Terrain-aware cross-attention Swin transformer for digital bathymetric model super-resolution

High-resolution digital bathymetric models (HR DBMs) are crucial for accurate seafloor representation and are widely used in seabed geomorphology, marine mapping, and underwater exploration. However, the acquisition of HR DBMs is costly and limited, prompting interest in DBM super-resolution (SR) techniques. Existing convolutional neural network (CNN)-based SR methods often struggle to preserve detailed seabed terrain structures, particularly in regions with complex morphology such as seamounts, ridges, and trenches. To address these challenges, we propose a novel framework named multiple seabed terrain transformer (MSTT) for DBM SR. MSTT integrates a cross-attention mechanism to guide slope-aware feature extraction aligned with bathymetric features and employs a spatial-gate feed-forward network (SGFN) within both spatial and cross-spatial transformer layers to adaptively regulate feature flow based on the local terrain distribution. Additionally, a DBM-aware loss function is designed to enhance the model’s ability to reconstruct fine-grained seafloor features. The DBM dataset used is constructed from EPOTO 2022, with preprocessing including rule-based patch cropping, terrain complexity stratification, and slope calculation via raster gradient. Extensive experiments conducted on 2× and 4× upsampling tasks show that MSTT consistently achieves state-of-the-art performance across quantitative and visual metrics. Notably, MSTT demonstrates superior reconstruction in complex terrain zones and robust scalability across resolutions, making it a reliable and generalizable solution for bathymetric enhancement.