An Improved Deep Learning-based Approach to Dual-Polarization Radar Rainfall Estimation for Landfall Typhoon in South China

Abstract

In recent years, the algorithms based on deep learning are proposed for quantitative precipitation estimation (QPE). Studies show that this approach is slightly better than the raindrop size distribution (DSD) fitting method. There is still room for improvement in these approaches, when estimating heavy rain. As we all know, precipitation is close related to geographical location, however these deep learning methods are location-independent. In this study, an innovative approach is proposed for dual-polarization radar (DPR) QPE, which is location-dependent. In particular, the entirely new radar-gauges dataset (RGD) is reconstructed for QPE using radar raw data, radar observational information, gauges measurements and digital elevation model (DEM) data of rain gauges during the landfall typhoons in South China. The deep neural network model is optimized and trained using the RGD. DPR data is fed into this model to estimate precipitation. Preliminary results show the promising performance of this novel method compared to previous QPE estimators of deep learning-based.