Search

In this study, a machine learning approach is applied to wind speed retrieval from vertical-vertical (VV) polarized synthetic aperture radar (SAR) during tropical cyclones (TC). More than 2400 dual-polarized (VV and vertical-horizontal (VH)) Sentinel-1 (S-1) images acquired in interferometric wide (IW) and extra wide (EW) mode are collected in 2016–2022. Along-track observations from stepped-frequency microwave radiometer (SFMR) and soil moisture active passive (SMAP) are available for several images. The azimuthal cut-off wavelength (ACW) is derived from VV-polarized image and the Doppler centroid anomaly (DCA) is provided from Ocean (OCN) product. It is found that SFMR wind speed linearly relates with ACW and DCA. Following this rationale, a machine learning method, called eXtreme Gradient Boosting (XGBoost), is trained for constructing SAR wind speed retrieval geophysical model function (GMF) through abundant matchups from 2000 images collocated with wind speeds from soil moisture active passive (SMAP) microwave radiometer and SFMR, denoted as TCWIND2-S1. Then GMF TCWIND2-S1 is applied for more than 400 images and the validation against SMAP and SFMR products up to 70 m s−1 shows a 3.01 m s−1 root mean squared error (RMSE) with a 0.16 scatter index (SI) and a 0.92 correlation (COR). Retrieval is also compared with CyclObs wind products derived from more than 400 dual-polarized images, yielding a 2.66 m s−1 RMSE with a 0.19 SI and a 0.93 COR. Our work provides an approach for VV-polarized SAR wind retrieval at wind speed ranged from 0 to 70 m s−1 during TCs. This work is conveniently adopted for TerraSAR-X without operating at cross polarization channel. [ABSTRACT FROM AUTHOR]