Frontiers in Environmental Science (Jan 2023)
Long lead-time radar rainfall nowcasting method incorporating atmospheric conditions using long short-term memory networks
Abstract
High-resolution radar rainfall data have great potential for rainfall predictions up to 6 h ahead (nowcasting); however, conventional extrapolation approaches based on in-built physical assumptions yield poor performance at longer lead times (3–6 h), which limits their operational utility. Moreover, atmospheric factors in radar estimate errors are often ignored. This study proposed a radar rainfall nowcasting method that attempts to achieve accurate nowcasting of 6 h using long short-term memory (LSTM) networks. Atmospheric conditions were considered to reduce radar estimate errors. To build radar nowcasting models based on LSTM networks (LSTM-RN), approximately 11 years of radar, gauge rainfall, and atmospheric data from the UK were obtained. Compared with the models built on optical flow (OF-RN) and random forest (RF-RN), LSTM-RN had the lowest root-mean-square errors (RMSE), highest correlation coefficients (COR), and mean bias errors closest to 0. Furthermore, LSTM-RN showed a growing advantage at longer lead times, with the RMSE decreasing by 17.99% and 7.17% compared with that of OF-RN and RF-RN, respectively. The results also revealed a strong relationship between LSTM-RN performance and weather conditions. This study provides an effective solution for nowcasting radar rainfall at long lead times, which enhances the forecast value and supports practical utility.
Keywords
