Dual-path modeling of global swell and local wave features for short-term significant wave height spatio-temporal forecasting using hybrid convolutional networks

Ocean waves exhibit significant multi-scale characteristics due to nonlinear interactions among multiple physical factors, including long-period global swells and short-period local wind–waves. The core challenge in accurately predicting significant wave height (SWH) lies in simultaneously capturing the spatio-temporal evolution of both global swell dynamics and local wind–wave dynamics. Existing single-architecture models based on Convolutional Neural Network (CNN) or Graph Convolutional Network (GCN) struggle to balance local features and global patterns due to limited receptive fields or insufficient temporal sensitivity, particularly under highly dynamic short-term sea conditions. To address these limitations, this paper proposes a novel Dual-Path Hybrid Convolutional Neural Network (DHCNN) that integrates a local spatial feature extraction module with a global graph structure modeling unit, overcoming the scale adaptability bottleneck of traditional models. Quantitative evaluations on the East Asia-Western Pacific Region (EAWPR) and Northwest Pacific Region (NPR) datasets demonstrate that our proposed DHCNN model achieves substantial improvements over state-of-the-art baselines. Compared to the strongest CNN-based baseline (MA-ConvLSTM), DHCNN reduces the 24-hour root mean square error (RMSE) by 5.68% on the EAWPR dataset and by 21.31% on the NPR dataset, while yielding consistent Pearson correlation coefficient (PCC) improvements of 0.58% on the EAWPR dataset. Relative to the best-performing GCN-based baseline (STGCN), DHCNN further decreases the 24-hour RMSE by 6.67% and 21.95% on the EAWPR and NPR datasets, respectively, while maintaining high fidelity in both global contours and local details. Typhoon case studies confirm DHCNN’s capability to accurately characterize intense wave phenomena under extreme conditions. The results indicate that DHCNN’s multi-scale feature fusion mechanism significantly enhances SWH prediction accuracy and robustness in complex marine environments, providing critical technical support for offshore engineering safety.