{BN}}_2$ denote batch normalization layers, $\delta (·)$ is ReLU activation, $\mathrm{MaxPool}(·)$ represents the max pooling operation, $\sigma (·)$ is sigmoid activation, and ⊕ denotes element-wise addition.

Equations (5)–(9) represent a modified version of the SE mechanism formulated by Hu et al. [52]. The original SE block utilizes GAP followed by fully connected layers. But, the proposed architecture introduces convolutional layers, batch normalization, and max pooling to enhance spatial encoding prior to channel recalibration. This module uses element-wise addition in place of element-wise multiplication.

The DW layer (represented in Figure 1e) applies spatial filtering to each channel, allowing lightweight, channel-specific feature learning without cross-channel mixing. It ensures minimal overhead in scaling the network, thereby preserving attention-weighted features from the CBAM block. Given an input feature map $F\in {\mathbb{R}}^{C\times H\times W}$, the output of the depth-wise convolution block can be formulated as follows:

$$F_{\mathrm{out}}=\mathcal{B}\left(\mathcal{D}\left(\mathcal{P}\left(F\right)\right)\right)$$

(10)

where $F\in {\mathbb{R}}^{C\times H\times W}$ is the input feature map, $\mathcal{P}(·)$ denotes the max pooling operation, $\mathcal{D}(·)$ denotes the depth-wise convolution operation, $\mathcal{B}(·)$ denotes the batch normalization operation, and $F_{\mathrm{out}}$ is the output feature map of the DW block. Equation (10) represents a lightweight feature extraction block inspired by the depth-wise convolutional concept introduced by Chollet et al. [53] in MobileNet, with adaptations (preceding pooling step and batch normalization) for lightweight features in the proposed architecture.

The ConvNeXt module (consisting of convolution, batch normalization, and ReLU activation layer) helps capture the hierarchical features for better discrimination of crop diseases. The ConvNeXt block used in the proposed DNN model is illustrated in Figure 1f. The output of the ConvNeXt block for an input feature map $Y\in {\mathbb{R}}^{C\times H\times W}$ can be formulated as follows:

$$Y_{\mathrm{out}}=Y+\delta \left(\mathrm{BN}\left(\mathrm{Conv}\left(Y\right)\right)\right)$$

(11)

where $Y\in {\mathbb{R}}^{C\times H\times W}$ is the input feature map, $\mathrm{Conv}(·)$ denotes the convolution operation, $\mathrm{BN}(·)$ denotes the batch normalization operation, $\delta (·)$ denotes the activation function (ReLU), ⊕ denotes element-wise addition (skip connection), and $Y_{\mathrm{out}}$ is the output feature map of the ConvNeXt block. Equation (11) represents a modified variant of the ConvNeXt block, adapted from the architecture originally proposed by Liu et al. [54]. The original formulation includes depth-wise convolutions, layer normalization, and GELU activation. But, the proposed ConvNext architecture employs standard convolution, batch normalization, and ReLU for computational efficiency in the target domain. Equations associated with the proposed MaxViT architecture are mentioned in Equation (12). In this case, I and $I_{out}$ represent input and output of the MaxViT module, respectively.

$$\left.\begin{array}{c}I=Input\left(MaxViT\right)\hfill \\ I_1=Conv\left(Conv\left(I\right)\right)\hfill \\ I_2=CBAM\left(I_1\right)\hfill \\ I_3=DW\left(I_2\right)\hfill \\ I_3=B\left(D\left(P\left(I_2\right)\right)\right)\hfill \\ I_4=CBAM\left(I_3\right)\hfill \\ I_5=MaxPool\left(I_4\right)\hfill \\ I_6=Resize\left(I_5\right)\hfill \\ I_7=I_6+I\hfill \\ I_8=SE\left(I_7\right)\hfill \\ I_9=ConvNext\left(I_8\right)\hfill \\ I_{out}=I_8+\delta \left(BN\left(Conv\left(I_8\right)\right)\right)\hfill \end{array}\right\}$$

(12)

3.1. Dataset

The experimental evaluation of the proposed approach was conducted using images sourced from publicly accessible repositories and real-field samples collected from local paddy fields in the Lalgudi region of Tiruchirapalli, Tamil Nadu, India. Images of paddy crop diseases were partly sourced from the Mendeley data platform, comprising four categories, i.e., bacterial blight, blast, brown spot, and tungro, with a resolution of $300\times 300$ pixels. These were combined with locally acquired field images and categorized into eight disease classes, resulting in a unified dataset of 7857 samples. To ensure the reliability and integrity of the newly curated paddy crop disease dataset, a systematic validation process was undertaken.

Labeling accuracy: All images were initially labeled based on expert agronomist annotations and then cross-verified by a secondary reviewer to ensure consistency. Disagreements regarding labels were resolved through consultation with domain experts. This two-stage verification minimized labeling errors and enhanced the overall annotation quality.

Class distribution: The dataset consists of eight distinct classes of paddy crop images, including seven disease categories and one healthy class. The distribution of images was examined to avoid severe class imbalance. Although minor variations exist due to natural data availability, the class frequencies were sufficiently balanced to enable fair learning across categories. Detailed statistics for the dataset are provided in

Table 4. Details of custom paddy crop disease dataset.

Sl. No.	Class	No. of Images	Description
1	Bacterial blight	981	Collected from local paddy fields. Camera specifications: - Canon EOS 1200D (Canon Inc., Tokyo, Japan), - $3456\times 2304$ pixels, - 55 mm focal length Augmentation: - Rotation ($-5^{\circ }$ to $+5^{\circ }$)
2	Brown spot	963
3	Leaf blast	993
4	Narrow brown spot	975
5	Neck blast	981
6	Rice hispa	989
7	Tungro	977
8	Healthy leaves	998
	Total images	7857

.

Image quality assessment: To evaluate image clarity, resolution, and content relevance, all images were visually inspected, and images with low-resolution samples, noisy backgrounds, and duplicates were removed. Only images with sufficient visual information regarding the symptomatic region of the plant were retained. This ensured high-quality inputs to the training model and mitigated the risk of misleading patterns. These validation steps collectively enhance the reliability of the dataset and ensure its suitability for paddy disease classification. A summary of the dataset is provided in Table 4.

The generalizability of the model for prediction of other crop diseases is evaluated using a wheat disease dataset (5000 images, 5 classes) sourced from the Kaggle platform. The distribution of images in the wheat dataset is presented in

Table 5. Dataset description for wheat crop diseases.

Sl. No.	Class	No. of Images	Description
1	Black point	1000	Downloaded from Kaggle platform. Image resolution: - $1080\times 1440$ pixels Augmentation: - Rotation ($-5^{\circ }$ to $+5^{\circ }$)
2	Blast	1000
3	Fusarium root rot	1000
4	Leaf blight	1000
5	Healthy	1000
	Total images	5000

.

Each image of paddy or wheat illustrates the distinct visual patterns, color variations, and lesion characteristics associated with each disease class. The diseased regions are highlighted with a pink color boundary. These images represent the visual complexity and similarity of different disease types. Images corresponding to each class of paddy and wheat are shown in Figure 2 and Figure 3.

To evaluate the impact of the train–test split ratio on model performance, experiments were carried out with four different configurations: 90:10, 80:20, 75:25, and 70:30. For each train–test split ratio, the images in the dataset were randomly shuffled. Among the various train–test ratios, the 80:20 split ratio provided the highest training and testing accuracies, with a minimal difference between them. Other train–test ratios resulted in lower performance. Therefore the 80:20 train–test ratio was adopted for performing the final experiments.

3.2. Performance Metrics

The classification ability of the proposed network in identifying diseases in paddy and wheat crops was systematically evaluated using a comprehensive set of performance metrics. Key indicators include the Matthews correlation coefficient (MCC) and several measures derived from the confusion matrix. The confusion matrix provides the distribution of prediction outcomes—namely, true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), which are used to compute accuracy, sensitivity (recall), specificity, mean accuracy, precision, and F1-score [55,56,57]. These metrics collectively provide a robust assessment of the model’s effectiveness, ensuring its suitability for deployment in precision agriculture settings. The evaluation metrics, along with their corresponding mathematical expressions, are summarized in

Table 6. Performance metrics.

Sl. No.	Metrics	Mathematical Representation	Range	Remarks
1	Accuracy (Acc)	${\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$	[0, 1]	Classification success rate; Fails to predict individual class performance
2	Sensitivity (recall)	${\displaystyle \frac{TP}{TP+FN}}$	[0, 1]	Rate of correctly predicted positive instances
3	Specificity	${\displaystyle \frac{TN}{TN+FP}}$	[0, 1]	Rate of correctly predicted negative instances
4	Mean accuracy (MAcc)	${\displaystyle \frac{\mathrm{Sensitivity}+\mathrm{Specificity}}{2}}$	[0, 1]	Identify classification performance for both positive and negative classes
5	Precision	${\displaystyle \frac{TP}{TP+FP}}$	[0, 1]	Correctly identified positive samples to total predicted positive samples
6	F1-score	${\displaystyle \frac{2\ast TP}{\left(\right(2\ast TP)+FP+FN)}}$	[0, 1]	Harmonic mean of precision and recall
7	MCC	${\displaystyle \frac{(TP\ast TN)-(FP\ast FN)}{\sqrt{(TP+FP)\ast (TP+FN)\ast (TN+FP)*(TN+FN)}}}$	[$-1$, 1]	Correlation between observed and predicted

Note: $TP$: True positive; $TN$: true negative; $FP$: false positive; $FN$: false negative. Higher metric value represents better classification performance.

.

4. Results and Discussion

The proposed DNN model, consisting of MaxViT, CBAM, SE, DW, and ConvNeXt modules, was developed in MATLAB R2023b installed on a machine having the hardware configuration described in

Table 7. Hardware and software configurations.

Sl. No.	Resource	Specifications
1	CPU	Intel (R) Core (TM) i7-14700 processor @ 2.10 GHz, with 32 GB RAM
2	GPU	NVIDIA T1000 GPU with 4 GB RAM
3	System type	64-bit operating system, x64-based processor
4	Operating system	Windows 11 Pro
5	Software	MATLAB R2023b

.

Experimental evaluation for the prediction of crop diseases was carried out by adjusting the learning rates (from 0.00001 to 0.001), varying the epochs (from 2 to 6 with a step size of 2), batch size (32), L2 regularization, early stopping (enabled), and the Adam optimizer. The classification accuracies with different hyperparameters are listed in

Table 8. Experimental evaluation of the proposed DNN architecture for the classification of wheat and paddy crop diseases.

Sl. No.	Optimizer	Epochs	Learning Rate	Paddy		Wheat
				Overall Accuracy (%)	Training Time (minutes)	Overall Accuracy (%)	Training Time (minutes)
1	Adam	2	0.001	88.80	4.67	82.30	2.37
			0.0001	91.59	4.47	89.30	2.40
			0.00001	82.76	4.45	72.00	2.47
2	Adam	4	0.001	94.34	8.57	86.30	5.03
			0.0001	97.07	8.28	91.20	4.98
			0.00001	92.75	8.57	80.20	4.82
3	Adam	6	0.001	97.52	16.68	91.70	7.48
			0.0001	99.43	12.62	94.00	6.42
			0.00001	94.27	12.65	84.40	6.62

.

From Table 8, it is inferred that the highest accuracies, 99.43% (for paddy dataset) and 94% (for wheat dataset), were attained using six epochs with a learning rate of 0.0001 and the Adam optimizer.

The class-wise performance metrics (such as accuracy, sensitivity, specificity, precision, F1-score, and MCC) were evaluated on the paddy and wheat datasets with the selected hyperparameters (epoch: 6; LR: 0.0001; optimizer: Adam; and batch size: 32). During the training process, overall accuracies of 99.43% and 94% were reported for the paddy and wheat datasets. For the paddy dataset, the highest performance (100%) was attained for the brown spot, leaf blast, narrow brown spot, and neck blast classes, whereas the lowest accuracy (99.43%) was reported for the bacterial blight class. When experimenting on the wheat dataset, the highest accuracy (98.20%) was recorded for the blast and healthy classes, whereas the lowest accuracy was observed for fusarium root rot. The performance metrics for the paddy and wheat datasets during training are presented in

Table 9. Comparison of results on crop datasets during training.

Sl. No.	Species	Class	Accuracy (%)	Sensitivity (%)	Specificity (%)	Mean Accuracy (%)	Precision (%)	F1-Score (%)	MCC ($-1$ to +1)	Overall Accuracy (%)
1.	Paddy	Bacterial blight	99.43	95.92	99.93	97.93	99.47	97.66	0.9736	99.43
		Brown spot	100	100	100	100	100	100	1
		Healthy	99.94	100	99.93	99.97	99.5	99.75	0.9972
		Leaf blast	100	100	100	100	100	100	1
		Narrow brown spot	100	100	100	100	100	100	1
		Neck blast	100	100	100	100	100	100	1
		Rice hispa	99.75	100	99.71	99.86	98.02	99	0.9886
		Tungro	99.75	99.49	99.78	99.64	98.48	98.98	0.9884
2.	Wheat	Black point	97.80	90.00	99.75	94.88	98.90	94.24	0.9304	94
		Blast	98.20	98.00	98.25	98.13	93.33	95.61	0.9452
		Fusarium root rot	96.80	91.00	98.25	94.63	92.86	91.92	0.8993
		Healthy	98.20	95.00	99.00	97.00	95.96	95.48	0.9436
		Leaf blight	97.00	96.00	97.25	96.63	89.72	92.75	0.9095

.

Table 10. Comparison of results on crop datasets during testing.

Sl. No.	Species	Class	Accuracy (%)	Sensitivity (%)	Specificity (%)	Mean Accuracy (%)	Precision (%)	F1-Score (%)	MCC ($-1$ to +1)	Overall Accuracy (%)
1.	Paddy	Bacterial blight	99.36	96.94	99.71	98.33	97.94	97.45	0.9707	98.47
		Brown spot	99.87	98.96	100	99.48	100	99.48	0.9941
		Healthy	99.87	99	100	99.50	100	99.5	0.9943
		Leaf blast	98.98	100	98.83	99.42	92.52	96.12	0.9563
		Narrow brown spot	99.75	100	99.71	99.86	97.98	98.98	0.9884
		Neck blast	100	100	100	100	100	100	1
		Rice hispa	100	100	100	100	100	100	1
		Tungro	99.11	92.78	100	96.39	100	96.26	0.9947
2.	Wheat	Black point	97.60	96.00	98.00	97.00	92.31	94.12	0.9264	92.8
		Blast	97.00	88.00	99.25	93.63	96.70	92.15	0.9045
		Fusarium root rot	97.20	95.00	97.75	96.38	91.35	93.14	0.9141
		Healthy	96.60	86.00	99.25	92.63	96.63	91.01	0.8915
		Leaf blight	97.20	99.00	96.75	97.88	88.39	93.40	0.9186

provides the classification performance metrics (class-wise) evaluated on the paddy and wheat datasets during the testing process. Overall testing accuracies of 98.47% and 92.8% were recorded for the paddy and wheat datasets. In the testing process, the highest performance (100%) in all metrics was attained for the neck blast and rice hispa classes. The lowest classification accuracy, 98.98%, was noted for the leaf blast class in the paddy dataset. When experimenting on the wheat dataset, the highest accuracy (97.6%) was obtained for fusarium root rot and the lowest accuracy (96.60%) was reported for the healthy class. The confusion matrix plots obtained during testing on the paddy and wheat databases are illustrated in Figure 4.

Spider charts are presented to display multiple evaluation metrics (such as accuracy, sensitivity, specificity, MAcc, precision, and F1-score) together to compare the overall performance for all classes at a glance. The radial visualization highlights variations in diagnostic efficacy. Spider charts for both training and testing of the model are presented in Figure 5.

All the metrics were found to be above 90%. The gap between training and testing performance was very small, representing the ideal classification model for classifying paddy crop diseases. During testing, the model demonstrated reduced sensitivity and precision for tungro and leaf blast, relative to the remaining paddy crop disease classes. The overlapping of training and testing accuracies in the spider chart reveals model consistency and generalization gaps. Similarly, the spider charts for wheat crop disease classification for five classes are presented in Figure 6. Visualization of the performance metrics shown in Figure 6 illustrates the model’s classification ability for other databases. Sensitivity and precision values were found to be less than 90% for the blast and healthy classes, respectively.

The training and validation loss curves demonstrate the model’s convergence behavior and its robustness to over-fitting. These plots provide clear evidence that the model converged effectively and maintains a consistent gap between training and validation losses, indicating that over-fitting was minimal. The training and validation loss curves are presented below in Figure 7.

4.1. Cross-Fold Validation

To evaluate the generalization capability and robustness of the proposed model, a five-fold cross-validation strategy was adopted. This step was taken to ensure that the performance estimates were not biased by a single train–test split ratio, especially for datasets with limited sample sizes or class imbalance. Initially, the custom paddy dataset comprising eight distinct classes was randomly partitioned into five equal and mutually exclusive folds, maintaining the original class distribution in each fold (i.e., stratified splitting). In each of the five iterations, one fold was reserved as the validation set, and the remaining four folds were used for model training. This procedure ensured that every image in the fold within the dataset was used once for validation and four times for training. The model was trained and then evaluated on the corresponding validation set. The performance metrics (such as accuracy, sensitivity, specificity, precision, F1-score, and MCC) were recorded individually for each class in every fold. Subsequently, for each metric, both per-class averages and macro-averaged scores across all folds were calculated. The mean and standard deviation (SD) of each metric across the five folds are reported in

Table 11. Performance evaluation metrics of the proposed model for 5-fold cross-validation.

S. No.	Class in Paddy Crop	Sensitivity (%)	Specificity (%)	Precision (%)	F1-Score (%)	MCC ($-1$ to +1)	Overall Testing Accuracy (%)
1.	Bacterial blight	98.57 ± 01.99	99.03 ± 00.63	93.70 ± 03.87	96.01 ± 01.31	0.9550 ± 0.0146	97.31 ± 01.63
2.	Brown spot	92.22 ± 08.13	99.96 ± 00.06	99.67 ± 00.50	95.64 ± 04.51	0.9528 ± 0.0476
3.	Healthy	97.30 ± 06.04	99.65 ± 00.49	97.74 ± 03.09	97.39 ± 02.97	0.9710 ± 0.0325
4.	Leaf blast	99.50 ± 00.87	99.78 ± 00.24	98.53 ± 01.57	99.00 ± 00.70	0.9886 ± 0.0080
5.	Narrow brown spot	100.00 ± 00.00	99.59 ± 00.50	97.30 ± 03.28	98.61 ± 01.69	0.9843 ± 0.0191
6.	Neck blast	100.00 ± 00.00	99.96 ± 00.10	99.70 ± 00.67	99.85 ± 00.34	0.9983 ± 0.0039
7.	Rice hispa	95.84 ± 06.41	99.30 ± 01.48	96.03 ± 08.33	95.60 ± 04.60	0.9518 ± 0.0501
8.	Tungro	94.98 ± 03.09	99.67 ± 00.25	97.63 ± 01.76	96.25 ± 01.27	0.9577 ± 0.0142
	Macro average	97.30 ± 01.63	99.62 ± 00.23	97.54 ± 01.37	97.29 ± 01.64	0.9699 ± 0.0178

to assess the consistency and reliability of the model’s performance across different data splits. This cross-validation approach provided a statistically sound evaluation of the model in multi-class settings, where each class’s contribution to the overall performance was equally weighted through macro-averaging.

The performance metrics summarized in Table 11 indicate that the proposed model achieved consistently high classification performance across all eight rice disease classes. With a macro-averaged sensitivity of 97.30%, specificity of 99.62%, and F1-score of 97.29%, the model demonstrated strong capability in both identifying diseased samples and minimizing false detections. The high value of macro-averaged MCC (i.e., 0.9699) further reflected robust predictive reliability across class distributions. These results affirm the model’s effectiveness and generalizability for multi-class rice disease classification. The low standard deviations further suggested the stability and consistency of the model across different validation folds.

4.2. Grad-CAM Visualization

Gradient-weighted class activation mapping (Grad-CAM) is an explainable AI technique proposed by Selvaraju et al. to highlight the class-discriminative regions in convolutional neural networks using the gradients of target outputs. It generates localization maps that visually explain model decisions, aiding in the interpretability of deep models. Grad-CAM is particularly effective in image classification and object detection by providing spatial insight into feature importance. Grad-CAM visualizations are generated by computing the gradients of the predicted class score in relation to the activation maps from the final convolutional layer. It utilizes a color-coded scheme to convey the significance of different regions. The colors are determined by the magnitude of the gradient values, which reflect the contribution of each region to the predicted class. Regions exhibiting high gradient values are depicted in red, indicating a substantial contribution to the predicted class. Regions with moderate gradient values are marked in yellow/green, signifying a moderate level of importance to the predicted class. Regions with low gradient values are represented in blue, signifying low importance, denoting a relatively minor contribution to the predicted class.

In this manuscript, Grad-CAM was used to identify the regions contributing to crop disease classification by the proposed model. To facilitate clear qualitative comparison, the original image, along with its corresponding Grad-CAM heatmap (highlighting the key regions used by the model for crop disease classification), is presented in Figure 8.

4.3. Comparative Analysis with Pre-Trained Model

The robustness of the model in the classification of crop (paddy and wheat) diseases is analyzed by comparing the proposed DNN with several existing pre-trained networks (such as AlexNet, GoogLeNet, ResNet50, MobileNetV2, InceptionV3, VGG-16, ResNet18, and SqueezeNet) [58]. The results are compared, along with their parameters (learnable parameters, number of layers, accuracy, and training time), in

Table 12. A comparison of the proposed method with existing pre-trained networks.

Sl. No.	Network	Learnable Parameters	Layers	Paddy (8 Classes)		Wheat (5 Classes)
				Accuracy (%)	Training Time (minutes)	Accuracy (%)	Training Time (minutes)
1	AlexNet	60.9 M	25	99.75	14.65	94.68	8.20
2	GoogLeNet	6.9 M	144	100	18.10	95.12	12.30
3	ResNet50	25.5 M	177	99.76	67.65	96.87	33.23
4	MobileNetv2	3.5M	154	100	24.95	96.95	13.15
5	Inceptionv3	23.8 M	315	100	77.85	96.12	42.76
6	VGG-16	134.2 M	41	98.85	422.37	93.26	208.45
7	ResNet18	11.6 M	71	99.68	15.75	95.06	9.70
8	SqueezeNet	1.2 M	68	99.68	13.80	94.60	7.68
9	Proposed	814.7 K	85	99.43	12.62	94.00	6.42

.

Among the pre-trained networks, SqueezeNet and VGG-16 have the minimum (1.2 M) and maximum (134.2 M) number of learnable parameters. On the other hand, the proposed model integrated with the MaxViT transformer has far fewer parameters (i.e., 814.7 K), and hence, training takes much less time (i.e., 12.62 min) compared to the other pre-trained networks. Similarly, for the wheat dataset, the proposed model attained 94% accuracy with a training time of 6.42 min. From the above comparative analysis, it can be inferred that the proposed DNN architecture offers a lightweight solution for categorization of crop diseases with less training time, making it suitable for real-time implementation. The comparison of accuracy and training time of various pre-trained models with the proposed method is visually represented in Figure 9.

4.4. Ablation Study

Ablation studies rigorously evaluate the model components by removing or modifying them to assess their impact on performance. This method isolates critical elements, distinguishing essential features from redundant ones. Such analyses enhance interpretability and ensure robust, well-justified model architectures. The ablation study results presented in

Table 13. An ablation study of the proposed DNN.

Sl. No.	Ablation Study	No. of Layers	Total Learnable Parameters (K)	Accuracy (%)
				Paddy	Wheat
1	Removal of SE block and ConvNeXT block from MaxViT (1 stage)	34	807.3	84.61	86.90
2	Removal of ConVNeXt block from MaxViT (1 stage)	42	808.5	86.32	89.12
3	Removal of all SE and ConvNeXt blocks from MaxViT (2 stages)	60	811.1	86.77	89.86
4	Removal of all CBAM blocks from MaxViT (2 stages)	47	809.0	83.65	83.80
5	With only 1 stage of MaxViT	46	809.1	88.10	87.80
6	Proposed (with 2 stages of MaxViT)	85	814.7	99.43	94.00

demonstrate how different architectural components contribute to the performance of MaxViT for paddy disease classification. The classification accuracy (with two-stage MaxViT block) for paddy and wheat crops was 99.43% and 94%, respectively. Eliminating one stage of MaxViT reduces the accuracy to 88.10% and 87.80% for paddy and wheat crops, respectively. Removing the CBAM from the MaxViT blocks led to significantly lower accuracy (i.e., 83.65% for paddy and 83.80% for wheat). From this analysis, the importance of the CBAM becomes evident for improving classification performance. Other blocks (such as SE block and ConvNeXt block) were removed one at a time from the proposed model to assess their impact on classification performance. A visual representation of the ablation study is represented in Figure 10.

4.5. Statistical Analysis of Proposed Model Performance

To reinforce the robustness and reliability of the results, a 95% confidence interval (CI) was computed, offering a probabilistic approximation of the range within which the true test accuracy is likely to fall with 95% certainty. The CI was determined using the mean, the standard error, and the t-distribution, as defined in Equation (13).

$${\mathrm{CI}}_{95\%}=\overline{x}\pm t·{\displaystyle \frac{s}{\sqrt{n}}}$$

(13)

where ‘$\overline{x}$’ represents the sample mean of the test accuracies, ‘t’ is the critical t-value for the chosen confidence level, ‘s’ denotes the standard deviation, and ‘n’ denotes the number of independent trials. The proposed model was assessed across five independent runs using the paddy disease dataset to ensure variability in testing conditions.

The resulting performance (presented in

Table 14. Statistical analysis of proposed model performance and confidence interval calculation.

Sl. No.	Run Test	Test Accuracy (%)	Mean Test Accuracy (%)	Standard Deviation	95% Confidence Interval
1	Run test_1	97.19	98.09	0.5369	[97.43, 98.76]
2	Run test_2	98.34
3	Run test_3	98.23
4	Run test_4	98.60
5	Run test_5	98.11

) revealed a high degree of consistency, with a mean test accuracy of 98.09%, standard deviation of 0.5369, and 95% confidence interval of [97.43%, 98.76%]. This narrow confidence interval indicates that the model’s performance is accurate, stable, and reproducible. This statistical evaluation supports the claim that the model generalizes well across different test instances, affirming its suitability for real-world disease classification tasks.

4.6. Assessment of Model Robustness

To evaluate the robustness of the proposed model, Gaussian noise with zero mean and variance (0.001 and 0.0001), and Gaussian blur were added to the original crop images, and the noisy images (represented in Figure 11 and Figure 12) were fed to the network for classification.

For the original images (without blurring and noise), the testing accuracy was 98.47%. The addition of Gaussian noise (mean = 0, variance = 0.001) to the paddy crop images led to a testing accuracy of 97.13%. This value was increased to 97.79% by reducing the variance of Gaussian noise to 0.0001 from 0.001. For blurred images, the obtained testing accuracy was 97.77%. These experimental results, presented in

Table 15. Performance of the model for noisy images.

Sl. No.	Input	Accuracy (%) for Paddy Crop
		Training	Testing
1.	Images without noise	99.43	98.47
2.	Images with Gaussian noise (mean = 0, variance = 0.001)	98.06	97.13
3.	Images with Gaussian noise (mean = 0, variance = 0.0001)	98.61	97.79
4.	Images with blur (standard deviation = 1.5)	98.52	97.77

, indicate that the model retained a high level of accuracy even under degraded input conditions. The results demonstrate minimal reductions in both training and testing accuracies, thereby substantiating the robustness of the model to common visual distortions encountered in practical agricultural scenarios.

4.7. Comparison of the Proposed Model with Pre-Existing Studies

The proposed model has been validated using a custom paddy dataset and a wheat dataset (collected from Kaggle database). The results of the proposed method have been compared with pre-existing studies in the literature that utilized the same dataset. The comparison is presented in

Table 16. Comparison of disease classification by various methods.

S. No.	Author, Year	Techniques	Testing Accuracy (%)
			Paddy	Wheat
1.	Pai et al. [32] (2025)	Twin CNN Feature fusion	96.24	91.16
2.	Singh et al. [39] (2024)	DL	96.84	91.24
3.	Proposed	MaxViT (CBAM, SE, DW, and ConvNeXt)	98.47	92.80

.

4.8. Limitations of the Proposed Study

Despite the promising performance of the proposed model, a key constraint addressed in this study is the effect of varying train–test split ratios on model performance. To evaluate this, experiments were conducted using four commonly adopted train–test configurations: 90:10, 80:20, 75:25, and 70:30. Each train–test split involved random shuffling of the dataset to ensure unbiased sampling. The results indicate that the 90:10 train–test split yielded minimal deviation (0.9%) between training (91.21%) and testing accuracies (90.31%). The overall performance was suboptimal. The 80:20 train–test split outperformed others, achieving the highest training (99.43%) and testing (98.47%) accuracies with only a 0.96% difference/gap, indicating effective generalization. Conversely, the 75:25 and 70:30 splits indicated stronger over-fitting effects, as evidenced by larger performance differences/gaps (3.51% and 3.59%, respectively). Based on this comparative analysis, the 80:20 split was adopted for subsequent experiments due to its balance between training efficacy and testing reliability. Although the proposed model demonstrated high testing accuracy in rice disease classification (98.47%), its performance declined to 92.8% for wheat disease classification. To address this, future work can explore the integration of domain adaptation techniques, such as adversarial training or feature alignment, to enhance the model’s cross-crop generalizability. Additionally, training the model on multi-crop datasets and employing domain generalization strategies may help the model learn invariant features across plant types.

The DL models typically require vast amounts of data for strong generalization. However, publicly available datasets for various plant diseases are limited, which can negatively impact model performance. In this study, several disease samples were collected from local paddy fields under bright sunlight, introducing variations in lighting conditions. Moreover, some disease classes had significantly fewer samples, leading to class imbalance. This imbalance biases the model toward majority classes and reduces its ability to accurately recognize rare diseases. The other crop diseases are lower in number for a few classes, which remains a significant challenge for crop disease recognition. Addressing this issue through advanced data augmentation, synthetic sample generation, or class-rebalancing techniques will be a crucial direction for future research.

5. Conclusions

In this study, a lightweight DNN (having 814.7 K learnable parameters and 85 layers) is developed (by introducing the MaxViT module, which consists of a CBAM module, SE module, DW module, and ConvNeXt module) for the classification of crop (i.e., paddy and wheat) diseases. The attention mechanism is implemented in the CBAM block. The SE block focuses on important channels and captures channel-wise dependencies to improve feature extraction. The DW block reduces the number of parameters and the computational cost. The ConvNeXt module aids in robust feature extraction. This work performed experiments with a custom paddy disease dataset (collected from local paddy fields in Tiruchirapalli) and a publicly available wheat database. Performance indicators (such as accuracy, sensitivity, specificity, mean accuracy, precision, F1-score, and MCC) were computed to demonstrate the model’s robustness in classifying different diseases of paddy and wheat crops. The proposed model achieved overall accuracies of 99.43% and 98.47% for training and testing, respectively, in classifying seven disease classes in paddy crops. The method obtained overall accuracies of 94.00% and 92.80% during training and testing, respectively, for wheat crops. The proposed model’s performance was compared with pre-trained networks. The training and testing performance of the model was analyzed by spider plots to understand its generalizability. Overlapping of training and testing metrics indicates the model’s ability to classify diseases in unknown samples. The model required the least training time and had a higher overall accuracy due to having the lowest number of learnable parameters compared to other pre-trained networks and DL models. Due to its low number of parameters and short training time, the model can be used for real-time disease detection in crops using unmanned aerial vehicles. In our future research, the model can be generalized to identify diseases in vegetables and fruits, along with grading the severity of damage.