DYNAMIC SUPER-RESOLUTION NETWORK (DSRNET) FOR DOWNSCALING OF PRECIPITATION DATA

Abstract

There are many environmental applications that require precise and comprehensive precipitation values. However the satellite and climate model precipitation data are generally available at coarse spatial resolutions. This work investigates the reduction of gridded precipitation data to higher spatial resolutions by employing deep learning-based superresolution techniques. We utilized the latest advanced deep learning model for enhancing single images, known as the dynamic super-resolution network (DSRNet). We have utilized the DSRNet architecture for the precipitation downscaling task and have evaluated its performance in comparison to the previously employed super-resolution models. These results of our experiments are carried on the IMD precipitation dataset, which shows that DSRNet performs better than existing methods. We have achieved higher accuracy metrics while downscaling precipitation from coarse to fine resolutions using DSRNet.We have also incorporated meteorological factors as extra input channels, which enhances its capacity to depict the significant spatial patterns in the data. Our findings show the promise of utilizing advanced deep learning techniques for downsizing precipitation in the analysis of climate data.