An Explainable Deep Learning Framework for Plant Leaf Disease Detection Using a Custom CNN Model

Abstract

Agricultural sustainability relies heavily on the early detection of plant pathologies. However, manual diagnosis remains challenging even for experts. This study proposes a lightweight Custom Convolutional Neural Network (CNN) architecture for automated leaf disease detection. The model was evaluated against state-of-the-art frameworks, MobileNetV2 and EfficientNetB0, using a dataset of 15,649 images that integrates global data with locally sourced samples from Libya. To ensure robustness, k-Fold Cross-Validation was implemented under standardized conditions. The proposed Custom CNN achieved a competitive accuracy of 97.6%, closely matching EfficientNetB0 (98.4%). Despite the slight accuracy advantage of transfer learning models, the Custom CNN demonstrated superior computational efficiency and a significantly smaller architectural footprint. These results position the proposed model as an ideal candidate for deployment in resource-constrained environments and mobile-based diagnostic systems.