Fusion-Mamba: High-resolution spatiotemporal fusion of multi-source meteorological precipitation data with delayed-mode station calibration over complex terrain

Reliable 3-hourly precipitation information is difficult to obtain in complex terrain because available products differ in spatial detail, temporal continuity, and bias. We develop Fusion-Mamba, a delayed-mode, station-calibrated multi-source fusion framework for regional hydroclimate assessment in Qinghai Province. The model is trained with gauge observations and, at inference, uses grid-available precipitation products, reanalysis fields, terrain attributes, and time factors to generate estimates on a 0.01° output grid. It combines point- and neighbourhood-scale representations, temporal modelling, and station-wise affine calibration. On the chronological test evaluation at 52 stations, Fusion-Mamba achieved CC = 0.8313, RMSE = 0.493 mm/3 h, MAE = 0.058 mm/3 h, and RB = 0.0275. It delivered the strongest pooled performance among five baseline models, although its comparison with the closest baseline was metric dependent. A separate station-disjoint experiment indicated metric-dependent transfer of the common neural backbone. In an additional production-consistent evaluation, five held-out stations were treated as ungauged target locations. Their predictions used the four nearest training stations and normalized inverse-distance-squared blending, yielding CC = 0.8254, RMSE = 0.3677 mm/3 h, MAE = 0.0538 mm/3 h, and RB = − 0.0215 across 210,040 samples. Heavy-rain results ( N = 234 ) require caution. The framework is not intended for near-real-time delivery, and latency depends on the slowest input plus preprocessing and quality control. The production-consistent test directly validates the blending rule at independent station locations, but broader station and operational validation remains necessary.