A dual-TransUNet deep learning framework for multi-source precipitation merging and improving seasonal and extreme estimates

China (18.0–53.5°N, 73.5–135.0°E), focusing on eastern China, where heavy-rain hazards are most severe, and the Qinghai–Tibet Plateau, where sparse gauges limit precipitation assessment. We develop DDL-MSPMF, a dual-stage deep-learning framework that merges and bias-corrects daily 0.25° precipitation over China during 2001–2020 using six multisource precipitation products and four ERA5 near-surface predictors. A classifier first identifies rainfall occurrence, and a regressor then predicts rainfall amount. Performance is evaluated against the original products and deep-learning/hybrid baselines using CN05.1 across China and TPHiPr over the Qinghai–Tibet Plateau, with additional tests for heavy-rain detection and the July 2021 Zhengzhou rainstorm. Separating rainfall occurrence from rainfall amount improves precipitation merging across contrasting hydroclimatic regions of China. The best model, TransUNet–TransUNet, achieves the highest seasonal skill (R = 0.7512; RMSE = 2.6954 mm/day), improves heavy-rain detection across much of eastern China, and better reproduces the Zhengzhou event than the original products and simpler baselines. These gains also extend to the data-scarce Qinghai–Tibet Plateau. Interpretation shows that rainfall-occurrence information suppresses false light-rain signals and reduces heavy-rain underestimation, while surface pressure is the most influential auxiliary predictor. The framework provides an interpretable route to improving precipitation forcing for flood monitoring and hydroclimatic risk assessment in China. • A two-stage model fused six MSPs with ERA5 predictors for daily correction. • TransUNet-TransUNet gave the best seasonal skill among tested hybrids. • The framework improved heavy-rain detection across much of eastern China. • Skill improvements extended to the data-scarce Qinghai-Tibet Plateau.