Establishment of DIC concentration model and analysis of influencing factors of groundwater in a typical karst area based on machine learning

Dissolved inorganic carbon (DIC) is a key concentration term in hydrochemical-runoff estimates of karst carbon sinks, but dense and accurate DIC observations are difficult and costly to obtain. In this study, 692 groundwater and spring samples from the karst region of Guangxi, China, were combined with multi-source environmental predictors to develop a machine-learning framework for regional DIC mapping. Five machine-learning models were evaluated, and the three best-performing models (XGBoost, BRT, and RF) were further integrated using a stacking ensemble. The stacked model achieved a test-set R² of 0.83, with an RMSE of 0.45 mmol/L, MAE of 0.30 mmol/L, and MSE of 0.20. Predicted DIC showed a persistent spatial pattern. The northwestern regions revealed high values with a maximum of 3.83 mmol/L in 2015, while the southeastern region was low with a minimum of 3.48 mmol/L in 2004. The distribution of DIC concentration is governed by several factors, with elevated concentrations in the southwest region, forested areas, and higher altitudes. The proposed model provides a reliable basis for regional DIC mapping and can support subsequent karst carbon sink estimation when combined with hydrological flux data.