A two-stage near-infrared framework for screening and quantification of trace soil hexachlorocyclohexane under strong...

Hexachlorocyclohexane (HCH) is a persistent legacy pesticide whose presence in soils threatens environmental quality and food-chain safety. Although near-infrared (NIR) spectroscopy is attractive for high-throughput, reagent-free soil analysis, its application to trace HCH remains challenged by weak analyte-related signals and their strong overlap with variable soil-matrix absorption and scattering. These conditions destabilize direct concentration modeling and, in practical datasets, are further aggravated by a substantial fraction of samples reported as 2 = 0.8643, RMSE = 1.94, and RPD = 2.73. Beyond predictive performance, the study clarifies what the models are learning. SHAP analysis showed that the classification was driven not by an isolated HCH fingerprint, but by matrix-conditioned multiband spectral patterns mainly in the 1300-1450 nm and 1600-1750 nm regions. Stacking contribution analysis further showed that robust prediction arose from the selective fusion of complementary learners with distinct inductive biases. Together, these findings advance soil spectroscopy for trace contaminants toward a more interpretable framework for regional soil HCH screening, prediction, and risk-oriented assessment.