DINCP: Dual Interaction Network for Digital Bathymetric Model Superresolution via Codebook Priors

Digital Bathymetric Models (DBMs) are essential for marine exploration but often suffer from insufficient resolution due to the high costs of bathymetry measurements. Recently, deep learning-based super-resolution (SR) methods have been widely applied to reconstruct high-resolution (HR) DBMs. However, achieving high-quality reconstruction remains a significant challenge due to the ill-posedness of the SR problem. To address this challenge, we propose the dual interaction network (DINet) via codebook priors (DINCP). Instead of directly decoding low-resolution (LR) input to generate HR DBMs, we first pre-train a discrete codebook to capture high-quality terrain features (HTFs). These features are then used to replace the features extracted from LR DBM through feature matching, ensuring accurate reconstruction. Additionally, we introduce terrain gradient features and utilize the proposed feature fusion model (FFM) to alternately integrate spatial domain and the gradient domain information, thereby achieving refined terrain feature extraction. Extensive experiments demonstrate that DINCP outperforms existing DBM SR methods (CRAFT, RGT, SwinIR, STFET) across all evaluation metrics (elevation: 10.94 meters, slope: 3.85°, aspect: 61.96°). Specifically, DINCP improves elevation accuracy by 17.3% to 22.4%, slope accuracy by 22.0% to 24.8%, and aspect accuracy by 7.3% to 15.8%. Moreover, further analysis reveals that DINCP demonstrates superior robustness across various complex terrains and exhibits strong generalization ability. This study holds significant potential to advance oceanographic research that relies on HR DBMs, such as investigations into the generation of tidally driven internal waves, tsunami propagation, and sediment transport.