A new approach for quantitative precipitation estimation from radar reflectivity using a gated recurrent unit network

Abstract

Accurate quantitative precipitation estimation (QPE) utilizing radar data plays a vital role in weather monitoring, hydrological modelling and prediction. In previous research, some regression techniques such as the reflectivity (Z) - rainfall (R) relationship and artificial intelligence methods were employed for radar QPE; however, the estimation results remain limited the spatial and temporal variability of rainfall results in large errors in radar QPE. Recently, an emerging deep learning approach that can be applied in hydrological areas for QPE, namely, the gated recurrent unit (GRU) network method, has been used. GRU networks have evolved from long short-term memory (LSTM) networks, with improved training speed and modelling accuracy. In this research, the GRU network is applied to QPE using the three-dimensional structure of reflectivity from the Constant Altitude Plan Position Indicator (3D CAPPI) for an urban catchment in Seoul, South Korea. Simultaneously, a “complete” 3D CAPPI based on a cubic grid of 4 × 4 km2 that is highly suitable for the study area is proposed. Based on the comparison of the results of the GRU model with those of the artificial neural networks (ANNs), recurrent neural networks (RNNs), and LSTM networks, and the Z-R relationship, namely the least-squares model (LSM) for 2 testing strategies including traditional test strategy and leave-one-out cross-validation (LOOCV) strategy, the results indicate that the GRU model is superior to other models in radar QPE, with improvements by 3.3 ∼ 46.2%, 25 ∼ 70% and 1.2 ∼ 44.8% in the root mean square error (RMSE), bias and correlation coefficient (CC). This study presents a valuable method for improving the accuracy of radar QPE that can be actively employed in the future.