Abstract: Reliable shallow-sea bathymetry data plays a crucial role in marine ecological research, marine process analysis, and the support of human marine activities. Traditional bathymetric technologies, such as acoustic sounders and lidar, suffer from high costs and limited accessibility. Synthetic Aperture Radar - based bathymetry is susceptible to interference from environmental factors. Although the satellite-derived bathymetry (SDB) method using satellite multispectral imagery offers better cost-effectiveness, existing SDB models still have significant limitations: physical models rely on high-quality data and involve complex solving processes; traditional machine learning empirical models are constrained by the difficulty in acquiring in-situ measured data; and deep learning models such as convolutional neural networks (CNNs) and vision transformers (ViTs) struggle to simultaneously capture local details (e.g., shoals and reefs) and characterize long-range global trends. Additionally, most current algorithms lack verification of cross-island transfer generalization, which restricts their practical applications.To address the above issues, this study proposes an SDB model integrating deformable convolution and multi-scale features. The Internimage model is adopted as the backbone: its deformable convolution layers adaptively adjust the receptive field to accurately capture the local spatial heterogeneity of complex terrains, while the cross-attention mechanism is utilized to model long-range dependencies for optimizing global trend characterization. After feature extraction by Internimage, an improved Upernet model is employed to fuse multi-scale features, and a downstream bathymetry inversion head is constructed to further balance the perception of local fine-grained prediction and global trend consistency.In the data preprocessing stage, cloud masking is performed on multi-temporal Sentinel-2 L2A surface reflectance data, and in-situ measured elevation data are used as labels. Experiments were conducted in six typical global study areas (Oahu Island and Kauai Island in Hawaii, St. Croix Island, Vieques Island, Saipan Island, and Tinian Island) for independent fitting tests. Moreover, a "pre-training and few-shot two-stage training" paradigm was used to verify the cross-island generalization capability of the proposed model. The results demonstrate that the proposed model achieves significantly higher inversion accuracy than other baseline models. Furthermore, it exhibits excellent generalization performance on new islands through few-shot fine-tuning, providing a more efficient and reliable technical approach for practical bathymetric mapping in shallow-sea areas.

Key words: Bathymetric Mapping, Internimage, Upernet, satellite derived bathymetry, Generalization, Deep Learning