Applications of Deep Learning-Based Super-Resolution for Sea Surface Temperature Reconstruction

Abstract

Deep learning-based super-resolution (SR) methods have been widely used in natural images; however, their applications in satellite-derived sea surface temperature (SST) have not yet been fully discussed. Hence, it is necessary to analyze the validity of deep learning-based SR methods in SST reconstruction. In this study, an SR model, including multiscale feature extraction and multireceptive field mapping, was first proposed. Then, the proposed model and four other existing SR models were applied to SST reconstruction and analyzed. First, compared with the bicubic interpolation method, the SR models can improve the reconstruction accuracy. Compared with four other SR models, the proposed model can achieve the lowest mean squared error (MAE) in the East China Sea (ECS), in the northwest Pacific (NWP) and in the west Atlantic (WA), the second-lowest MAE in the southeast Pacific (SEP); the lowest root mean squared error (RMSE) in ECS and WA, the second-lowest RMSE in NWP and SEP. Additionally, ODRE model can acquire the highest or the second-highest peak single-to-noise ratio and structural similarity index in ECS, NWP, and SEP. Moreover, the number of missing pixels and SST variety are two essential factors in the SR performance. The proposed multiscale feature extraction process can enhance the SR performance, especially for small regions and stable SST regions. Finally, while a deeper network can be helpful in achieving SR performance, the approach of simply adding more dilation convolutions may not enhance the reconstruction accuracy.