Abstract: Red tide is an ecological disaster caused by the excessive proliferation of photosynthetic algae in the ocean. The prediction on red tide occurrence can reduce its continuous damage, and thus is of great importance. This study proposed an integrated model that combines a graph convolution network (GCN) with long-short-term memory (LSTM) for red tide prediction in inter-connected coastal regions by jointly utilizing the of the spectral feature, temporal feature, and topological feature. The multi-band spectral indices in different observation points are used to form topological graphs, of which the topological features are obtained through a GCN. Consequently, the time-series topological graphs are analyzed based on the LSTM module to determine the temporal trends. The proposed model was tested using the in-situ hyperspectral observation and achieved the accuracy at about 91% using the input period of 5 days before the target date of prediction. The contribution of each module is evaluated via an ablation experiment, and the results indicated that both the topological and the temporal features play significant roles in the prediction task of red tide outbreaks. The proposed model can realize red tide prediction only using hyperspectral remote sensing technology, and is expected to provide information support to red tide monitoring and prevention.