Identification of tobacco leaf diseases using hyperspectral imaging and machine learning with SHAP interpretability...

Area of Science:

• Plant Pathology
• Remote Sensing
• Agricultural Science

Background:

• Tobacco diseases significantly impact crop yield and quality, necessitating advanced diagnostic methods.
• Current hyperspectral imaging (HSI) studies often focus on single diseases, lacking generalized frameworks for multi-class identification.

Purpose of the Study:

• To develop and validate an efficient, interpretable HSI framework for the simultaneous diagnosis of multiple tobacco diseases.
• To compare various preprocessing, wavelength selection, and machine learning techniques for optimal performance.

Main Methods:

• Collected hyperspectral images of healthy and diseased tobacco leaves (brown spot, wildfire, TMV, PVY).
• Constructed a balanced, leaf-independent dataset, strictly partitioning data at the leaf level.
• Systematically evaluated preprocessing (Savitzky-Golay), wavelength selection (SPA), and machine learning classifiers (ANN, Transformer), utilizing SHAP analysis for interpretation.

Main Results:

• An Artificial Neural Network (ANN) model with Savitzky-Golay preprocessing and SPA wavelength selection achieved high performance using minimal wavelengths.
• A Transformer model showed slightly better accuracy but required full-spectrum data and higher computational resources.
• Pixel-level predictions allowed for severity estimation of leaf-spot diseases, and SHAP analysis identified key spectral regions linked to physiological changes.

Conclusions:

• The developed HSI framework provides an efficient and interpretable method for multi-disease tobacco diagnosis.
• This approach supports the creation of practical hyperspectral or multispectral systems for real-time crop monitoring and management.
• The study highlights the potential of HSI in advancing precision agriculture for disease detection.