Multi-scale boundary-enhanced diffusion network for high-resolution sea surface height reconstruction

Accurately reconstructing information-rich high-resolution sea surface height (SSH) from low-resolution SSH data has long been a critical challenge in oceanography. Despite significant advances in generative methods, most existing approaches fail to adequately capture the multi-scale nature of oceanic phenomena and struggle to resolve high-frequency features such as small-scale vortices and boundary currents, particularly at high sampling factors. To address these challenges, we propose a multi-scale boundary-enhanced diffusion network (MBD-Net) for SSH super-resolution. The key innovation of our method lies in the design of the contextual squeeze and excitation pyramid pooling module, which efficiently captures local and global contextual information across multiple scales, enabling the model accurate reconstruction of fine-scale structures while preserving large-scale oceanic patterns. Additionally, we propose the boundary enhanced channel attention block, which improves the model's sensitivity to high-frequency details (particularly around complex vortex boundaries) while strengthen its robustness by mitigating noise as well. Experimental evaluations show that MBD-Net outperforms existing methods, achieving average structural similarity indexes of 0.983 for 4× and 0.962 for 8× super-resolution across multiple ocean regions. These results demonstrate the effectiveness and versatility of MBD-Net, establishing it as a promising tool for high-fidelity SSH reconstruction in ocean environment.