Enhanced Sea Surface Salinity Estimates Using Machine-Learning Algorithm With SMAP and High-Resolution Buoy Data

Abstract

Despite the recent advances in satellite-based L-band microwave radiometry and retrieval algorithms to provide a unique capability to measure sea surface salinity (SSS) from space and explore its utility for understanding mesoscale dynamics, global ocean circulation, vertical mixing, evaporation rates and climate oscillations, SSS retrieval from L-band microwave radiometric data from the NASA-SMAP (Soil Moisture Active Passive) mission is often biased with systematic errors on larger temporal, spatial scales and random errors on short-time and length scales. To improve the SSS retrieval from SMAP data, we developed a robust algorithm based on a machine-learning approach with high-resolution in-situ data from the Ocean Moored Buoy Network in the Northern Indian Ocean (OMNI), which includes the Arabian Sea (AS) and Bay of Bengal (BoB)). The new algorithm was rigorously trained, tested and validated using the in-situ SSS time series measurements from the AS and BoB. Several sensitive variables were examined and used to improve the SSS estimatessuch as radiometric and ancillary data from the satellite observations, sea-surface wind and precipitation from the ERA5 data, and SSS/ SST from the OMNI buoy measurements. The OMNI timeseries measurements provided the spatially averaged satellite products to characterize the variability and gross features of SSS in BoB and AS waters on weekly, monthly, seasonal and annual time scales. The systematic validation of SMAP SSS products on a range of spatio-temporal scales showed that the new algorithm improved the SSS estimates by more than 15% in open ocean waters and 25% in river-discharge and precipitation-dominated regions in BoB and AS, when compared to the standard (operational) algorithm. Further analysis demonstrated that the new algorithm reduced clear biases and significant anomalies in SMAP SSS retrievals in the regions of river runoff, surface-freshened ocean and intense tropical cyclones, and captured synoptic/mesoscale SSS features and their seasonal variations in the North Indian Ocean.

Continuity and Enhancements in Sea Surface Salinity Estimation in the East China Sea Using GOCI and GOCI-II: Challenges and Further Developments

“…This comparative analysis indicated an improvement in the coefficient of determination (R 2 ), with a reduction in both the root mean square deviation (RMSD) and relative RMSD (rRMSD) for the adjusted GOCI-II Rrs (Table 2 and Figure 3). Machine learning approaches have been applied extensively across various domains, including satellite data analysis for SSS estimation [10,14,27,[39][40][41]. In this study, we employed a random forest (RF) model, the most widely used tree-based model in machine learning approaches.…”

Forecasting Buoy Observations Using Physics-Informed Neural Networks

“…Instead of directly solving intractable formulations like Naiver-Stokes or other prognostic equations for ocean modeling, a data-driven surrogate model is trained using the substantial amounts of historical training data available via numerical models or raw observations [12]. The use of observation assimilated models to train deep learning surrogates has been seen multiple times using both HYCOM [13] [14] and ERA5 [15] [16][17] models.…”