AsyDM-SST: Asynchronous Diffusion Model based on Schrödinger Bridge for Sea Surface Temperature reconstruction

Sea Surface Temperature (SST) is vital for monitoring marine environments and studying climate change. However, satellite-derived SST images are frequently compromised by extensive cloud cover and noise, which challenge conventional interpolation and data assimilation methods. In recent years, deep learning-based approaches have achieved remarkable progress in the field of image inpainting, with diffusion models emerging as a particularly powerful class of generative models. However, traditional diffusion models like DDPM are inherently designed to learn a reverse process from Gaussian noise. In conditional settings for data restoration, the damaged input is used only to guide the denoising of random noise, rather than serving as the iterative starting point. This mismatch often leads to global structural deviations and local inconsistencies in the generated output. Additionally, their uniform time-step denoising strategy applies the same intensity of restoration to all pixels simultaneously. This inflexibility often leads to over-smoothed zones and a loss of critical small-scale features that are essential in SST. Thus, this paper introduces an Asynchronous Diffusion Model based on Schrödinger Bridge for SST reconstruction, denoted as AsyDM-SST. Our method comprises two main modules: The Data Preprocessing Module initializes missing regions using the weekly mean SST and generates an anomaly weight map to guide the subsequent asynchronous diffusion process; The Asynchronous Diffusion Schrödinger Bridge Module utilizes the anomaly map to assign different time steps to each pixel, enabling asynchronous region-adaptive reconstruction. This mechanism governs pixel-level forward and backward diffusion coefficients, establishing a direct joint distribution path between the initial and target SST data. Within this module, a U-Net predicts the score function under the constraint of the weekly mean SST, which is then used to compute the current reconstructed SST, enabling iterative refinement of the intermediate states toward the final reconstruction. We evaluated our method on the publicly available Level-4 SST products provided by the National Satellite Ocean Application Service. Specifically, under 68% cloud cover with Noise-to-Signal (N/S) ratios of 0.1, 0.2, and 0.3, our method achieves R 2 improvements of 3.0%, 3.4%, and 4.4%, respectively, along with R M S E reductions of 27.0%, 25.6%, and 29.9%, compared to the second-best approach. Furthermore, comprehensive ablation studies validate the contribution of each component in our framework.