Author(s): Xiaoyu Liu; Huanyu Qi; Xuan Wang

Linked Author(s):

Keywords: Sea surface temperature; Multi-scale; Downscaling analysis; Empirical orthogonal function; Generative adversarial network

Abstract: The data obtained from in situ observations, satellite imagery, and numerical modelling serve as the primary means for understanding and analyzing the oceans. The resolution of these data determines the level of detail available for oceanic information. However, various ocean phenomena exhibit wide range of scales, from large-scale circulation to mesoscale eddies and small-scale turbulence. Understanding these diversities typically requires data at varying resolutions. Nevertheless, high-resolution data are often not widely accessible due to constraints such as technological limitations and computational power. One potential solution to this issue is downscaling analysis. To extrapolate high-resolution data from low-resolution sources, it is crucial to ascertain the homogeneity between the two datasets in terms of scale and governing evolution laws. Taking sea surface temperature (SST) as an example, this paper decomposes a multi-year SST dataset into spatial modes and temporal series through the Empirical Orthogonal Function (EOF) method. The inner spatial structure between data with disparate resolutions is then identified and mapped through the Generative Adversarial Network (GAN). This mapping structure, which is a pivotal aspect of downscaling analysis, can further serve as the foundation for deriving high-resolution datasets from low-resolution information.

DOI: https://doi.org/10.64697/978-90-835589-7-4_41WC-P2078-cd

Year: 2025