High-resolution climate prediction in mountainous terrain using a ConvLSTM-XGBoost hybrid model with dynamic bayesian weighting

To address the challenge where the interplay between spatiotemporal dynamics and topographic effects complicates climate modeling over complex terrain, we propose a hybrid ConvLSTM-XGBoost model incorporating dynamic Bayesian weighting, and demonstrate its capacity for high-precision climate prediction through a case study in the Hongyuan Mountain region of Yunnan, China (22.5°–23.5°N, 102.5°–103.5°E); specifically, the ConvLSTM network captures spatiotemporal evolution patterns (e.g., propagation of the southwest monsoon front) from the 0.25° resolution CN05.1 climate dataset, while XGBoost quantifies the nonlinear modulation effects of 90-m SRTM DEM-derived topographic features (elevation, aspect) on precipitation phases, with an innovatively integrated Bayesian Model Averaging (BMA) framework dynamically calibrating model weights—XGBoost at 0.68 ± 0.05 during dry seasons and ConvLSTM at 0.72 ± 0.07 during monsoons—to enhance responsiveness to extreme events. Validation using 1961–2022 climate data shows the hybrid model reduces precipitation prediction mean absolute error (MAE) by 30.5% compared to CMIP6 (achieving an MAE of 0.0089 [specify units, e.g., mm/day]), improves the F1-score for identifying extreme precipitation (> 50 mm/day) by 20%, achieves 96.53% accuracy in maximum temperature (Tmax) predictions (errors ≤ 3%), and reduces high-temperature dispersion by 52%, thereby serving as a 1-km resolution decision-support tool for mountain climate risk management, supporting drought warning and hydropower scheduling in Yunnan’s Climate Adaptation Plan 2035, and offering a scalable framework for global mountain climate modeling.