Prior-guided spectral reconstruction and spectral-spatial learning for non-destructive origin authentication of Pinellia ternata.

Accurate origin authentication of Pinellia ternata requires non-destructive models that are reliable and interpretable. This study proposes a prior-guided spectral reconstruction and spectral-spatial ROI cube framework using a Vis-NIR spectral image dataset of 800 ROI samples from four origins. First, each ROI was represented by a mean spectrum and classified under raw, Savitzky-Golay, multiplicative scatter correction, and standard normal variate preprocessing. To enhance discriminative spectral regions, we introduce RSDQ, a reinforcement-style reconstruction algorithm that learns class-aware priors from training spectra; SHAP-derived band importance then guides where reconstruction is concentrated, and a validation-controlled multiplier limits over-reconstruction. Second, co-registered ROI patches were preserved as spectral-spatial cubes and classified by ResNet2D, Inception-ResNet2D, and an improved Inception-ResNet2D ensemble with test-time augmentation. SHAP-RSDQ increased the 50-run baseline macro-F1 from 0.8848 to 0.8963, and adaptive reconstruction reached 0.8975. The best reconstructed MSC-MLP route achieved 0.9738 macro-F1. The spectral-spatial ensemble achieved the highest five-seed macro-F1 of 0.9850. Grad-CAM, channel-gradient and band-occlusion analyses confirmed that performance gains were supported by interpretable spectral and spatial evidence.