Intelligent retrieval of sea surface wind fields in the typhoon area of the Northwest Pacific Ocean from dual-polarization SAR data

The Northwest Pacific Ocean is prone to typhoons; hence, high-resolution wind field data are essential for understanding their formation and aiding disaster prevention in China's coastal areas. Spaceborne synthetic aperture radar (SAR) is currently the only means for detecting large-scale, fine-grained typhoon wind fields. Traditional SAR techniques have limitations, such as challenges in determining wind direction and reliance on external inputs for wind speed. Most methods utilize single-polarization data, restricting their ability to capture a broad range of wind speeds. Although deep learning has shown promise in SAR wind field retrieval, research in this area remains preliminary and often neglects the benefits of dual polarization images or the specific challenges posed by typhoon conditions. Therefore, it is necessary to further explore the application potential of dual-polarization SAR data and deep learning technology in obtaining sea surface wind fields in typhoon sea areas covering a wide range of wind speeds.

In this paper, we propose a wind field retrieval model based on deep learning using dual-polarization Sentinel-1 SAR data. We attempt to effectively capture spatial features at various positions in SAR images, enhance the significance of key features, mitigate interference from irrelevant information, and improve retrieval efficiency. We integrate attention mechanisms such as SKNet, ECANet, and CBAM into the ResNet18 architecture to develop an E-SKNet_wind model. The performance of the developed models under different polarizations (VV, VH, and VV+VH) is systematically evaluated through comparisons with the ResNet18_wind models and the results reported in the literature.

The statistical results show that the precision of both types of deep learning models (E-SKNet_wind and ResNet18_wind) is higher in VV+VH dual polarization data than in either VV or VH single polarization data. Further, the dual-polarization E-SKNet_wind model performs better than the dual-polarization ResNet18_wind model. For wind speed retrieval, the root mean square error (RMSE) of the dual-polarization E-SKNet_wind model is 1.49 m · s^-1, which is smaller than that of the dual-polarization ResNet18_wind model (1.86 m · s^-1). For wind direction retrieval, the dual-polarization E-SKNet_wind model has an RMSE of 19.03°, which is smaller than that of the dual-polarization ResNet18_wind model (22.38°). In addition, the dual-polarization E-SKNet_wind model performs better than nearly all traditional methods and most existing machine learning and deep learning wind speed retrieval models. However, a few models may outperform our model potentially because of the narrower wind speed range or the inclusion of external wind direction data as an input parameter. The results of a case analysis show that the retrieval results of the typhoon wind field from the dual-polarization E-SKNet_wind model follow a trend consistent with the wind field from the ERA5 reanalysis across almost all regions. However, in the regions characterized by exceptionally low wind speeds, such as near the center of the typhoon, there is a notable and significant overestimation of wind speed values. This discrepancy results in a discontinuity in the retrieved wind speed profile. Future research and solutions are necessary to address these issues.

The dual-polarization E-SKNet_wind model effectively leverages spatial texture features from dual-polarization SAR images to precisely extract wind speeds and directions without any external input, thus overcoming the limitations of single-polarization data. This model accurately extracts sea surface wind fields from SAR images for the Northwest Pacific Ocean typhoon sea area.