Correcting the bias of daily satellite precipitation estimates in tropical regions using deep neural network

Abstract

The high spatiotemporal variability of rainfall in tropical regions has posed a great challenge for generating satisfactory satellite precipitation products (SPPs). Most of previous studies have found a modest performance of various SPPs in estimating daily rainfall in tropical regions such as Malaysia. In-depth research on effective ways to correct the bias of SPPs in the tropical region is urgently needed. This study aims to establish a bidirectional long short-term memory recurrent network (Bi-LSTM) framework for the bias correction of SPPs, and apply it to correct daily rainfall estimates of the early runs of Integrated Multisatellite Retrievals for the Global Precipitation Measurement (IMERG-E) from 2010 to 2016 in the Kelantan River Basin, Malaysia. After optimization and statistical comparison, Bi-LSTM with the covariates of daily maximum and minimum temperature (Bi-LSTM-T) was determined to be the best model for bias correction. Annually, Bi-LSTM-T could raise the correlation coefficient (CC) of IMERG-E by 26.7%, while reducing its root mean square error (RMSE) and mean absolute error (MAE) by 23.9% and 21.7% in the Kelantan River Basin. In the four seasons, it increased the CC of IMERG-E by 19.7–27.5%, while decreasing its RMSE and MAE by 18.4–30.0% and 20.9–23.2%. Multiple statistical tests confirmed that Bi-LSTM-T significantly outperformed two benchmark methods, namely ratio bias correction (RBC) and cumulative distribution function (CDF) matching, in correcting the bias of IMERG-E for all four seasons. This suggests that the Bi-LSTM-T model may work as a promising framework of great potentials for correcting the bias of SPPs in tropical regions, where adequate precipitation data are in great need for diverse purposes such as water-related disaster prevention and mitigation.