An attention mechanism based convolutional network for satellite precipitation downscaling over China

Precipitation is a key part of hydrological circulation and sensitive indicator climate change. The Integrated Multi-satellitE Retrievals for the Global Measurement (GPM) mission (IMERG) datasets are widely used global regional precipitation investigations. However, their local application limited by relatively coarse spatial resolution. Therefore, in this paper, an attention mechanism based convolutional network (AMCN) proposed to downscale GPM IMERG monthly data from 0.1° 0.01°. method end-to-end network, which consists cross-attention module, multi-factor residual comprehensively considering potential relationships between complicated surface characteristics. In addition, degradation loss function on low-resolution designed physically constrain training, improve robustness under different time scale variations. experiments demonstrate that significantly outperforms three baseline methods. Compared with in-situ measurements, normalized root-mean-square error decreased 0.011–0.045 real-data experiment. Finally, geographic difference analysis introduced further downscaled results incorporating measurements high-quality fine-scale estimation.