Abstract: Sea Surface Temperature (SST) is a pivotal driver of weather and climate. Across the vast oceans, satellite-derived retrievals provide the primary source of global SST data. Yet frequent cloud cover can result in poor or invalid SST data if the retrievals are not well designed and work under all weather conditions. In addition, the inherent imager gaps between polar-orbiting satellite scan swaths leave extensive SST voids. To fill these SST data voids, the present work proposes a Contour-Aware Reconstruction Network (CARNet) specifically designed for improving spatial SST distributions. To achieve this, the continental contour is first utilized to eliminate the land-sea mixing field of views. The morphological dilation operators are then developed to enhance edge-gradient learning capabilities for extensive missing regions. The zonal and meridional gradient constraints are integrated into the loss function to preserve latitudinal-longitudinal SST distribution patterns. This algorithm is applied to FY-3D MWRI-derived SST products from January to December 2023 to reconstruct a gap-free, twice-daily SST dataset spanning 60°S—60°N with global coverage. Evaluation demonstrates that the proposed reconstruction method accurately restores mesoscale eddy features in dynamic current systems (e.g., tropical instability waves, Gulf Stream) while reducing positive biases in tropical cold tongues. Specifically, spring SST daily mean errors decrease from 0.8℃ to approximately 0.5℃, with a concurrent reduction of 0.1℃ in global daily SST standard deviation. The reconstructed dataset characterizes the evolutionary dynamics of the 2023 El Niño event, with computed Nino3.4 indices demonstrating high consistency with reanalysis benchmarks. This approach significantly enhances the operational usability of SST products retrieved from FY-3D MWRI data, providing reliable data support for ocean-atmosphere research.