DC²Net: An Asian Soybean Rust Detection Model Based on Hyperspectral Imaging and Deep Learning.

Area of Science:

• Agricultural Science
• Plant Pathology
• Computer Vision

Background:

• Asian soybean rust (ASR) causes significant global yield losses, necessitating early and accurate detection methods.
• Hyperspectral imaging and deep learning show promise for crop disease detection, but current models struggle to extract both spatial and spectral features effectively.
• Existing deep learning architectures have limitations in capturing complex spatial and spectral information from hyperserspectral images, hindering detection accuracy.

Purpose of the Study:

• To develop an advanced deep learning model for early and accurate detection of Asian soybean rust (ASR).
• To improve the extraction of spatial and spectral features from hyperserspectral images for enhanced ASR identification.
• To leverage attention mechanisms and feature importance analysis for robust and interpretable ASR detection.

Main Methods:

• Proposed a novel deformable convolution and dilated convolution neural network (DC²Net) for ASR detection.
• Utilized deformable convolutions for spatial feature extraction and dilated convolutions for spectral feature extraction.
• Incorporated Shapley value and channel attention methods to identify critical wavelengths for ASR detection.

Main Results:

• The DC²Net achieved a high overall accuracy of 96.73% in detecting ASR, outperforming existing state-of-the-art methods.
• Demonstrated the capability of DC²Net for early, asymptomatic detection of ASR, even before visual symptoms are present.
• Shapley Additive exPlanations (SHAP) analysis confirmed the model's ability to identify key wavelengths, enabling performance with reduced data.

Conclusions:

• The proposed DC²Net significantly enhances the accuracy and timeliness of ASR detection using hyperspectral imaging.
• Early and asymptomatic detection of ASR is achievable, providing critical advance warning for disease management.
• The integration of attention mechanisms and feature interpretability offers a path towards more efficient and reliable crop disease surveillance systems.