Precipitation Data Fusion Method Incorporating Local Structural Enhancement Strategy

Precipitation data fusion is a critical approach to improving both the accuracy and spatial continuity of precipitation estimates in complex regions. However, existing fusion methods often struggle to balance high accuracy and spatial consistency in areas characterized by rugged terrain and sparse gauge networks, particularly due to their limited capacity to effectively capture and utilize local spatial structures. This limitation reduces the applicability of fused datasets in watershed hydrological modeling and disaster monitoring. To address this issue, we propose a local structural enhancement–based precipitation data fusion (LSE-PDF) to further improve the accuracy and spatial fidelity of precipitation fusion. First, a local structured enhancement and virtual station generation (LSE-VS) module is developed, which exploits the “central peak with outward smooth decay” advantage of the two-dimensional Pascal’s array in spatial weighting, integrates an adaptive spatial attention mechanism to reinforce key local features in remote sensing precipitation, and couples these enhanced features with station series data through convolutional neural networks with long short-term memory (CNN-LSTM) to generate virtual enhanced stations. Second, a collaborative Yang-ChiZhong interpolation algorithm is adopted to effectively merge the virtual and observed stations, yielding continuous high-accuracy precipitation fields. The proposed method was validated in two contrasting regions of China: Shandong Province, representing a station-dense plain area, and Southwest China, representing a station-sparse mountainous area with complex terrain. The experiments considered three observational scenarios: dense (1308 observations), moderate (109 observations), and sparse (32 observations). The results show that LSE-PDF consistently improves precipitation estimates under all densities, with root mean square error (RMSE) reduced to 42.6696 mm, 8.3630 mm, and 9.7043 mm, and mean absolute error (MAE) reduced to 19.0572 mm, 5.3566 mm, and 8.6750 mm for dense, moderate, and sparse conditions, respectively. Spatial analysis shows that LSE-PDF preserves precipitation continuity and captures fine-scale details, effectively mitigating over-smoothing and the bullseye effect in both flat and complex terrains. These results highlight the method’s potential for accurate precipitation mapping, supporting applications in hydrological modeling and disaster early warning.