A Two-Stage Machine Learning Framework for High-Resolution Multi-Source Precipitation Fusion in Complex Terrain: A Case Study of Shaoxing, China

High-resolution precipitation fields are essential for flash-flood forecasting and hydrological risk management, especially in small and medium-sized basins, yet single-source precipitation products often show limited accuracy over complex terrain. This study develops a two-stage machine-learning framework for 1 km/1 h multi-source precipitation fusion over Shaoxing, China, during the 2025 flood season. In the first stage, a machine-learning classifier identifies precipitation occurrence and reduces zero-inflated noise; in the second stage, an optimized tree-based residual-regression model corrects precipitation estimates for rainy samples. A 61-dimensional feature set was constructed by integrating satellite precipitation estimates, weather-radar precipitation estimates from the Zhejiang radar network, temporal-lag and accumulation statistics, neighborhood descriptors, cyclic time variables, and terrain-derived interaction features, with gauge observations used as the training target. After quality control, the dataset comprised 41,458 hourly station samples from 72 rain gauges. The stations were divided at the station level into a 57-station development set and a fixed 15-station held-out spatial test set containing 8637 hourly samples. Station-blocked fivefold cross-validation within the development set was used for model selection, hyperparameter tuning, and probability-threshold selection, whereas the held-out stations were used only for final performance evaluation. On the fixed held-out test set, the occurrence classifier achieved an overall accuracy of 0.947, with a probability of detection of 0.806, a false alarm ratio of 0.158, a critical success index of 0.700, and an F1 score of 0.823. For quantitative estimation, the two-stage fusion product reduced root mean square error from 2.342 mm for satellite precipitation estimates to 1.189 mm, corresponding to a 49.22% reduction, and decreased mean absolute error from 0.712 mm to 0.262 mm, while increasing the coefficient of determination to 0.685. The fused precipitation product also improved the detection of intense rainfall events, with probability of detection and critical success index reaching 0.511 and 0.442, respectively, for events exceeding 10 mm/h, while reducing false weak precipitation and showing closer agreement with observed station-level spatial variability. By separating precipitation-occurrence identification from rainfall-intensity correction, the framework reduces zero-inflated bias, improves heavy-rainfall representation, and demonstrates predictive skill at gauges excluded from model development during the 2025 flood season.