Machine Learning-Based Wind Direction Retrieval From Quad-Polarized Gaofen-3 SAR Images

In this article, an intelligent method for inverting wind direction from quad-polarized Gaofen-3 (GF-3) synthetic aperture radar (SAR) images is proposed. Specifically, 11300 acquired in wave (WAV) mode are used to retrieve the wind directions using a spectrum-transformation approach and prior information from European Centre for Medium-Range Weather Forecasts reanalysis at version 5 (ERA-5) data at 1-h intervals with 0.25° grids. The dependence of the wind direction on the polarimetric correlation coefficient (PCC) between the co- (vertical-vertical and horizontal-horizontal) and cross-polarization [vertical-horizontal and horizontal-vertical] channels is studied. It is found that the PCCs in four combination polarizations have asymmetric characteristics with respect to the wind direction with correlation coefficients of greater than 0.4 or less than −0.4. Following this rationale, the scheme for inverting wind direction from quad-polarized SAR is trained according to machine learning, in which the matrix PCCs, wind directions, azimuthal angles, and slopes from SAR intensity spectra at the peaks are used as inputs. Subsequently, this intelligent approach is applied to 1300 images in quad-polarization stripmap mode, and the retrieval results are validated against advanced scatterometer (ASCAT) measurements. The statistical analysis shows that the root-mean- squared error (RMSE) of the wind direction is 17.7°, the COR is 0.98, and the scatter index is 0.11. In addition, the wind speeds inverted using a geophysical model function CSARMOD-GF are compared with well-calibrated ASCAT products, resulting in an RMSE of 1.85 m/s, a COR of 0.78, and an SI of 0.28 for the wind speed. Thus, this article provides an automatic scheme for inverting wind from quad-polarized GF-3 SAR images without any external information.