Robust daily satellite sea surface salinity reconstruction using deep learning in low-salinity coastal regions

Abstract

Spatiotemporally continuous sea surface salinity (SSS) is essential for monitoring rapid changes in the physical and biogeochemical characteristics of oceans and plays a crucial role in effective coastal environment management. Traditional physics-based SSS methods often oversmoothed salinity variations and struggle to capture sharp gradients, leading to reduced accuracy in river-dominated areas. In addition, the long revisit times of satellite data limit real-time monitoring. To address these challenges, we propose the multi-scale aware interpolation network (MAIN), a self-supervised deep neural network designed to generate near-real-time, gap-free daily SSS data without relying on future observations. The input data consisted of soil moisture active passive (SMAP) L2B swath data and the bias-corrected SMAP data that were enhanced using in situ measurements. Validation against in situ measurements and existing satellite-derived SSS products demonstrates that MAIN significantly improves salinity estimation accuracy. In the Amazon region, operational products showed a root mean square error (RMSE) of 1.04 psu, while MAIN reduced this to 0.71 psu. In the East Asian region, the RMSE for operational products was 0.64 psu, whereas MAIN improved it to 0.51 psu. Feature representation analysis revealed that MAIN effectively captured the spatial expansion of low-salinity water over time and significantly reduced the salt-and-pepper noise typically associated with L-band measurements. These findings demonstrate that MAIN is a robust and scalable framework that enhances gap-free SSS estimation, offering new opportunities for monitoring transient oceanic processes in diverse marine environments.