Enhancing Bias Correction and Downscaling of Rainfall Pattern Over Taiwan with a Deep Learning Neural Network Over Complex Terrain

This study evaluates the application of a deep learning approach employing a multi-head attention mechanism within a deep neural network (DNN) framework to enhance bias correction and downscaling of the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis rainfall datasets. The proposed Encoder-Decoder with multi-head Attention (EDA) model leverages gridded 5-km daily rainfall observations and auxiliary inputs, such as surface wind data and high-resolution topography, to generate local-scale daily rainfall estimates across Taiwan—a mountainous subtropical island with complex terrain.

The model's performance is assessed using mean rainfall patterns, rainfall statistics, extreme climate indices, and interannual variations during Taiwan's rainy seasons. Results demonstrate that the EDA model effectively corrects biases in low-intensity rainfall and resolves inaccuracies in orographic rainfall placement present in reanalysis datasets, outperforming conventional quantile-mapping methods. Additionally, the integration of auxiliary surface wind information significantly improves the model's downscaling accuracy across various metrics.

This study highlights the potential of deep learning architectures, particularly those incorporating attention mechanisms and auxiliary data, for statistical bias correction and downscaling in regions characterized by intricate interactions between synoptic and local circulations modulated by topography.