MGCNet: a multi-scale gated convolutional network for SST remote sensing image inpainting

Existing image inpainting methods are prone to boundary artifacts and texture distortion when dealing with cloud cover in sea surface temperature (SST) satellite remote sensing images. To solve the above problems, this paper proposes a multi-scale gated convolution network (MGCNet) for high- quality SST image inpainting. The backbone of the network is composed of eight multi-scale gated convolution (MGC) blocks, each of which can realize the feature learning at different scales of the image. Building upon this network, an expandable feature enhancement (EFE) module is designed to enhance the perception of crucial features such as temperature gradients and land-sea boundaries through convolution-based feature expansion. A location attention (LA) module is also designed to learn the correlation between local sea temperature and specific sea location by coding the whole target sea area in pixel-based2Dposition, especially to capture the distribution of local sea temperature that is higher or lower than the normal due to specific monsoon or ocean current. We conducted experiments on the European Centre for Medium-Range Weather Forecasts (ECMWF) SST remote sensing daily reanalysis dataset for the South China Sea and the Gulf of Mexico regions. Experimental results show that our method has better inpainting effect compared with the mainstream T-former, MISF, CTSDG, PiiGAN and so on. The source code and datasets are released and publicly available via https://github.com/Hanlin-Zhai/MGCNet.