POSTER Session 2

DEEP LEARNING-BASED RECONSTRUCTION OF SEA SURFACE TEMPERATURE VIA BLENDING MULTI-SATELLITE DATA

Spatiotemporal continuous sea surface temperature (SST) data is crucial for understanding bio-geophysical ocean phenomena. Existing operational gap-free SST products blend multi-satellite observations, leveraging statistical interpolation techniques such as optimal interpolation and variational methods. However, meteorological conditions significantly influence satellite data acquisition, resulting in data unavailability. Furthermore, conventional interpolation methodologies are subject to blurriness, thereby imposing constraints on attaining high-resolution, gap-free SST. To solve the above issues, this study presents a deep learning (DL)-based blending method to reconstruct high-resolution SST using multi-satellite observations. The DL blending model combined five satellite-based SST datasets from AHI, AVHRR, MODIS, VIIRS, and AMSR2 sensors that were onboard on Himawari-8, MetOp, AQUA, Suomi NPP, JPSS, and GCOM-W1, respectively. The proposed model consists of two stages: 1) a self-attention transformer network that reconstructs individual SST datasets; and 2) a novel U-Net for fusing reconstructed SSTs. The reconstructed SST exhibited fine feature-scale ocean phenomena such as fronts and eddies, demonstrating that it can simulate the high spatial variability of SSTs with a spatial resolution of 2km. The results showed high validation accuracy, resulting in a root mean square error of 0.55 °C when compared with randomly occluded data. The proposed blending method achieved promising reconstruction results and provided the potential to adapt to other ocean optical parameters, such as chlorophyll concentration.

*Sihun Jung, Ulsan National Institute of Science and Technology, Republic of Korea, https://orcid.org/0000-0002-3583-1707

Jungho Im, Ulsan National Institute of Science and Technology, Republic of Korea

Code of Conduct
Privacy Policy