An unsupervised adaptive fusion framework for satellite-based precipitation estimation without gauge observations

Satellite-based precipitation estimation plays a crucial role in climate change assessment and water resource management, benefiting from its wide coverage. However, the systematic bias and random errors of satellite precipitation product impose limitations on its application, making it necessary to use gauge observation based correction methods to improve the precipitation estimation. While correction methods are effective, they are limited to the gauged regions and pose challenges for sparsely gauged and ungauged regions. To address these limitations, we propose a novel unsupervised adaptive fusion framework that fuses multi-source satellite precipitation data via an adaptive fusion network in an unsupervised manner. Specifically, the proposed framework involves an unsupervised optimization manner to optimize the network parameters, leveraging unsupervised learning without ground-based gauge observations. The adaptive fusion module is designed to dynamically select the most informative features from different precipitation data, ensuring optimal satellite precipitation data fusion. Experiments using precipitation data in China from 2015 to 2019 demonstrate that the proposed framework improves the quality of satellite precipitation data. The fused precipitation product exhibits enhanced spatial accuracy and better consistency with gauge observations, surpassing the performance of the original products and even matching the quality of gauge observation corrected product. It improves the precipitation estimation without gauge observations, and thus provides valuable insights for climate and water resource management in ungauged regions.