A Method to Estimate Sea Surface Wind Vectors Using Geostationary Satellites

Abstract

Sea surface wind (SSW) is an essential physical parameter in the ocean and atmosphere, playing an irreplaceable role in hydrological and energy cycles, as well as in global and local climate systems. Polar-orbiting satellite instruments can gather a large amount of SSW information by observing surface roughness or wave height. Although observations from polar-orbiting satellites can cover the globe, there is a significant temporal gap for observing a fixed region. However, a geostationary satellite enables a relatively high observation frequency to accomplish this mission. Due to limitations in resolution, power consumption, and other factors, it is difficult for geostationary satellites to directly retrieve SSW. Nevertheless, it can obtain wind vectors at different altitudes by tracking the movement of clouds or clear-sky water vapor gradients in continuous satellite imagery, which is called atmospheric motion vector (AMV). There is a strong correlation between low-level AMV and SSW, and SSW could be estimated from low-level AMV. Previous studies estimated SSW based on low-level AMV mainly by empirical methods, which is unable to take the variation of AMV with height and latitude into consideration. Therefore, a new method based on a fully-connected neural network (FCNN) is proposed to address this issue. The theory of atmospheric dynamics, which explains how wind varies with altitude and latitude, is referenced, and FCNN is constructed by selecting physical parameters with strong causal relationships from geostationary satellite AMV information. The experiment is performed using GOES-16 advanced baseline imager (ABI) visible band AMV with a resolution of 0.5 km. After completing the wind estimation and comparing it with data from 93 National Data Buoy Center buoys nearshore or offshore North America from 1 January to 31 December in 2021, results show that root mean square error (RMSE) of the estimated wind speed from FCNN is less than 1.5 m·s-1. This represents a reduction of up to 0.24 m·s-1 compared to the empirical model. Additionally, the estimated wind direction shows slight improvement compared to AMV. After applying the model to the vicinity of a hurricane, and comparing it with reanalysis information for a total of 13 hours for 3 North Atlantic hurricanes and 3 Eastern Pacific hurricanes in 2022, results show that RMSE of wind speed estimated from FCNN is less than 1.1 m·s-1, reduced up to 0.04 m·s-1 compared to the result of the traditional empirical model. Additionally, there is no systematic bias in the low wind speed range.

Keywords

• sea surface wind vector
• atmospheric motion vector
• fully-connected neural network
• low-level atmosphere