Downscaling and Bias Correction of ERA5-Land Precipitation over the Tibetan Plateau Using Terrain-Aware XGBoost

High-resolution and accurate precipitation data are essential for regional climate and hydrological assessments across the TP. Although the ERA5-Land reanalysis dataset provides valuable spatial coverage, it frequently exhibits systematic wet biases and spatial smoothing over this topographically complex region. To address these limitations, this study employs a terrain-aware XGBoost downscaling and bias-correction framework. By integrating ERA5-Land monthly precipitation with in-situ observations from 71 meteorological stations spanning 1981 to 2017, the model incorporates geographical coordinates and critical topographical features, specifically elevation, slope, northness, and eastness. Utilizing Bayesian hyperparameter optimization through the Optuna framework, a corrected monthly precipitation dataset with a 0.01° spatial resolution spanning 1981 to 2025 was generated. Independent testing from 2012 to 2017 demonstrated substantial accuracy improvements over the original ERA5-Land product. Specifically, the RMSE decreased by 68.3% from 72.59 mm to 23.02 mm, the MAE decreased by 70.4% from 44.05 mm to 13.04 mm, and the R2improved from -0.48 to 0.85. This terrain-aware framework effectively mitigates systematic biases and spatial mismatches, providing a reliable, finely resolved precipitation product for ungauged alpine regions to support downstream hydrological modeling, water resource analysis, and hazard assessment.