POSTER Session 3

A NOVEL DEEP LEARNING FRAMEWORK FOR HOURLY SEA SURFACE TEMPERATURE RECONSTRUCTION USING GEOSTATIONARY SATELLITE DATA

The diurnal variation of sea surface temperature (SST) is associated with changes in biological variables such as chlorophyll-a and primary productivity. However, discerning the diurnal cycle of SST from diverse heat flux sources requires substantial computational demand and complex parameterization based on numerical models. This paper introduces a novel generative adversarial network (GAN)-based reconstruction framework called physics-adjusted reconstruction adversarial network (PARAN), which effectively synergizes geostationary satellite data and physical knowledge from numerical models to reconstruct high-resolution hourly SST. When compared to drifting and mooring buoys, this framework has demonstrated high accuracy, as evidenced by correlation coefficients of 0.994 and 0.982, biases of 0.007 and -0.165, and root mean square errors of 0.435 and 0.766, respectively. The capability of PARAN for the accurate reconstruction of SSTs with high spatial variability, utilizing satellite-based data at a spatial resolution of 2 km, is highlighted. The precise reconstruction of a diurnal warming (DW) signal through the integrated use of satellite data and numerical models further underscores the high robustness of the reconstructed SSTs. The proposed framework can be extended to various ocean optical properties, including chlorophyll concentration, with diurnal characteristics dependent on solar insolation and SST.

Jungho Im, Ulsan National Institute of Science and Technology, Republic of Korea, [email protected], https://orcid.org/0000-0002-4506-6877

Sihun Jung, Ulsan National Institute of Science and Technology, Republic of Korea, [email protected]

Code of Conduct
Privacy Policy