Prediction of soil selenium content and model optimization for a watershed in the Taihang Mountains based on machine learning models
Abstract Accurate spatial prediction of soil selenium (Se) in mountainous regions is constrained by limited field sample sizes and unreasonable model validation strategies, which complicates co‑management of selenium‑rich agricultural production and heavy‑metal ecological risk mitigation. In this work, we investigated a metamorphic mountain watershed on the eastern Taihang Mountains, Hebei Province. A total of 107 topsoil samples were collected, and 17 soil geochemical indicators were measured. Two geostatistical models (ordinary kriging, OK; regression kriging, RK) alongside four machine-learning algorithms: Elastic-Net, radial-basis-function support vector machine (RBF-SVM), Honest Random Forest (Honest RF), and XGBoost were benchmarked. We adopted a dual nested cross-validation scheme to separate model-fitting performance from spatial extrapolation capacity. The results indicated that our spatial‑block cross‑validation avoids inflated accuracy caused by random sample division. Among all competing models, Honest RF achieved the optimal performance, exhibiting minimal residual spatial aggregation and stable predictive performance for unmeasured watershed areas. Permutation feature importance (PFI) and non‑linear response curves consistently identify soil organic matter (SOM), lead (Pb), and mercury (Hg) as dominant correlative predictors closely associated with spatial variation in soil Se, with distinct threshold-saturation responses for each driving factor. We put forward a tentative tri‑element coupling hypothesis explaining Se–Hg–Pb co‑enrichment driven by rock weathering, atmospheric deposition and organic‑colloid interfacial reactions. Our combined analytical pipeline—multi‑model comparison, spatial cross‑validation and interpretable driver analysis—offers a practical reference for point-wise out-of-sample prediction of soil elements in mountain catchments with limited sample sizes.
Authors
- Weiming Xie (ORCID: https://orcid.org/0000-0002-8714-4140)
- Deng Furong (ORCID: https://orcid.org/0000-0002-2801-4389)
- Zhao Xianglei
- Zhang Lufei
- Li Le
- Tian Shixiong
Institutions
- Hebei Normal University (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1038/s41598-026-75232-6
- Primary Topic
- Soil Geostatistics and Mapping
- Type
- article
- Field-Weighted Citation Impact
- 0.00