Systematic scatterometer wind errors near coastal mountains

Satellite scatterometers provide the only regular observations of surface wind vectors over vast swaths of the world oceans, including coastal regions, which are of great scientific and societal interest but still present challenges for remote sensing. Here we demonstrate systematic scatterometer wind errors near Hawaii's Big Island: two counter-rotating lee vortices, which are clear in the ICOADS ship-based wind climatology and in aircraft observations, are absent in the Jet Propulsion Laboratory (JPL) and Remote Sensing Systems (RSS) scatterometer wind climatologies. We demonstrate similar errors in the representation of transient Catalina Eddy events in the Southern California Bight. These errors likely arise from the non-uniqueness of scatterometer wind observations, i.e., an "ambiguity removal" is required during processing to select from multiple wind solutions to the geophysical model function (GMF). We test two strategies to improve the ambiguity selection near coastal mountains, both of which "learn" the relationship between the wake winds and surrounding winds from a regional atmospheric model: a regression model based on maximum covariance analysis (MCA) modes that relate winds in the wake region to the surrounding region; and a machine learning technique that utilizes a deep neural network to relate winds in the wake region to the surrounding region. The deep neural network shows substantial improvement over the regression model, but still struggles to place the reverse flow at the correct position, due to internal variability of the lee vortices.