SPECGAN: Extracting sensitive bands from plant disease spectra based on generative adversarial network.

Area of Science:

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

Background:

• Hyperspectral imaging offers detailed spectral data for non-destructive plant disease diagnosis.
• High spectral dimensionality and limited, imbalanced data hinder accurate diagnosis and model applicability.
• Extracting subtle pathological signals is challenging due to noise and data scarcity.

Purpose of the Study:

• To develop SPECGAN, a Generative Adversarial Network (GAN) framework for sensitive band extraction and data augmentation in plant disease diagnosis.
• To overcome limitations of high dimensionality and imbalanced datasets in hyperspectral plant disease analysis.
• To improve the efficiency and interpretability of early disease diagnosis models.

Main Methods:

• Proposed SPECGAN framework incorporating Temporal-Domain Feature Pyramid Fusion (TD-FPNF) and a residual attention mechanism.
• Utilized a multi-scale convolutional module for capturing both narrow-band biochemical and broad-band structural features.
• Employed gradient saliency analysis for identifying discriminatory spectral bands and GANs for synthetic data generation.

Main Results:

• Identified sensitive spectral bands in the green peak (520-550 nm) and red-edge (680-720 nm) regions, correlating with disease-induced physiological changes.
• Achieved 96.22% accuracy using only the top 20 discriminatory bands (8% of spectrum) with an MLP classifier.
• Generated synthetic data significantly improved model performance (6%-13%) under a severe data imbalance (14.6:1).

Conclusions:

• SPECGAN effectively extracts key spectral bands and augments data for robust plant disease diagnosis.
• The method enhances the accuracy and applicability of hyperspectral imaging for early detection of rice bacterial leaf blight.
• SPECGAN offers an efficient, interpretable, and data-efficient approach for agricultural disease management.