An Explainable Deep Learning Network With Transformer and Custom CNN for Bean Leaf Disease Classification

Abstract

Bean rust and angular leaf spot pose significant challenges to bean cultivation, impacting yields. Prompt disease identification maximizes productivity, but traditional methods need specialized expertise. This research presents an explainable deep learning model that combines the Pyramid Vision Transformer (PVT) and Group Context Aware Depthwise Shuffle Network (GCADSN). The PVT effectively models long-range dependencies, identifying disease patterns across larger leaf areas, while the GCADSN focuses on capturing nuanced, context-specific features. This combined approach leads to a richer representation of the input image, resulting in improved disease classification. Model explainability is provided through GradCAM visualizations, highlighting the image regions crucial for the model’s predictions and enabling transparent, class-specific insights. The model’s performance was rigorously tested using the IBean dataset, a collection of images depicting various bean leaf diseases, including common rust, angular leaf spot, and healthy leaves. Our proposed network achieved a high accuracy of 97.66%, outperforming many current state-of-the-art deep learning models. Furthermore, it demonstrated strong performance across other key metrics, with an F1 score of 97.67%, precision of 97.83%, and recall of 97.67%. Importantly, the model’s computational efficiency makes it well-suited for practical application in real-world agricultural scenarios.