Spatiotemporal Prediction of Harmful Algal Blooms Based on Spatiotemporal Attention and Transformer

Accurate HAB prediction is crucial for water pollution management, and the application of spatiotemporal prediction methods in HAB has not been fully explored. To further enhance prediction performance, considering the complex HAB motion patterns, a spatiotemporal prediction method named STA-MIMO based on spatiotemporal attention and transformer is proposed. The proposed spatiotemporal attention module (STA) comprises Temporal Attention Block (TA) and 4D Attention Block (4DA). TA is designed to select critical spatial features along the temporal dimension by assigning temporal attention weights. 4DA is designed to finely select critical spatiotemporal feature points by assigning 4D attention weights. The STA can adaptively highlight key local spatiotemporal features, benefiting the prediction of complex motion patterns. Experiment results on the Taihu HAB 2019 dataset demonstrate that the proposed method significantly outperforms the comparative methods in all metrics, validating its superior HAB prediction performance.