Abstract: The large amount of missing remote sensing data caused by clouds and aerosols has a serious impact on marine research. Reconstructing the missing remote sensing data to obtain high spatial coverage data is crucial to support large-scale ocean environmental monitoring and climate change studies. Based on the structure of Data Interpolating Convolutional Auto Encoder (DINCAE), improvements have been made by replacing fully connected layers with convolutional layers, adding skip connections, and introducing an attention mechanism module. These improvements strengthened the model’s ability to extract spatiotemporal features from the data, resulting in the development of a Data Interpolation Convolutional Autoencoder with Attention Mechanism (A-DINCAE) model. The A-DINCAE model was used to reconstruct spatially complete SST data in the South China Sea and adjacent waters from 2015-2020, based on satellite infrared SST data with low spatial coverage. Compared to the Data Interpolating Empirical Orthogonal Functions (DINEOF) method and the DINCAE model, the results indicate that from 2015 to 2020, the A-DINCAE model achieves a root mean square error (RMSE) improvement of 0.10, 0.19, 0.17, 0.16, 0.06 and 0.17 °C for SST data relative to the DINEOF method, and improvements of 0.02, 0.09, 0.07, 0.06, 0.04 and 0.05 °C compared to the DINCAE model, respectively. The SST data reconstructed by the A-DINCAE model showed higher accuracy, especially in recovering small-scale information and gradient trends.