In this paper, we use the deep neural networks (DNN)-based single image super-resolution (SISR) method for the super resolution of sea surface temperature (SST) data. In previous research, Convolutional Neural Networks (CNN), such as SRCNN, have been used for super resolution of precipitation data. By using state of the art DNN technology, we will be able to generate high quality super-resolution data. Then, we focus on RRDBNet and ESRGAN which are CNN and generative adversarial network (GAN) models for SISR, respectively. Actually, these models can generate much more high resolution data from ERA20C. In this evaluation, generated images are compared to OISST with the root mean square error (RMSE). We denoted comparison results of RRDBNet and ESRGAN as shown Figure 1. RRDBNet has a better RMSE than SRCNN and ESRGAN. But, CNN-based SISR model is not a faithful representation of the ocean currents of OISST. This means that CNN-based models does not recover the portion of the information removed by ERA20C. On the other hand, ESRGAN is able to represent the complex distribution of ocean currents, although its RMSE is less than that of CNN.

Comparison of CNN and GAN models for SST high resolution

Created in 1999, the Environmental & Water Resources Institute (EWRI) is the recognized leader within ASCE for the integration of technical expertise and public policy in the planning, design, construction, and operation of environmentally sound and sustainable infrastructure impacting air, land and water resources.

[Paper Proceedings](https://doi.org/10.1061/9780784483466)

Conference DOI: https://doi.org/10.26226/m.6329c39b68fb0e00131f77d2