An Explainable and Attention-Mechanism-Based Deep Learning Framework for Reliable Plant Disease Detection

Abstract

The escalating global food crisis, exacerbated by climate change-induced yield losses and the increasing impact of pests and plant diseases, has reached a critical level. The early and accurate detection of plant diseases is of strategic importance not only for ensuring sustainable agricultural production but also for reducing economic dependency and safeguarding food security. Although deep learning-based approaches proposed in the existing literature often achieve high classification accuracy, their decision-making processes largely remain opaque, thereby limiting model reliability and practical adoption. In this study, a MobileNetV2-based deep learning architecture was employed for plant leaf disease classification, and the Convolutional Block Attention Module (CBAM) was integrated to enhance model performance. By emphasizing salient regions within leaf images, CBAM improved classification accuracy while simultaneously reducing computational overhead associated with processing irrelevant features. Furthermore, to enhance model interpretability, the Grad-CAM technique was applied to visualize the specific features and image regions that influenced the model's predictions. The experimental results not only demonstrate the contribution of the attention mechanism to classification performance but also address a significant gap in transparency and reliability within deep learning-based agricultural decision support systems.