A Comparative Study of Deep Learning Architectures for Multi-Source Precipitation Data Fusion Towards Daily Precipitation Estimation in the Yellow River Basin, China

Accurate precipitation estimation is crucial for water management and flood reconstructions, yet remains challenging due to sparse ground observations and systematic satellite biases. Despite the rise of deep learning in multi-source fusion, controlled comparisons under a unified framework remain scarce, and the roles of window size and precipitation product combinations remain underexplored. This study evaluated five deep learning architectures, namely 1D CNN, 2D CNN, 1D–2D CNN, 3D CNN, and Transformer, for fusing three precipitation products (IMERG, MSWEP, and ERA5-Land) to estimate station-scale daily precipitation in the Yellow River Basin, China. Results showed that deep learning models consistently outperformed individual products, weighting methods, and machine learning methods, with Transformer achieving the best overall performance, followed by 3D CNN; MSWEP led the standalone products. Spatial improvements were more pronounced in densely gauged areas, but Transformer also performed well in the sparse western regions. All products and fusion models tended to overestimate no-rain and light-rain events and underestimate moderate-to-heavy, and extreme events, with Transformer yielding the largest reductions in conditional RMSE and BIAS for heavy and extreme events. Enlarging the spatial or temporal window individually improved performance, and combining both worked even better. Ablation experiments revealed that IMERG contributed little unique information, whereas incorporating more products generally improved Transformer performance. Overall, Transformer demonstrated superior performance in fusing precipitation products and enhancing estimation accuracy.