An interpretable stacking ensemble model for simulating ice thickness in shallow lakes: Lake Ulansuhai of Central Asia

Abstract

Central Asia lies in the interior of the Eurasian continent, dominated by a temperate continental climate with long and cold winters. As a typical lake type in this region, shallow lakes are characterized by shallow water depth and low heat capacity, making them highly sensitive to atmospheric forcing. The distribution and thickness evolution of lake ice cover not only exert significant impacts on the structure and function of lake ecosystems but also serve as crucial sensitive indicators to regional climate change. Therefore, accurately and efficiently simulating lake ice thickness and investigating the driving mechanism of its changes have become increasingly important. Using in situ measurements collected by the Floating Remote Observation System during the complete ice season of 2022–2023 in Lake Ulansuhai, Central Asia, this study proposed a stacking ensemble model (XRF-Stacking) that fuses XGBoost and Random Forest, with RidgeCV as the meta-model, integrated through tenfold cross-validation. Four benchmark models, including XGBoost, Random Forest, SVR, and LightGBM, were selected for comparison, and model performance was evaluated using three metrics: R², RMSE, and MAE. Results show that the XRF-Stacking model substantially outperforms all benchmark models, with an R² of 0.995, RMSE of 0.010 m, and MAE of 0.007 m, indicating high consistency between predicted and measured values. Further model interpretation based on SHAP reveals that water temperature, net shortwave radiation, and net longwave radiation were the three most influential variables on ice-thickness variation. This study can provide methodological references for the modeling of ice thickness in shallow lakes.