FC-ZSM: Spatiotemporal Downscaling of Rain Radar Data Using a Feature Constrained Zooming Slow-Mo Network

Abstract

This work studies the application of deep learning methods in the spatiotemporal downscaling of meteorological elements. Aiming at solving the problems of the single network structure, single input data feature type, and single fusion mode in the existing downscaling problem’s deep learning methods, a Feature Constrained Zooming Slow-Mo network is proposed. In this method, a feature fuser based on the deformable convolution is added to fully fuse dynamic and static data. Tested on the public rain radar dataset, we found that the benchmark network without feature fusion is better than the mainstream U-Net series networks and traditional interpolation methods in various performance indexes. After fully integrating various data features, the performance can be further improved.