Data-driven environmental monitoring of soil potentially toxic elements using multisource remote sensing and Machine Learning

Soil contamination monitoring in industrialized regions requires accurate, spatially continuous assessments. We present an integrated remote sensing framework for predicting concentration of soil Potentially Toxic Elements (Cd, Be, V, Cr, As, Co), which were selected based on their significant correlations with hyperspectral and multispectral signatures observed in preliminary exploratory chemical analysis. The framework integrates PRISMA hyperspectral, Sentinel-2 multispectral, and DEM-derived topographic data, tested near the industrial area of Taranto (southern Italy), a priority site for environmental risk assessment.

Our methodology integrates heterogeneous satellite data through systematic preprocessing, spectral index computation, morphometric feature extraction, and spatially feature selection. A correlation-based selection algorithm with a spectral distance constraint (∆λ<30 nm) was specifically implemented to mitigate multicollinearity inherent in high-dimensional hyperspectral data, ensuring the selection of non-redundant predictors. Machine learning regression models were trained on laboratory measured soil samples and validated via stratified cross-validation and independent holdout data.

Results demonstrate differential model performance across PTEs: R² = 0.75–0.82 (training) and 0.58–0.68 (validation). Feature importance analysis revealed complementary contributions from hyperspectral bands, multispectral indices, and terrain morphology, with hyperspectral data providing the strongest discriminative power. Single-sensor approaches (Sentinel-2 only) yielded notably lower performance, confirming the value of data integration. High-resolution maps identified the most polluted areas, validating the framework's capability for spatial assessment of soil contamination hotspots.