The proposed AgriX-SENet framework effectively combines high classification performance with model interpretability, addressing a key limitation of existing CNN-based plant disease detection systems and making it a promising solution for scalable agricultural diagnostics.
Abstract
Introduction Timely and accurate detection of plant diseases is essential for ensuring global food security and supporting sustainable agriculture. Conventional diagnostic approaches, such as manual inspection and laboratory testing, are often time-consuming, labor-intensive, and impractical for large-scale or remote agricultural environments. Although deep learning models, particularly Convolutional Neural Networks (CNNs), have significantly improved automated plant disease classification, they often lack interpretability and struggle to generalize under diverse field conditions. Methods This study proposes AgriX-SENet, an explainable deep learning framework that integrates Squeeze-and-Excitation (SE) blocks with a DenseNet121 backbone to enhance disease classification performance. The SE blocks recalibrate channel-wise feature responses to emphasize disease-relevant information while suppressing background noise. To improve model transparency, Grad-CAM, SHAP, and LIME were incorporated to provide visual and feature-level explanations of the model’s predictions. The framework was trained and evaluated using the Plant Pathology 2020 dataset containing four classes: healthy, rust, scab, and multiple diseases. Results AgriX-SENet achieved a training accuracy of 97.47% and a validation accuracy of 95.07%, outperforming fourteen state-of-the-art deep learning models. The classification report demonstrated high precision and recall across most disease categories, although the scab class exhibited comparatively lower recall, indicating an opportunity for further improvement. The explainability analyses consistently showed that the model focused on pathologically relevant regions of leaf images, validating the reliability of its predictions. Discussion The proposed AgriX-SENet framework effectively combines high classification performance with model interpretability, addressing a key limitation of existing CNN-based plant disease detection systems. Its ability to provide accurate and explainable predictions makes it a promising solution for scalable agricultural diagnostics. Future work will focus on improving classification performance for challenging disease categories and optimizing the framework for deployment on mobile and edge computing devices to enable real-time field applications.
Rice diseases and pest-related damage severely threaten agricultural productivity and global food security. While deep learning has advanced automated crop diagnostics, distinguishing visually similar disease symptoms and damage patterns remains challenging due to subtle visual variations and complex backgrounds. To address this challenge, we introduce DASO-RiceNet (Dual-Attention Semantic Optimization Network), a deep learning framework for fine-grained rice disease and damage classification. The architecture utilizes a multi-stage residual backbone for feature extraction and a sequential dual-attention module combining channel and spatial attention to emphasize diagnostically relevant features while suppressing background information. Evaluated on a ten-class rice disease and damage dataset under a consistent experimental protocol, DASO-RiceNet achieved an accuracy of 0.965, macro precision of 0.963, macro recall of 0.966, macro F1-score of 0.964, and micro F1-score of 0.965, outperforming the evaluated CNN- and transformer-based baseline models. Furthermore, Grad-CAM and LIME provided qualitative insights into model predictions, with the examined examples showing attention to visually apparent symptom-related regions. These results demonstrate the potential of DASO-RiceNet for automated rice disease and damage classification and precision-agriculture applications.
Plant diseases are still posing a challenge to the productivity, quality of crops, and food security, especially in locations where field diagnosis is based on manual visual inspec-tion. This paper assesses deep learning network-based automated classification of plant leaf diseases on public RGB leaf-image datasets, such as the Kaggle New Plant Diseases Dataset (Augmented) and PlantVillage images. They investigated four archi-tectures: EfficientNetV2B0, ResNet152V2, DenseNet201, and one hybrid Vision Trans-former (ViT)-based model. The steps of the experiment involved loading the dataset, exploratory analysis, preprocessing, resizing, normalizing, augmentation, transfer learning, independent model training, and evaluation metrics such as accuracy, preci-sion, recall, F1-score, training curves, testing results, and confusion matrices. The hy-brid ViT-based model was reported to have the best accuracy of 99.5%. On the smaller seven class subset of PlantVillage, EfficientNetV2B0 scored 98.11%. On the 38-class dataset, DenseNet201 improved test accuracy (97.34) and validation classification ac-curacy (around 98). ResNet152V2 scored 97.01 on the 38-class test set. The results demonstrate that CNN and transformer-based models can help to recognize plant diseases accurately whereas hybrid attention-based structures provide a promising path to enhance fine-grained classification. Since the model notebooks had varying class settings and splits, the comparison is seen as a model-structured assessment as opposed to a precisely identical benchmark across all architectures.
Najaf Ali, Ferdi Sonmez· Veredas do Direito· 0 citations
This study proposes a hybrid technique that integrates attention-weighted exponential pooling (AWEP) with CNN and Vision Transformer (ViT) to enhance feature representation and significantly improve classification performance and highlights that ViT improves embedding separability through t‑distributed stochastic neighbor embedding (t-SNE), thereby reducing overfitting and producing fewer misclassifications in visually similar classes.
Maddassar Jalal, Amandeep Kaur· Applied Fruit Science· 0 citations
In the field of precision agriculture, one of the major hurdles is the early and accurate identification of plant diseases. Farmers may face serious irreversible loss in yield if there is a delay in diagnosis by even a few days. The CNN model has helped in improving the classification of plant diseases but faces difficulty in distinguishing fine-grained symptoms, has limited generalizability, and is not easily interpretable. Though the deep CNN model performs well in classifying plant diseases, the current methods have several flaws. Most importantly, existing algorithms misclassify visually complex samples due to their limited spatial differentiation of disease-specific morphological features, including lesion borders and necrotic regions. Classification reliability is further compromised by low inter-class embedding separability for visually comparable illness phenotypes. Apart from these representational problems, training instability is still a major problem for attention-based models. Combining randomly initialized attention modules with pretrained backbone networks causes this instability, which still limits practical application.
We suggest EfficientNet-CBAM-Prototype (ECP-Net), a unique end-to-end deep learning architecture, to overcome these constraints. Three complementary techniques are combined into a single framework by ECP-Net. First, parameter-efficient multi-scale feature extraction is done using an EfficientNetB0 backbone. Second, joint channel-wise and spatial feature recalibration is performed using a stabilized convolutional block attention module (CBAM). Third, a dynamic prototype memory layer uses cosine similarity-based categorization and exponential moving average (EMA) updates to maintain class-representative embedding vectors. To overcome the instability caused by randomly initialized attention weights, we further propose a two-phase training strategy wherein CBAM is frozen during phase 1 to allow prototype stabilization and then jointly fine-tuned with the learning rate in phase 2.
Evaluated on the PlantVillage tomato subset comprising 10 disease classes across a class-balanced split of 10,000 training, 500 validation, and 500 test samples, ECP-Net achieves 98.6% test accuracy, 98.59% F1-score, and 98.65% precision with only 4.80M parameters and 85.04 ms average inference time. These results outperformed baselines including VGG16 (97.00%), ResNet50 (81.20%), MobileNetV2 (81.20%), and CNN (70.00%).
Generalization is further validated on 35 real-field tomato leaf images captured under natural, uncontrolled conditions, confirming practical deployment potential.
E. Jansi, Kavitha Br· Frontiers in Plant Science· 0 citations
