Large-scale continuous estimation of soil heavy metal content using 3DT-NET and DS-corrected hyperspectral imagery.

This study proposes a novel framework for soil heavy metal pollution monitoring and source apportionment, namely "multi-source data fusion-spectral correction-feature selection-inversion mapping-uncertainty assessment-source apportionment." In this study, GF-5A and Sentinel-2B images were fused using the 3DT-NET algorithm, and the fused imagery was subsequently corrected using the direct standardization (DS) spectral correction method. Fractional-order differentiation was then applied to enhance soil spectral features, while the Boruta algorithm was employed to select informative bands for heavy metal modeling. Machine learning models were developed to predict heavy metal concentrations, and the associated prediction uncertainty was quantitatively assessed. Finally, soil heavy metal pollution sources in specific sampling areas were identified through correlation analysis. Compared with GF-5A imagery alone, the fused and DS-corrected spectra increased the best prediction accuracy (R2) from 0.476 to 0.611 for Pb (+28.4%), from 0.448 to 0.551 for Zn (+23.0%), and from 0.654 to 0.764 for Cu (+16.8%). Relative to Sentinel-2B imagery, the improvements were even greater, with R2 increasing from 0.285 to 0.611 for Pb (+114.4%), from 0.215 to 0.551 for Zn (+156.3%), and from 0.458 to 0.764 for Cu (+66.8%). Source apportionment results further showed that Zn is jointly controlled by elevation and slope, whereas Pb and Cu are mainly affected by slope. Overall, the proposed framework provides an effective and transferable solution for overcoming the spectral-spatial trade-off in satellite remote sensing, and offers strong support for regional pollution screening, risk zoning, source tracing, and precision land management.