Characteristic wavelength selection for rice blast based on hyperspectral remote sensing and deep convolutional...

Area of Science:

• Agricultural remote sensing
• Plant pathology
• Machine learning for agriculture

Background:

• Hyperspectral remote sensing is crucial for detecting rice blast, but current methods struggle with data redundancy and interpretability.
• Existing dimensionality reduction techniques often fail to extract the most informative spectral features for disease severity assessment.

Purpose of the Study:

• To develop an advanced feature wavelength selection method for hyperspectral data.
• To integrate deep learning with model attribution analysis for precise rice blast detection.
• To extract key spectral features across varying disease severity levels.

Main Methods:

• A novel Dilated Convolution and Deformable Convolution-Residual Network (DCR-ResNet) was developed to analyze spectral features.
• Integrated Gradient (IG) and Gradient-weighted Class Activation Mapping (Grad-CAM) were combined for spectral wavelength selection.
• Statistical and modeling analyses were used to validate the method's effectiveness.

Main Results:

• The DCR-ResNet combined with IG-GradCAM identified spectral features with high separability and compactness.
• Models using selected wavelengths outperformed those using conventional methods, achieving up to 86.2% accuracy.
• Classification performance was significantly improved compared to traditional dimensionality reduction techniques.

Conclusions:

• The DCR-ResNet and IG-GradCAM method enhances hyperspectral feature extraction accuracy for rice blast.
• This approach offers an efficient and feasible solution for precise rice blast identification using remote sensing data.