Decoupling geogenic and anthropogenic drivers of groundwater natural background levels using machine learning and high-resolution maps

Accurate determination of natural background levels (NBLs) of groundwater hydrochemistry is critical for setting water quality standards, yet traditional approaches often fail to capture spatial heterogeneity, particularly in large-scale regions with uneven monitoring networks and human activities. To bridge the gap, this study proposes a novel framework that integrates machine learning models (TabPFN and XGBoost) with spatial grid resampling to model NBLs and their controlling factors (strata, land use/land cover, topographic slope, population, and nitrogen fertilizer use), thereby generating high-resolution NBL maps. A total of 1250 groundwater samples from the northwest Sichuan Basin, China, were used to validate this framework. The results showed that TabPFN outperformed XGBoost in NBL prediction across all parameters, yielding higher R2 (0.933–0.979) and reduced uncertainty by 15%–44%. The high-resolution NBL maps revealed that, although lithology remained the primary control on geogenic indicators (e.g., Ca2+, Mg2+), anthropogenic drivers reshaped the baseline patterns of hydrochemical parameters (e.g., NO3-N, Cl−, and Na+). The dissolution of mirabilite, dolomitic strata (Cretaceous), and calcareous formations (Jurassic) explained the elevated SO42−/Na+ in the groundwater environmental unit (GEU) Ⅰ, Mg2+ in GEU Ⅱ, and Ca2+ in GEUs Ⅲ–Ⅵ, respectively. Slow groundwater runoff increased hydrochemical concentration in GEUs Ⅰ and Ⅲ–Ⅵ. Domestic sewage and nitrogen fertilizer application were identified as the primary sources of elevated Na+/Cl− and NO3-N levels, respectively, which notably affected GEUs Ⅰ and Ⅲ–Ⅴ. The maps characterized the spatial heterogeneity of NBLs, thereby enhancing the understanding of baseline hydrochemistry. Contemporary NBL should be regarded as a present-day baseline modified by modern human activities. These findings provide an effective approach for determining groundwater NBLs across large, unevenly sampled areas. © 2026 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.