Leveraging ResUnet, Oceanic and Atmospheric Data for Accurate Chlorophyll-a Estimations in the South China Sea

Weiwei Fang ¹ Ao Li ^2,3* Haoyu Jiang ^3,4* Chan Shu ^5* Peng Xiu ^1*

• ¹ Xiamen University, Xiamen, China
• ² China University of Geosciences Wuhan, Wuhan, Hubei Province, China
• ³ Shenzhen Research Institute, China University of Geosciences, Shenzhen, China
• ⁴ College of Life Sciences and Oceanography, Shenzhen University, Shenzhen, Guangdong Province, China
• ⁵ Huanggang Normal University, Huanggang, Hubei Province, China

Chlorophyll-a (Chl-a) plays a vital role in assessing environmental health and understanding the response of marine ecosystems to physical factors and climate change. In situ sampling, remote sensing, and moored buoys or floats are commonly employed methods for obtaining Chl-a in marine science research. Although in situ sampling, buoys, and floats could provide accurate data, they are limited by the spatial and temporal resolution. Remote sensing offers continuous and broad spatial coverage, while it is often hindered by cloud cover in the South China Sea (SCS). This study discussed the feasibility of a predictive model by linking the physical factors (e.g., wind field, surface currents, sea surface height (SSH), and sea surface temperature (SST)) with surface Chl-a in the SCS based on the ResUnet. The ResUnet architecture performs well in capturing non-linear relationships between variables, with the model achieving a prediction accuracy exceeding 90%. The results indicate that (1) the combination of oceanic dynamical and meteorological data could effectively estimate the Chl-a based on deep learning methods; (2) the combination of meteorological and SST effectively reproduces Chl-a in the northern SCS, while adding surface currents and SSH improves model performance in the southern SCS; (3) With the addition of surface currents and SSH, the model effectively captures the high Chl-a patches induced by eddies. This research presents a viable method for estimating surface Chl-a concentrations in regions where they are highly correlated with dynamic factors, using deep learning and comprehensive oceanic and atmospheric data.