Inpainting of cloud-occlusion sea surface temperature image from a novel generative network using multi-scale physical constraints

Sea surface temperature (SST) is considered as an important environmental indicator, which has a wide impact on climate systems, marine ecosystems, and human activities. Due to the limitations of observation conditions, occlusion caused by clouds, and so on, SST data are often missing. Compared with traditional interpolation-based methods which reconstruct SST data only relying on current data, inpainting-based methods recently have successfully carried out the advantages of utilizing historical images to train a deep neural network (DNN), such as denoising diffusion model (DDM). However, although historical SST data can be used to better refill the missing data by considering SST data reconstruction as an image inpainting task that is constrained by semantics, SST images are mainly constrained by physical laws but semantics. Thus, this paper proposes a DDM-based inpainting network with multi-scale physical constraints. The proposed framework mainly consists of three modules including Global Average Estimation Module (GAEM) to generate an initial estimation of corrupted SST images by considering historical SST images as a global physical constraint, Local Deviation Repair Module (LDRM) to obtain the bias between the initial estimation and current SST image as local physical constraint, and Multi-Scale Fusion Module (MSFM) apply a multi-scale decoupling scheme to guarantee the efficient fusion of the "global" and "local" constraints and back to a complete SST image. Sufficient experiments have been conducted to verify the effectiveness and physical consistency compared with prior SOTA methods on the public AVHRR Pathfinder SST dataset and the results indicated the superiority of the proposal.