Adaptive attention and severity estimation framework for robust pearl millet leaf disease identification

Abstract

Abstract Pearl millet is an important crop in arid regions, but its yield is reduced by foliar diseases like Downy Mildew and Rust. Traditional and deep learning methods struggle with accurate lesion detection, severity estimation, and robustness under complex field conditions, and often lack interpretability for practical agricultural deployment. To address these challenges, this study proposes the Adaptive Severity-Aware Swin Attention Network (ASA-SAN), an integrated framework designed for disease segmentation, classification, and severity estimation in pearl millet leaves. The proposed architecture combines a Swin Transformer encoder for hierarchical feature extraction with a ResUNet++ decoder for accurate lesion segmentation. This is further enhanced using Adaptive Channel Attention to improve feature discrimination and a dual-stream classification network to jointly capture local lesion characteristics and global contextual information. Additionally, an Adaptive Disease Severity Index (ADSI) is introduced to quantitatively assess disease progression based on lesion area ratio, color degradation, edge irregularity, and texture variations. Experimental evaluations conducted on a pearl millet leaf dataset demonstrate that the proposed method achieves a Dice score of 97.8%, IoU of 95.6%, classification accuracy of 98.3%, and F1-score of 98.2%, outperforming several state-of-the-art methods. Furthermore, Grad-CAM visualizations enhance model interpretability by highlighting disease-relevant regions. Overall, the ASA-SAN framework provides a robust, interpretable, and severity-aware solution for automated pearl millet disease analysis, enabling early detection and supporting precision agriculture practices for improved crop protection and yield optimization.