IMPACT OF EXPLAINABLE AI AND CNN MODELS ON BANANA LEAF DISEASE DETECTION

Abstract

Plant leaf disease detection plays a crucial role in precision agriculture, enabling early intervention and sustainable crop management. Traditional convolutional neural network (CNN) classifiers often depend on softmax outputs, which may produce overlapping feature spaces and overconfident predictions for visually similar classes. This study introduces an end-to-end CNN integrated with a Support Vector Machine (SVM) classifier that jointly optimizes feature extraction and margin-based classification using a categorical hinge-loss objective. The model was trained on the BananaLSD dataset comprising Cordana, Sigatoka, Pestalotiopsis, and Healthy banana leaves. The proposed architecture achieved an overall test accuracy of 73.2%, outperforming conventional CNN baselines in class-margin consistency. Furthermore, explainable AI (XAI) techniques based on Gradient-weighted Class Activation Mapping (Grad-CAM) were applied to visualize disease-affected regions, enhancing transparency and trust in model decisions. The combination of SVM-based learning and Grad-CAM interpretability demonstrates that integrating discriminative and explainable mechanisms can yield reliable and interpretable deep-learning solutions for agricultural disease diagnostics.