Deep Learning for Super-Resolution of Meteorological Data

Abstract

For any country, understanding and interpreting precipitation dynamics is of great importance. In India, rainfall patterns profoundly influence daily livelihoods and economic development. Thus, it is essential to get cognizance of local rainfall for better policy making. Often, downscaling methods are used to produce high-resolution projections from low-resolution GCM outputs or observation data. This study applies a deep generative model called SRGAN to statistically downscale precipitation data from IMD over the Indian region. Our analysis shows that SRGAN performs comparatively better than other deep learning methods used for downscaling. We used SRGAN to downscale the precipitation data from 0.25° to 0.125° and 0.0625° resolutions and found that the downscaling results closely matched station observations. We also introduce a custom trained VGG based feature extractor that can act as a backbone for other DL models using meteorological data. Our study establishes that SRGAN can be used as a reliable statistical downscaling model. SRGAN yields result faster than the dynamical RCMs, allowing for more practical real-time applications.