Deep learning-based precipitation nowcasting integrating radar echoes and rain gauge data

For flood protection and disaster mitigation, reliable and accurate short-term heavy rainfall forecasts are essential. However, precipitation nowcasting is a very difficult task due to the great temporal and spatial variability of precipitation at small spatial scales. To improve precipitation nowcasting accuracy, a radar echo extrapolation model (SAAR-UNet) and quantitative precipitation estimation model (RST-RainNet) were developed. By using an ASPP module, residual convolution, and a spatial attention mechanism, the SAAR-UNet greatly improved the extrapolation ability for the variation of strong echoes. In order to accomplish high-precision precipitation nowcasting, RST-RainNet innovatively employed a dual-branch strategy to perform residual correction on the quantitative precipitation nowcasting results of the extrapolated echoes. Extensive comparative experiments demonstrated that SAAR-UNet outperformed other attention-based and convolutional variants in echo extrapolation. RST-RainNet outperformed the optimized Z-R relationship, machine learning models (MLP, RF, XGBoost), and the single-branch UNet-ConvLSTM model in precipitation nowcasting tests. RST-RainNet best captured the regional characteristics and temporal dynamics of precipitation while maintaining the highest correlation coefficients and lowest errors for both one-hour and two-hour lead times. The results illustrated that deep learning techniques which combine rain gauge observation data with radar echoes can greatly improve precipitation nowcasting accuracy.