An Attention-Enhanced Lightweight Deep Learning Framework for Banana Leaf Disease Detection Using CBAM-MobileNetV2 with Grad-CAM Visualization

Abstract

 Fungal diseases like Sigatoka and nutrient deficiencies like potassium deficiency greatly affect farmer income and yield of bananas. The traditional visual inspection methods are subjective, labor-intensive and subject to errors. An improved deep learning model is proposed to identify banana leaves diseases efficiently by introducing a Convolutional Block Attention Module (CBAM) and MobileNetV2 architecture. The proposed method is based on a small labeled dataset consisting of 2,801 images across three classes (healthy, Sigatoka infected and potassium deficiency) and pre-processing techniques and incorporates visual explanations in the form of Gradient-weighted Class Activation Mapping (Grad-CAM). The training set is small but an independent test set of 1307 images from different agricultural farms has been used only for the final evaluation of the primary training set. The independent set was not used for training or validation providing an independent assessment of the model's ability to generalize to unseen real-world data. Overall, the developed system attained an accuracy of 94.4%, recall value of 97% for Sigatoka and 89% for potassium deficiency. The model's light weight stress and attention guided feature selection allows for accurate real-time diagnosis in low resource smallholder farms for timely interventions and better crop management.