Prediction of red tide outbreaks in inter-connected coastal environments using time-series hyperspectral data and transformer-based graph convolution network

The accurate predictions on the red tide outbreaks in coastal regions can reduce their negative impacts on the marine environment and human life. Currently, the red tide prediction is generally accomplished by monitoring some related key factors, which are difficult to obtain on large spatial scales. Combining a transformer encoder with a graph convolution network (GCN), this study proposed an integrated model for red tide prediction that makes comprehensive use of the time-series hyperspectral data obtained through remote sensing methods. The topological graphs are constructed based on the multi-band spectral indices in the interconnected observation points, which are further analyzed using a GCN to obtain the topological features. After that, the temporal features of such topological graphs are extracted based on a transformer encoder, which are used for red tide prediction. The results show that the proposed model achieves reasonable predictions using the input period of 3 d before the date of red tide outbreaks, and the accuracy can reach about 92% with the input period of 5 d. The ablation experiments indicate that both the topological features obtained by the GCN and the temporal features obtained by the transformer encoder play significant roles in the prediction task of red tide outbreaks. The proposed model achieves the red tide prediction in interconnected coastal environments through the fusion of spectral-, topological-, and temporal features, and is expected to provide early alarms on red tide outbreaks for maritime and oceanic agencies.