High-resolution precipitation downscaling in mainland Southeast Asia: A novel integration of BMA and U-Net CNN

Abstract

This study introduces a hybrid approach that combines Bayesian Model Averaging (BMA) with a U-Net Convolutional Neural Network (CNN) to improve precipitation estimates in mainland Southeast Asia. The method addresses key limitations of Global Climate Models in capturing fine-scale variability, particularly in topographically complex. An ensemble of five GCMs, supplemented by ERA5 reanalysis data, was used to produce high-resolution downscaled precipitation estimates. Compared to the original BMA, the hybrid model significantly improved performance, increasing the Symmetric Concordance Correlation Coefficient from 0.68 to 0.82 and reducing the Root Mean Squared Error from 1.63 to 1.27 (validated against ERA5). Validation using TRMM and IMERG data revealed similar enhancements. Additionally, Wasserstein distance analysis confirmed improved distributional similarity between model outputs and observed data. The most notable improvements occurred in mountainous areas, especially in northern Myanmar. This approach enhances the utility of climate data for water resource management and adaptation planning in Southeast Asia.