Skip to main content

Inferring Soil Trace Metal Contamination via Spectroscopy: A Proxy-Based Machine Learning Framework for Spatial Prediction in Complex Terrains

This study addresses a fundamental challenge in environmental geochemistry: how to reliably map trace metal contamination in heterogeneous terrains where conventional sampling and laboratory analyses are spatially limited and economically restrictive. We develop a generalizable, spectroscopy-driven learning framework that integrates VIS–NIR–SWIR reflectance data (350–2500 nm) with multiple regression models (ANN, SVR, PLSR) to quantify how preprocessing governs predictive performance for trace metals. Using 110 soil samples and rigorous validation (10-fold cross-validation and independent testing, n = 25), we show that predictive performance is element-dependent, with highest reliability for cadmium (R²ₚ ≈ 0.92; RPD > 3), while lower accuracy for arsenic, chromium, and copper reflects variability in mineral–metal coupling strength. The key innovation lies in formalizing proxy-based environmental inference as a transferable modeling paradigm, where trace metals spectrally inactive are estimated through their associations with mineralogical carriers (e.g., iron oxides, clay minerals). This reframes soil spectroscopy from direct detection to relationship-driven prediction, explicitly accounting for context dependency and confounding effects. By coupling machine learning outputs with geostatistical interpolation (RMSE ≈ 0.82–1.07), the framework produces spatially continuous contamination fields that resolve localized hotspots consistent with geological structure, without assuming causality. Beyond this case study, the framework provides a scalable global template for screening-level environmental monitoring, enabling rapid prioritization of sampling, early identification of contamination risk zones, and cost-efficient deployment in data-limited regions. This establishes a methodological bridge between high-resolution laboratory geochemistry and large-scale environmental decision-making, with direct implications for sustainable land management and contamination risk assessment worldwide.

Details

Author
Amir Joukar1, Saeid Pourmorad2, Samira Abbasi3
Institutionen
1Islamic Azad University, North Tehran Branch, Department of Earth Sciences, Iran; 2University of Coimbra, Centre of Studies in Geography and Spatial Planning (CEGOT), Department of Geography and Tourism, Portugal; 3Universidade de Coimbra, MARE - Centro de Ciências do Mar e do Ambiente / ARNET - Rede de Investigação Aquática, Earth Sciences Dep. - FCTUC, Coimbra, Portugal
Veranstaltung
GeoMinBochum 2026
Datum
2026
DOI
10.48380/wcap-kd13