Machine Learning for an Advanced Level-4 Data Interpolation for Satellite Sea Surface Height Anomaly Using Satellite Altimetry and NCEI-Archived in situ DART Water-Level Equivalent Ocean Bottom Pressure and Tide Gauge Observations

Satellite altimeters provide global coverage of sea surface height anomalies (SSHA) with coarse time and spatial resolutions. As satellite altimeters have revisit periods greater than 10 days, a given day contains large gaps in a full two-dimensional field due to the limited number of satellite observations. For the reconstruction of SSH anomalies, an interpolation method such as optimal interpolation (OI) has to be used to derive the large gaps within the field of satellite observations. While popular, the OI method is prone to over-smoothing and faces challenges in resolving mesoscale features not observed by a passing satellite altimeter. In addition, it does not take advantage of the relationships of SSHA and other variables determined by the physical atmospheric and oceanic processes to improve interpolation. To overcome this issue, a machine learning algorithm, the random forest, is used for the improvement of interpolating satellite altimeter SSHA observations by deriving SSHA observations from the learned relationships with correlated variables instead of using the relationships between SSHA and other climate and oceanic observations. SSHA observations from multi-satellites (e.g., JASON 2 & 3, Sentinel 3A & 3B, CRYOSAT2, and SARAL, etc.) are used for training the model over the last five years. The SSHA from these satellites show significant correlations with the sea surface temperature (SST), surface pressure, wind stresses, and in situ tide gauge and water column level observations from the NOAA Deep-ocean Assessment and Reporting of Tsunamis (DART) records. The model will learn the relationships between the altimeter SSHA and these related observations. The study focuses on a region off the southern coast of Alaska where DART records are shown to have high correlations with the high frequencies of SSHA observations and shows that the DART records can play an important role in deriving SSHA. This presentation will show preliminary results of the model’s ability to interpolate SSHA using the atmospheric and oceanic variables with significant correlations to the altimeter SSHA.