A Self-Attention Causal LSTM Model for Precipitation Nowcasting
Lei She, Chenghong Zhang, Xin Man, Xuewei Luo, Jie Shao
Abstract
Precipitation nowcasting is a critical task that can facilitate multiple applications including urban warnings and traffic. Deep learning methods combining convolutional neural networks and recurrent neural networks have been investigated for this problem. However, they always capture local and ineffective spatial dependencies through convolutional layers, while long-range spatial dependencies are critical for spatial applications. To improve long-range representations, we propose a spatiotemporal prediction model called Self-Attention Causal LSTM (SAC-LSTM), which utilizes a self-attention mechanism to model channel correlations. SAC-LSTM aggregates sequence features by extracting spatial features with global and local dependencies so that the visual details of each frame can be significantly preserved. Furthermore, SAC-LSTM is trained by a generative adversarial network with a learned perceptual loss, which improves the perceptual quality of predictions. Experimental results show that the proposed SAC-LSTM outperforms other models. Code is available at https://github.com/LeiShel/SAC-LSTM-MindSpore.