Calibration transfer via standard reference materials: A chemometric strategy for multi-element screening of soil...

Visible-Near Infrared (VNIR) spectroscopy for soil heavy metal detection is often hindered by severe matrix interference and limited regional samples. To address these challenges, this study proposes an Adaptive Prior-guided Spectral Cross-Attention Network (AP-SCAN). By utilizing National Standard Reference Materials (SRMs) as a knowledge source, the framework employs a cross-domain attention mechanism and a Masked PolyLoss function to address spectral calibration across domains and class imbalance under data-scarce conditions. Evaluated on a dataset comprising 21 SRMs and 74 field samples, AP-SCAN demonstrated superior performance in the simultaneous screening-risk identification of six priority heavy metals (As, Cd, Cr, Cu, Ni, and Pb), achieving a Macro F1 of 0.952 and an Exact Match Ratio of 0.913. Interpretability analysis further showed that the discriminative spectral regions identified by the model were broadly consistent with known soil spectral characteristics associated with iron oxides, organic matter, and hydroxyl-bearing minerals. Although constrained by the limited physical sample size and specific pedogenic backgrounds, this study preliminarily validates that introducing SRMs as structured prior knowledge under limited regional sample conditions can effectively compensate for the representational deficiencies of small-sample models. This provides a promising strategy for Vis-NIR-based multi-element screening under local cross-scene conditions.