Combining Causal Inference with Machine Learning for Reconstructing Mountain Snow Water Equivalent Data

Snow Water Equivalent (SWE) is a key variable for evaluating hydrological processes and the impacts of climate change in mountainous regions such as the Qilian Mountains. Passive microwave remote sensing provides large-scale SWE estimates, but its coarse spatial resolution and coverage gaps pose limitations, particularly in complex terrain with heterogeneous snow distribution. This study integrates multi-source data from 2018 to 2024, combining ground-based observations with multiple meteorological factors to develop a high-resolution SWE reconstruction model tailored to the Qilian Mountains. Eight machine learning algorithms—Support Vector Machine (SVM), CatBoost, LightGBM, XGBoost, Random Forest, AdaBoost, ElasticNet, and Bayesian Ridge Regression—were systematically compared, with LightGBM achieving the best performance on the test set. During feature selection, Granger causality inference was applied to screen input variables, resulting in an optimized reconstruction model with a mean absolute error (MAE) of only 1.984 mm, a root mean square error (RMSE) of 4.656 mm, and a coefficient of determination (R2) of 0.973. Model interpretability was enhanced using SHAP (Shapley Additive Explanations), which revealed that snow depth, surface soil temperature and moisture, and precipitation were the primary driving factors, with varying contributions to the model. The model generates SWE reconstruction sequences at 30 min intervals. This high-resolution dataset provides crucial support for studying snow dynamics in complex mountainous regions and contributes to improved water resource management and climate change assessments in the Qilian Mountains.