A transfer learning-based contourlet residual-driven terrain-aware method for Arctic seafloor DEM super-resolution

Arctic seafloor Digital Elevation Models (DEMs) serve as fundamental data for marine scientific research, navigation route planning, and submarine resource exploration. However, high-resolution bathymetric measurement data are extremely scarce in Arctic regions due to severe environmental constraints. Additionally, the Arctic mid-ocean ridges and their peripheries exhibit diverse highly fragmented landforms, including rift structures, fault scarps, and iceberg plough marks. These conditions pose significant challenges for DEM super-resolution reconstruction. To address this issue, we propose a contourlet residual-driven terrain-aware super-resolution framework for Arctic seafloor DEMs. It first pre-trains on a global coastal terrain dataset to acquire generic terrain representation capabilities. Subsequently, it fine-tunes on the Arctic mid-ocean ridge dataset with hierarchical learning rates to adapt to specific tectonic-glacial geomorphic patterns and mitigate overfitting caused by data scarcity. For network design, we propose the Terrain-Aware Feature Aggregation Block (TAFAB), which explicitly embeds the spatial continuity and multi-attribute correlation characteristics of DEMs into a dual-branch attention mechanism to achieve collaborative extraction of multi-scale terrain structures. Meanwhile, we construct the Contourlet Refinement Gating module (CRG), which enhances direction-sensitive high-frequency edge information in the frequency domain through terrain complexity-adaptive Laplacian pyramid decomposition and learnable directional filter banks. Furthermore, we construct a joint optimization loss function for elevation-slope-structure and introduce an adaptive weight learning mechanism based on homoscedastic uncertainty to dynamically balance the optimization contributions of different terrain features. Experimental results on the Arctic mid-ocean ridge dataset demonstrate that compared to current state-of-the-art methods, our framework achieves 3.5% − 4.7% reduction in RMSE, 17.5% − 25.6% reduction in MAE, and 0.17–0.44 dB improvement in PSNR, providing an effective technical approach for refined seafloor terrain modeling in data-scarce regions.