Answers are grounded in this paper and cite their sources.
Hyperspectral imaging offers a powerful approach for agricultural quality sensing, yet its practical deployment remains constrained by high acquisition cost, spectral redundancy, and limited model interpretability. To address these issues, this study developed a slice-based framework for cassava amylose grading that integrates hyperspectral reflectance analysis, interpretable machine learning, and RGB-to-selected-band spectral reconstruction. Hyperspectral images of cassava storage roots were acquired, and samples were grouped into high, medium, and low amylose classes according to reference measurements. Based on full-spectrum reflectance, multiple preprocessing strategies and representative classifiers were systematically compared, and the XGB achieved the highest classification accuracy of 83.99% using either RAW or Savitzky–Golay-preprocessed spectra. Shapley Additive Explanations (SHAP) were further used to quantify spectral contributions and identify task-relevant wavelengths associated with amylose discrimination. To reduce sensing cost while preserving downstream utility, an RGB-to-selected-band reconstruction network was developed to estimate the SHAP-selected task-relevant spectral bands from RGB images for subsequent grading. The reconstructed selected-band images showed good agreement with the measured data, achieving an MRAE of 0.0697 ± 0.0013, an RMSE of 0.0471 ± 0.0008, and a PSNR of 26.57 ± 0.12 dB. More importantly, spectral reflectance extracted from the reconstructed images at the selected feature bands retained practical grading capability, with the XGB classifier yielding an accuracy of 71.75%. These results demonstrate that task-oriented selected-band reconstruction can retain useful trait-relevant spectral information for downstream classification while reducing reliance on dedicated hyperspectral acquisition. The proposed framework provides an interpretable and cost-aware technical solution for agricultural quality sensing and high-throughput phenotyping of starch-related traits.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: