Explainable and Robust Deep Learning for Banana Leaf Disease Classification: A Comparative Study of MobileNet, ResNet50, and EfficientNetB0 With Gradient-weighted Class Activation Mapping (Grad-CAM) and Statistical Validation

Banana cultivation is significantly affected by fungal and foliar diseases, leading to substantial yield losses worldwide. Early and accurate detection is therefore essential for sustainable crop management. This study presents a comparative evaluation of three convolutional neural network architectures - MobileNet, ResNet50, and EfficientNetB0 - for automated classification of banana leaf diseases using the BananaLSD dataset. The dataset includes four classes: Cordana leaf spot, Pestalotiopsis leaf spot, Black Sigatoka, and healthy leaves. To enhance robustness, Focal Loss and a Convolutional Block Attention Module are integrated into an improved MobileNet variant. Model interpretability is achieved using Gradient-weighted Class Activation Mapping (Grad-CAM), while robustness is evaluated through stratified 5-fold cross-validation. Statistical significance is assessed using McNemar’s test and Cochran’s Q test. Results show that ResNet50 achieves the highest test accuracy (99.24%) and macro-area under the curve (0.9997), while MobileNet achieves comparable performance (99.08% accuracy; macro-area under the curve = 0.9993) with significantly lower computational cost. Cross-validation confirms the stability of MobileNet (0.9915 ± 0.0066). Statistical analysis reveals significant overall differences among models (Q = 29.65, p < 0.001) but no significant difference between MobileNet and ResNet50 (p = 0.625). Grad-CAM visualizations confirm biologically meaningful lesion localization. These findings highlight MobileNet as an efficient and reliable solution for real-time banana disease detection.