Integrating a novel feature set with the random forest model to improve satellite hyperspectral estimation and...

Satellite hyperspectral remote sensing has proven to be reliable and cost-efficient for estimating the concentration levels and spatial distribution of potentially toxic elements (PTEs) in soil, thereby supporting contaminated soil management and environmental sustainability. However, satellite hyperspectral estimation of soil PTEs remains challenging because it is difficult to obtain an effective set of spectral features for model development, as PTEs occur at trace levels and their spectral signals are often mixed with those of major soil elements and environmental factors under field conditions. To address this problem, we conducted controlled spectral experiments using artificially prepared Pb-contaminated soil samples, in which interfering factors affecting the spectral features of soil Pb were removed. The spectral features of soil Pb were effectively identified using the SPA-SHAP algorithm, and a new spectral index for Pb-contaminated soil (Pb-SI) was developed. Based on these results, a novel spectral feature set (NSFS) for soil Pb was established. Subsequently, the NSFS was extrapolated to satellite hyperspectral datasets via the direct standardization algorithm. Building upon this, a satellite-based estimation model for soil PTEs was established by integrating the NSFS with the random forest (RF) algorithm. Results shown that (1) the NSFS consisted of eight spectral parameters: R₄₇₉, R₇₁₈, R₈₄₃, R₁₇₂₇, R₂₀₁₅, R₂₁₉₀, R₂₃₅₃ and Pb-SI; (2) the satellite-based soil Pb estimation model, constructed via the integration of NSFS and RF algorithm, attained R² = 0.76 and RPD = 2.07, showing a 25.45% improvement in RPD compared with models that use traditional spectral feature sets; and (3) high Pb concentrations are mainly observed in the northern part of the study area, which are closely related to mining sites and transportation routes. This study provides an effective satellite hyperspectral framework leveraging the synergy between the NSFS and RF model for the estimation and spatial characterization soil PTEs and is useful for soil pollution control and the protection of human health.