Banana Leaf Disease Identification Using SqueezeNet Architecture with Convolutional Block Attention Module

Abstract

Banana leaf diseases significantly reduce crop productivity and quality, while conventional visual inspection methods are often subjective, time-consuming, and inefficient for large-scale plantations. This study proposes an automated banana leaf disease identification approach using a lightweight Convolutional Neural Network (CNN) based on the SqueezeNet architecture integrated with the Convolutional Block Attention Module (CBAM). The dataset consists of four classes—Cordana, Healthy, Pestalotiopsis, and Sigatoka—with image augmentation applied to increase data variability. Several experimental scenarios were conducted to evaluate the impact of data augmentation and CBAM integration on model performance. The models were evaluated using accuracy, precision, recall, and F1-score metrics. Experimental results show that SqueezeNet combined with CBAM achieved superior performance compared to the baseline SqueezeNet model, particularly in non-augmented conditions, with an accuracy of 93.75% while maintaining a relatively small number of parameters. Although data augmentation alone led to performance degradation, the inclusion of CBAM mitigated this effect by enhancing spatial and channel-wise feature representation. These findings indicate that the proposed SqueezeNet–CBAM model offers an effective and computationally efficient solution for banana leaf disease identification, with strong potential for real-world agricultural applications.