Fluctuations in sea surface salinity (SSS) have significant impacts on marine ecosystems and global climate change. However, most existing SSS prediction models predominantly emphasized temporal dependencies while insufficiently accounting for spatial heterogeneity in SSS distributions, consequently constraining their predictive capabilities. To address this issue, this paper introduces a novel spatiotemporal Chebyshev graph attention network (SCGATN) model, which can model the intricate spatiotemporal interdependencies in SSS datasets, for enhanced SSS forecasting. The SCGATN model synergistically integrates spectrogram theory with a deep learning architecture and consists of three components, i.e. a graph neural network (GNN) module, an attention mechanism, and a Chebyshev polynomial filter. Among them, the GNN is used to construct an adaptive spatial correlation matrix, the attention mechanism is used to dynamically focus on important nodes, and the Chebyshev polynomial filter extracts local topological features through spectrogram convolution. Experimental studies were conducted in areas of both the South China Sea (SCS) and the East China Sea (ECS). Experiments conducted in areas in the SCS from 10 January 2010 to 29 September 2020 showed that the root mean square error (RMSE) of the SCGATN model over 1 to 12 days was reduced by approximately 21.66% to 83.16%, and by approximately 6.35% to 56.69%, compared with those of the graph convolutional network (GCN) model and the graph attention network (GAT) model, respectively. Experiments conducted in areas in the ECS showed that the RMSE of the SCGATN model over 1 to 12 days was reduced by approximately 2.17% to 89.69%, and by approximately 1.27% to 65.44%, compared with those of the GCN model and the GAT model, respectively. The experimental results demonstrated the promising capabilities of the proposed SCGATN model for SSS predictions.
Liu et al. (Fri,) studied this question.