Resolving Geogenic and Anthropogenic Sources of Soil Contamination in Central Tanzania Using Probabilistic and Machine Learning Approaches

Abstract

<p>Soils in mining terrains are subject to complex interactions between geological backgrounds and human activities, often resulting in elevated concentrations of Potentiality Toxic Elements (PTEs). This study applied an integrated framework combining probabilistic pollution indices, positive matrix factorization (PMF), and machine learning (Gradient Boosted Decision Trees and Artificial Neural Networks) to evaluate soil contamination in the Singida mining terrain of Tanzania. A total of 1,884 surface soil samples (0–20 cm) were analyzed for 12 PTEs. Concentrations showed strong heterogeneity, with right-skewed distributions indicating hotspot enrichment. Pb (mean 25.3 mg/kg; 70% > UCC) reflects regional background enrichment with possible localized anthropogenic enhancement, whereas Cd (0.13 mg/kg; 49% > UCC), and As (1.85 mg/kg; 5% > UCC) show stronger anthropogenic influence. Cr (62.6 mg/kg; 18% > UCC), Ni (23.4 mg/kg; 14% > UCC), and V (61.2 mg/kg; 16% > UCC) reflected lithogenic control from mafic–ultramafic lithologies. Probabilistic simulations (20,000 iterations) showed that most soils were low risk with Pollution Load Index (PLI) mean 0.60; Potential Ecological Risk Index (PERI) mean 59.5; and Nemerow Integrated Risk Index (NIRI) mean 29.5, yet ~21% of sites reached moderate to extreme risk categories. PMF resolved two dominant source factors: (i) a lithogenic Ba–Sr–Pb–Cd–Mn assemblage, and (ii) a ferromagnesian–sulphide Cu–Ni–Cr–V–Co–Zn–As assemblage. Machine learning reproduced these factor contributions with high fidelity (R2 = 0.96–0.99), enabling nonlinear sensitivity analysis and identification of dominant predictor elements rather than independent validation of the PMF solution. These findings demonstrate the effectiveness of a combinatorial approach in capturing both deterministic structure and stochastic uncertainty in soil contamination. The results highlight the need for hotspot-targeted remediation, region-specific baselines, and integration of probabilistic monitoring frameworks into environmental policy for mineralized terrains in Sub-Saharan Africa.</p>