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

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Pearl millet is vital for arid regions, but foliar diseases like Downy Mildew and Rust significantly reduce yield.
• Existing methods for disease detection and severity estimation in pearl millet lack accuracy and interpretability in real-world field conditions.

Purpose of the Study:

• To develop an integrated, interpretable, and severity-aware framework for automated disease analysis in pearl millet leaves.
• To enhance disease segmentation, classification, and severity estimation using advanced deep learning techniques.

Main Methods:

• Proposed the Adaptive Severity-Aware Swin Attention Network (ASA-SAN), combining a Swin Transformer encoder and ResUNet++ decoder.
• Incorporated Adaptive Channel Attention for feature discrimination and a dual-stream network for comprehensive feature capture.
• Introduced an Adaptive Disease Severity Index (ADSI) for quantitative disease assessment based on multiple visual parameters.

Main Results:

• Achieved high performance with a Dice score of 97.8%, IoU of 95.6%, classification accuracy of 98.3%, and F1-score of 98.2%.
• Demonstrated superior performance compared to state-of-the-art methods in disease segmentation and classification.
• Grad-CAM visualizations confirmed model interpretability by highlighting disease-affected areas.

Conclusions:

• The ASA-SAN framework offers a robust and interpretable solution for automated pearl millet disease analysis.
• This technology supports early disease detection and precision agriculture, leading to improved crop protection and yield optimization.