Precipitation Nowcasting Based on Radar Echo Images via Multiscale Spatiotemporal LSTM

Accurate precipitation nowcasting up to 2 h in advance is crucial for weather-sensitive decision-making, such as flood disaster warnings. Although deep learning-based radar echo extrapolation methods show advantages, they still face challenges, particularly in predicting local spatial variations and mitigating performance degradation with increasing prediction time horizon. This study proposes a new multiscale spatiotemporal (MS2T)-long short-term memory (LSTM) model to enhance precipitation nowcasting capabilities by fusing multiscale spatial features through parallel convolutional structures and improving spatiotemporal (ST) reasoning through stacked LSTMs. Notably, a multiscale spatiotemporal feature gate is introduced into the LSTM unit to transmit multiscale spatiotemporal features of radar echo image sequences, mitigating the issue of time-dependent performance degradation. The loss function for model optimization is carefully selected via a series of comparative experiments. The experiments show that MS2T-LSTM outperforms benchmark models in predicting precipitation events with different intensities. Specifically, in comparison to predictive recurrent neural network (PredRNN++), the $F1$ score, critical success index, Heidke skill score, and probability of detection (POD) for low-intensity precipitation threshold (5 mm/h) improved by 2.3%, 3.5%, 13%, and 10.6%. This study demonstrates the effectiveness of integrating multiscale spatiotemporal features to enhance precipitation forecasting, providing valuable insights for spatiotemporal prediction modeling and optimization.