Enhanced Adaptive Graph Convolutional Network for Long-term Fine-grained SST Prediction

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 16; pp. 1 - 11
• Main Authors: Peng, Han, Jin, Chang, Li, Wengen, Guan, Jihong
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation models; Artificial neural networks; Convolutional neural networks; Correlation; Disasters; Dynamic graph; Embedding; enhanced adaptive graph convolution; enhanced adaptive graph convolution (EAGC); Feature extraction; Marine environment; Mathematical models; Ocean temperature; Predictions; Predictive models; Sea surface; Sea surface temperature; sea surface temperature (SST); Spatial discrimination learning; SST prediction; static graph; Surface temperature
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2023.3308033

Sea surface temperature (SST) prediction is an important task in monitoring marine environment and forecasting ocean disasters, and has received tremendous popularity in recent years. However, existing methods for SST prediction usually focus on short term prediction (e.g., predicting the SST in future 7 days) or long-term coarse-grained prediction (e.g., predicting the monthly average SST in the next half year), and cannot work well for long-term fine-grained SST prediction (e.g., predicting the daily SST in future 60 days) due to their inability of learning long-term fine-grained spatial and temporal dependencies in SST. To address this issue, we proposed an enhanced adaptive graph convolutional network (EA-GCN) for long-term fine-grained SST prediction. First, EA-GCN introduces a static graph to capture the long-term correlations between the SST in different sea regions and a dynamic graph to learn the short-term correlations. Then, EA-GCN designs the multi-branch spatio-temporal embedding to learn the patterns in SST at different time granularities. Finally, the learned patterns are fused to make long-term fine-grained SST predictions. According to the experiments on two real SST datasets, EA-GCN largely outperforms state-of-the-art SST prediction methods when predicting the daily SST in future 60 days.