A hyperspectral co-design framework guided by occlusion sensitivity for early mould detection in bamboo

Area of Science:

• Materials Science
• Computer Science
• Agricultural Science

Background:

• Schizostachyum funghomii McClure bamboo is crucial for crafts but vulnerable to mold.
• Mold contamination degrades bamboo quality and poses health risks due to mycotoxins.
• Hyperspectral imaging (HSI) and deep learning (DL) show potential for mold detection, but feature selection and model design integration is challenging.

Purpose of the Study:

• To introduce a co-design engineering framework for integrating model optimization and feature selection in HSI-based mold detection.
• To leverage Occlusion Sensitivity (OS) analysis for diagnosing and refining deep learning models.
• To develop an efficient and interpretable hyperspectral system for bamboo quality assessment.

Main Methods:

• Implemented a two-stage framework using Occlusion Sensitivity (OS) analysis.
• Diagnosed baseline networks' over-reliance on spatial textures and refined a ResNet model (ResNet-HS) focusing on spectral features.
• Developed an OS-guided algorithm for selecting characteristic wavelengths and a lightweight model using these bands.

Main Results:

• Refined ResNet-HS model improved validation accuracy from 87.04% to 95.06% on full-spectrum data.
• An OS-guided algorithm identified seven characteristic wavelengths.
• The final lightweight model achieved 90.74% accuracy with a 97% data reduction, outperforming other methods.

Conclusions:

• The proposed co-design framework effectively integrates model optimization and feature selection using OS analysis.
• This approach enables the development of accurate, efficient, and interpretable hyperspectral systems for material quality assessment.
• The study demonstrates a novel method for enhancing mold detection in bamboo and potentially other agricultural products.