Deep learning-based gap filling for near real-time seamless daily global sea surface salinity using satellite observations

Abstract

Sea surface salinity (SSS) provides crucial information about ocean environments, influencing global hydrological cycles, thermohaline circulation, and climate change. Although L-band passive microwave radiometers have provided satellite-based SSS data, there are gaps due to the limited daily coverage of the sensors. This study proposes a U-Net-based spatial gap-filling model for global SSS using Soil Moisture Active Passive (SMAP) satellite data. The proposed model utilizes SSS swath data from the target and past days to generate daily global SSS maps with full coverage, incorporating only past temporal information. Additionally, bias-corrected data using gradient-boosted regression trees (GBRT) are employed to reduce inherent errors in the SMAP SSS data. We designed 24 schemes using data from the past 3, 5, and 7 days for both GBRT-corrected and original SMAP SSS data, with one to four times oversampling for low-salinity water, where the number of samples is significantly small. Validation results using masked-out pixels indicate that the gap-filling models with GBRT-corrected SSS data from the past 3 and 5 days and four times oversampling yielded the best performance, with root mean square errors (RMSEs) of 0.388 and 0.413 psu, respectively. Compared with the in situ Argo data for 2020, the RMSEs were 0.237 and 0.241 psu for the two models, respectively, significantly outperforming the SMAP Level 3 8-day SSS, which requires future data (RMSE of 0.456 psu). Notably, these models successfully filled gaps over coastal areas where low SSS (<30 psu) fluctuated due to freshwater discharge from inland areas. This study proposes an effective, novel gap-filling method for generating bias-corrected global daily SSS without delay, meeting operational needs. Moreover, the proposed gap-filling framework can be applied to other environmental swaths or even grid datasets.