MEPNet: a multielement prediction model for the Arctic sea area catering to spatiotemporal heterogeneity

The prediction of Arctic sea ice is of great significance for addressing global climate change, ensuring shipping safety, and promoting economic development. However, current methods still face two challenges: 1) It is difficult to capture spatio-temporal heterogeneity. Traditional networks have difficulty simultaneously taking into account the small-scale edge structures and large-scale spatial dependencies; 2) The cross-scale dependency relationships among multiple elements are complex. The differences in the importance of features from different modalities or branches are overlooked, making it difficult to adapt to dynamic scene changes. To this end, we propose a multi-element prediction model for the Arctic Sea area (MEPNet) based on spatiotemporal heterogeneity. Specifically, we design a dual-branch network, which includes a local detail modeling module, a frequency-aware dynamic filter module. The two branches are merged by an adaptive feature fusion module. The local detail modeling module enhances the ability to capture edges and tiny structures; the frequency-aware dynamic filter module, by introducing a frequency dynamic filter, captures the global context and structural information of images from the perspective of the frequency domain; the adaptive feature fusion module achieves deep feature fusion across scales and modalities for multi-modal information. We conducted a large number of experiments on the ARCTIC_MULTIYEAR_PHY_002_003 daily reanalysis dataset, and the results verified the effectiveness of the proposed network.