Development of a Deep Learning Super-Resolution Generative Adversarial Network (SGAN) for Improving the Resolution of Global Precipitation Measurement (GPM) Dual-Frequency Precipitation Radar (DPR) Data

Researchers at the University of Alabama in Huntsville's Information Technology and Systems Center (UAH/ITSC) are using use deep learning technologies to downscale (increase) the spatial resolution of the GPM Dual-frequency Precipitation Radar (DPR) data. This process is called super-resolution and can be used to improve rainfall retrieval estimates from GPM data in areas where ground-based scanning radar data and reliable precipitation gauge observations are lacking. The focus of this presentation will be the development of a deep learning application that uses a Generative Adversarial Network (GAN) that consists of convolutional autoencoders to learn features that can infer high-resolution information from low-resolution variables. Our research uses coincident low- and high-resolution rain rate gridded data to train the GAN where the low-resolution GPM data is used as input and the corresponding higher resolution ground radar derived data is the target output. The GAN automatically extracts hierarchical features and learns the mappings between the sets of corresponding low- and high-resolution data pairs. The trained GAN can then be effectively used for super-resolution improvement of the low resolution GPM data when no higher resolution ground radar data are available. This presentation will discuss the complexity of using radar data in deep learning applications.