On the Effect of Polarization and Incidence Angle on the Estimation of Significant Wave Height From SAR Data

Significant wave height is an extremely important descriptor of the ocean wave field. We have implemented the CWAVE algorithm using linear regression, with elastic net term selection, and single-layer feed-forward neural network using buoy observations and RADARSAT-2 Fine Quad image data as model inputs. We used a number of standard performance metrics and found that the neural network models comprehensively outperformed the regression models. We explored the effect of incidence angle and polarization on model performance and found that the most accurate models were implemented within incidence angle bins between 1° and 2°, rather than including incidence angle as an independent variable. We found that the performance of copol (horizontal–horizontal, vertical–vertical, and RL) and hybrid-pol (right-circular-horizontal and right-circular-vertical) channels was comparable, and that these channels outperformed cross-pol channels (horizontal–vertical and right-circular–right-circular). The accuracy of our $H_{s}$ estimates was significantly higher than other published linear regression and neural network results. We demonstrate that a major factor in improving the accuracy of $H_{s}$ estimation is to use buoy observations rather that operation wave model hindcasts as training data. We demonstrate an application of our model by creating two high-resolution $H_{s}$ maps.