Accurate Oceanic Parameter Predictions via Spatiotemporal Graph Neural Network With Incremental Learning

Accurate oceanic parameter predictions have occupied pivotal roles in the fields of ocean research, enabling to protect ocean environment, regulate global climate, and prevent extreme weather. However, the current observational techniques typically struggle to acquire real-time subsurface oceanic parameters, and most of prediction methods usually rely on an offline learning (OL) paradigm. In our work, a novel spatiotemporal graph neural network with incremental learning (IL-STGNN) model is proposed. First, we provide a spatiotemporal relationship extraction main module (STREMM) to extract multi-coupled spatial features, and capture nonlinear temporal dependencies, via employing cascaded graph sample and aggregate (GSAGE), and long short-term memory (LSTM) structure, respectively. Second, we design a spatiotemporal relationship extraction auxiliary module (STREAM) to reconstruct real-time unknown subsurface oceanic parameters, and it combines with observed sea surface parameters. Third, we adopt an incremental learning (IL) paradigm with experience replay and adaptive weight assignment in the STREMM, which incrementally trains the combined oceanic parameters. Finally, the experimental results demonstrate that the proposed model surpasses the other baseline models in terms of prediction performance, when we apply the OL. Moreover, the prediction accuracy adopted the IL paradigm is superior to the OL paradigm in the proposed model. Hence, the proposed model provides an invaluable and promising insight for ocean climate forecasting.