Quantifying heavy metal concentrations in industrial-transitional zone soils via integrated XRF and VIS-NIR spectroscopy.

Some suburbs urgently need to investigate soil heavy metal contamination to ensure a clean environment. Compared to traditional monitoring methods, XRF and VIS-NIR spectroscopy offer advantages such as rapid, non-destructive, cost-effective, and environmentally friendly analysis. In this study, we developed multiple estimation models for Cd and As content, evaluated the impact of different spectral preprocessing methods on model accuracy, and analyzed the distribution characteristics and for the feature wavelengths selected by competitive adaptive reweighted sampling (CARS). We compared the accuracy of estimation models constructed using partial least squares regression (PLSR) and backpropagation neural networks (BPNN), and elaborated on the advantages of spectral concatenation (SC), outer product analysis (OPA), and Granger-Ramanathan averaging (GRA) fusion strategies. The results demonstrated that among single-spectrum estimation models, the highest accuracy was achieved using XRF and VIS-NIR transformed by second derivative (SD) preprocessing. XRF spectra exhibited a larger number of feature wavelengths with uniform distribution, while VIS-NIR feature wavelengths were concentrated in the 450-1000 nm range. PLSR models outperformed BPNN models in terms of accuracy. Among fused-spectrum estimation models, the accuracy ranking was OPA > SC > GRA, with the OPA model combined with Pearson correlation coefficient (PCC) dimensionality reduction achieving the highest accuracy (R = 0.9920, RPD = 11.2020 for Cd estimation; R = 0.9852, RPD = 8.2134 for As estimation). These findings establish a technical framework for estimating soil heavy metal content based on XRF and VIS-NIR spectroscopy, and offer a novel monitoring approach for agricultural soils in industrial-urban-rural transition zones.