Radar precipitation nowcasting based on ConvLSTM model in a small watershed in north China

The spatial distribution and depth of precipitation are the main driving factors for the formation of flood disasters. Precipitation nowcasting plays a crucial role in rainstorm warning, flood mitigation and water resources management. However, high spatiotemporal resolution nowcasting is very challenging owing to the uncertain dynamics and chaos, especially at a small-scale region. In recent years, deep learning approaches were applied in precipitation nowcasting and achieved good performance in learning spatiotemporal features. In this paper, ConvLSTM model and sequences of radar reflectivity maps were used to forecast the future sequence of reflectivity maps with up to 2 h lead time in Liulin watershed with a small area of 57.4 km2. Dynamic hierarchical Z–I relationship was employed to calculate the forecasting precipitation and the forecasted spatiotemporal features were compared to the observed. The results indicated that the model can provide a well performance for the reflectivity above 10 dBZ with 0.70 of CSI for 30 min nowcasting and 0.57 for 2 h nowcasting, but was not good at forecasting the reflectivity above 30 dBZ with 0.38 of mean CSI for 30 min nowcasting and 0.12 for 2 h nowcasting, which have a decrease of 45.7% and 78.9%, respectively. The forecasted precipitation could truly show the details of precipitation spatial distribution and provide the accuracy of forecasting area with 49.2% for 30 min nowcasting. The satisfied areal precipitation depth could be offered basically with 26.3% of Bias for 30 min nowcasting in Liulin watershed.