Downscaling Spatial Patterns of Precipitation Data By Deep-learning-based Super Resolution

Abstract

Improving spatial resolutions of precipitation data is one of the most critical issues for disaster mitigations caused by heavy rainfall. In recent years, the deep-learning-based super-resolution attracts interests of researchers for predicting high-resolution image from coarser-resolution input data. This study proposes applying the super-resolution techniques for downscaling spatial patterns of precipitation data. Here we trained two deep learning-based super-resolution neural networks, SRCNN and SwinIR, so that the models predict the original high-resolution precipitation fields from coarsened rain fields. A series of experiments demonstrated that the two deep-leaning-based models outperformed the conventional bilinear interpolation method when sufficient training data are used for the training. Particularly, using SwinIR with the transformer, we successfully achieved super-resolution with higher accuracy than the SRCNN.