Precipitation nowcasting leveraging spatial correlation feature extraction and deep spatio-temporal fusion network

Precipitation nowcasting is crucial for various applications. However, existing deep learning models for meteorological applications face challenges regarding training efficiency, generalization of spatial features, and capturing long-range spatial dependencies. In particular, convolutional neural networks struggle to describe the complete spatial dependencies in radar echo reflectivity image sequences, making it difficult to model spatial features effectively. Additionally, current approaches using Encoder-Decoder structures based on recurrent neural networks have limited success in capturing global spatial dependencies and trajectory motion features in radar echo reflectivity images, especially for medium to high-intensity precipitation nowcasting. This paper addresses these issues by proposing a feature extraction method based on spatial correlation (FESC) and an end-to-end deep spatio-temporal fusion network (DST-FN) for precipitation nowcasting. FESC divides regions based on spatial correlation features extracted from radar echo reflectivity image sequences, improving the model’s understanding and prediction ability of meteorological data. We also introduce a Spatial Attention Mechanism (SAM) module into the TrajGRU model for better performance by adding a new memory channel. The proposed DST-FN framework utilizes the features extracted by FESC and temporal information, overcoming the limitations of encoding-decoding structures in precipitation nowcasting. Our approach demonstrates improved efficiency and effectiveness in capturing complex spatio-temporal dynamics compared to existing deep learning models.