Spatial prediction of soil organic carbon stocks across contrasting Andean basins, Peru

Abstract

Soil organic carbon stocks (SOCS) are critical components of the global carbon cycling and play a central role in climate change mitigation. However, their dynamics in high‐altitude Andean ecosystems remain poorly understood despite their importance for carbon sequestration. The significant spatial heterogeneity of SOCS in mountainous terrain makes accurate quantification and mapping challenging. This study evaluated the performance of geospatial regression and machine learning (ML) approaches for predicting SOCS in two Peruvian Andean basins: Torobamba and Coata. We compared Geographically Weighted Regression (GWR), GWR with collinearity analysis (GWRC), their kriging‐adjusted variants, and ML models (Random Forest, Gradient Boosting). Models were built using key SOCS covariates for each basin and validated through 5‐fold cross‐validation with Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R²). In Torobamba, GWRC markedly improved performance, reducing the RMSE by 79–90% and achieving R² up to 0.99. In contrast, Coata, showed only modest improvements (RMSE reductions of 7.8–9.8%, R² = 0.30–0.39). ML models performed poorly (negative R²), likely due to feature selection, parameter tuning, or limited sample size. Overall, locally weighted regression approaches (GWRK/GWRCK) outperformed conventional ML methods for SOCS prediction in complex mountain environments, particularly with small to medium sample sizes. These results highlight the importance of accounting for spatial non‐stationarity in SOCS and provide methodological guidance for SOCS mapping in Andean ecosystems.