Deep Learning Super Resolution of Sea Surface Temperature on South China Sea
John Julius Danker Khoo, King Hann Lim, Po Ken Pang · 2022
Surface temperature is one of the key observations to analyse the greenhouse effect on the Earth. The surface of the ocean can be captured using satellite sensors and transmitted to a meteorological center for real-time analysis. The use of the deep learning paradigm in super resolution has its potential in geoscience applications to increase the data transmission latency and enhance low-quality observation from remote sensing data. In this paper, the deployment of Generative Adversarial Network (GAN) architecture is studied to apply resolution reconstruction using the South China Sea sea surface temperature data. In addition, the development of spectral normalization is added to the Enhanced Super Resolution Generative Adversarial Network (ESRGAN) architecture to improve the training mechanism of generator and discriminator. This improved ESRGAN is compared with its super resolution performance against peak signal-to-noise ratio and structural similarity index evaluation metrics. The experiment shows that the low resolution of South China Sea data can be inferred to obtain a higher resolution with a more realistic resolution as compared to the conventional upsampling approaches.