Buoy Wave Height Measurements Improve SAR Estimates

Figure 18.  Comparison of Sentinel-1 (S-1) SAR deep-learning predictions of significant wave height Hs and buoy measurements. (b) scatter plot of SAR Hs vs buoy Hs. (d) RMS error of SAR prediction vs. buoy measurements as a function of Hs; error bars show one standard deviation. From Quach et al. (2020).

Synthetic Aperture Radar (SAR) sensors on satellites measure backscatter from the ocean surface and can be used to estimate wave height at very high spatial resolution (~10 m) relative to satellite altimetry. Two Sentinel-1 satellites of the European Space Agency (ESA) collected SAR measurements of the ocean surface from 2015-2018, together covering the entire globe every six days. Data-driven approaches to predicting significant wave height (Hs) from SAR have either used relatively limited in-situ data sets or used a wave model (e.g. WaveWatch-3) as the “training” data for a deep learning approach.  Quach et al.(2020) improve on previous approaches to estimation of Hs from SAR by creating a comprehensive in-situ observational record. They compiled data from the US National Data Buoy Center and Coastal Data Information Program, Canadian Marine Environmental Data Services, the international OceanSITES project, and the OOI. Surface wave data sets from the OOI Irminger Sea, Argentine Basin and Southern Ocean surface buoys were used. The authors note the importance of the Southern Ocean Array, where “many of the largest wave heights are recorded… [from] an under sampled region of the ocean.”

The comprehensive in-situ data set is split into separate training and validation segments. When SAR Hs from training data are compared to altimeter Hs from the validation segment, the deep learning algorithm shows root-mean-square (RMS) error of 0.3 m, a 50% improvement relative to prior approaches. Comparison with the buoy validation segment (Fig. 18) shows RMS error of 0.5 m. The authors attribute the increased error to the larger number of extreme sea states in the observations and the relative paucity of extremes in the training data.

Observational sea state information is critical for understanding surface wave phenomena (generation, propagation and decay), predicting wave amplitudes, and estimating extreme sea states. Thus, the improvement in RMS error using the deep learning technique notable. The availability of in-situ data from extreme environments such as those sampled by the OOI Irminger Sea and Southern Ocean Arrays are key to validation of these new approaches.