Interpretable MPA-Optimized ShuffleNetV2 for Plant Health Assessment

It is crucial to identify diseases affecting crops as soon as possible and as accurately as possible for sustainable agriculture. The proposed distinctive approach for detecting leaf diseases in black gram integrates ShuffleNetV2 for feature extraction, a hybrid method combining Principal Component Analysis (PCA) with the Marine Predators Algorithm (MPA) for optimal feature selection, and XAI (SHAP and Grad-CAM) for model interpretation. ShuffleNetV2 is a lightweight convolutional neural network (CNN) that extracts high-dimensional features refined further by MPA, mimicking the foraging behavior of marine predators, thus achieving a balance between exploitation and exploration. A support vector machine (SVM) classifies the selected features, which are then trained and evaluated using a 5-fold stratified cross-validation approach. The interpretation of SVMs’ decision-making processes is achieved by quantifying feature contributions by SHapley Additive exPlanations (SHAP) analysis, which renders the model more transparent. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) has been implemented in visualizing the areas of the input image that are the most important for the model’s predictions. Experimental findings validate that this novel pipeline attains competitive accuracy with low feature dimensionality, thus portraying the utility of MPA. The SHAP method allows further interpretation and qualifies the model as an excellent and fast disease-detection tool for real-world agricultural applications.