Local Structural Enhancement Strategy Reduces Errors in Complex Terrain Precipitation Fusion

A new local structural enhancement-based precipitation data fusion (LSE-PDF) framework improves precipitation estimation by integrating rain gauge observations with spatially continuous satellite products, according to a study evaluated in Shandong Province and mountainous Southwest China.

Representing local precipitation patterns remains a major challenge in regions characterized by complex terrain and sparse gauge networks. To address this issue, the proposed method introduces a local structural enhancement and virtual station generation (LSE-VS) module, which applies weights derived from a two-dimensional Pascal’s array to emphasize the central region of satellite precipitation patches while using an adaptive spatial attention mechanism to further adjust feature weights.

The architecture pairs these enhanced spatial features with a convolutional neural network coupled with long short-term memory (CNN-LSTM). By integrating the enhanced spatial features with station time series, the network generates virtual enhanced stations that augment the sparse ground observation network.

In the subsequent processing stage, the framework employs collaborative Yang-ChiZhong interpolation. This interpolation algorithm combines the virtual enhanced stations, the observed gauge stations, and the satellite precipitation data to construct continuous precipitation fields.

Researchers evaluated the framework across three experimental scenarios containing 1,308, 109, and 32 observations, respectively. The LSE-VS module achieved root mean square errors of 42.6696, 8.3630, and 9.7043 millimeters, alongside mean absolute errors of 19.0572, 5.3566, and 8.6750 millimeters, which were lower than the errors of the evaluated baseline models in each scenario. Visual inspections of the fused fields also demonstrated improved spatial continuity, with reduced over-smoothing and bullseye artifacts.

The authors noted that the framework offers potential utility for precipitation mapping in hydrological modeling and disaster monitoring. However, the learning-based estimates exhibited compressed prediction ranges at high precipitation intensities under certain evaluated conditions, limiting accuracy during extreme events. Future work will investigate intensity-aware training and calibration strategies using multi-year rainy-season and extreme-event samples to improve high-precipitation estimates.

The study was published in the Journal of Big Data, an open access publication, and is currently available as an early version of accepted, peer-reviewed research.

Subject of Research: Precipitation data fusion methodology

Article Title: Precipitation data fusion method incorporating local structural enhancement strategy

Article References: Yu, H., Li, R., Xiao, L., Feng, Q., & Lu, Y. (2026). Precipitation data fusion method incorporating local structural enhancement strategy. Journal of Big Data. https://doi.org/10.1186/s40537-026-01599-y

Image Credits: AI Generated

DOI: 10.1186/s40537-026-01599-y

Keywords: precipitation data fusion, satellite precipitation, rain gauges, deep learning, spatial interpolation, hydrological modeling, remote sensing, meteorology, Precipitation, data, fusion, method