Precipitation data fusion method incorporating local structural enhancement strategy

Precipitation data fusion combines gauge observations with spatially continuous satellite products to improve precipitation estimation. However, representing local precipitation patterns remains challenging in regions with complex terrain and sparse gauge networks. This study proposes a local structural enhancement-based precipitation data fusion (LSE-PDF) framework. First, a local structural enhancement and virtual station generation (LSE-VS) module applies weights derived from a two-dimensional Pascal’s array to emphasize the central region of satellite precipitation patches. An adaptive spatial attention mechanism further adjusts the feature weights. A convolutional neural network coupled with long short-term memory (CNN-LSTM) then integrates the enhanced spatial features with station time series to generate virtual enhanced stations. Second, collaborative Yang-ChiZhong interpolation combines virtual and observed stations with satellite precipitation data to produce continuous precipitation fields. The framework was evaluated in Shandong Province and a mountainous region of Southwest China. The three experimental scenarios contained 1,308, 109, and 32 observations, respectively. LSE-VS achieved root mean square errors (RMSEs) of 42.6696, 8.3630, and 9.7043 mm and mean absolute errors (MAEs) of 19.0572, 5.3566, and 8.6750 mm. These errors were lower than those of the evaluated baselines in each scenario. Visual comparisons of the fused fields indicate improved spatial continuity and reduced over-smoothing and bullseye artifacts in the examples examined. The results suggest that LSE-PDF has potential for precipitation mapping to support hydrological modeling and disaster monitoring. However, the learning-based estimates exhibited compressed prediction ranges at high precipitation intensities under some evaluated conditions, limiting the representation of extreme events. Future work will incorporate multi-year rainy-season and extreme-event samples and explore intensity-aware training and calibration strategies to improve high-precipitation estimation.