A two-stage model for spatial downscaling of daily precipitation data

Providing reliable and accurate high-resolution meteorological data is very significant to guide the rapid response to extreme weather conditions. However, due to the constraints of computing power and simulation time, the spatial resolution of existing all global climate models is low, and it is unable to provide meteorological data with more precise resolution at local scale. In this research, a deep learning downscaling model called two-stage multi-scale feature extraction network (TSMFN) is proposed. By combining ERA5 reanalysis data and terrain data, spatial downscaling of global precipitation measurement mission precipitation data is carried out. Specifically, in the first stage of the network, several multi-scale residual Inception blocks are used to extract multi-scale features of low-resolution precipitation data; several residual-based residual multi-scale cross blocks are used to fully excavate multi-scale features after the fusion of multiple data. In the second stage of the TSMFN, the output feature map after the fusion of the previous stage is fused with the high-resolution monthly average precipitation data, and several progressive multistage Swin Transformer blocks are constructed to overcome the problems that the reconstructed image of a general convolutional neural network is smooth and cannot reflect the real spatial distribution of precipitation. Finally, a hybrid loss function combining \(L_{1}\) $L_1$ loss function and focal frequency loss function is proposed to alleviate the ill-posedness of the downscaling. By comparing the proposed algorithm with some advanced deep learning downscaling algorithms, the results of the experiment show that the TSMFN model is significantly better in terms of generated image quality and multiple evaluation indexes, and the model has stronger generalization ability.