Wind direction retrieval from Sentinel-1 SAR images using ResNet

This paper introduces a novel approach to estimate the wind direction over the sea from Synthetic Aperture Radar (SAR) images without any external information. The method employs deep residual network (ResNet), a variant of Convolutional Neural Network, to obtain high resolution (2 km by 2 km) aliased wind direction fields. Forty-seven SAR images of the European Space Agency satellites Sentinel-1 have been processed with ResNet, previously trained with other fifteen images. The areas of interest are the Mediterranean Sea and the Persian Gulf, two regional seas where the SAR images often present complex patterns associated to the wind field spatial structure reporting traces of the interaction with coastal orography, hence valuable test sites to evaluate the performance of the methodology here proposed. Statistical analysis was carried out comparing the SAR-derived wind directions with those from ECMWF atmospheric model, ASCAT scatterometer and in-situ gauges. It reports biases ? of -1.1o, 2.4o and -4.6o respectively, and centered root mean square difference cRMSd<21o, consistent with the benchmark obtained comparing scatterometer with ECMWF wind directions over the areas imaged by SAR (? 2.1o, cRMSd=19o). These results are relevant because they include the coastal data not accounted in the benchmark. Analysis of selected cases showed that SAR-derived wind fields reproduce meteorological situations characterized by strong divergence. Notably, our ResNet is able to estimate the wind direction even in the absence of wind streaks on the SAR images and in presence of convective turbulence structures, atmospheric lee waves, and ships. Furthermore, the model is also able to derive the wind field over small areas, as the example of Venice lagoon has shown. Detailed analysis of selected cases raised the issue of the lack of data with true spatial resolution of ?2 km and within half hour from the satellite pass time necessary for exhaustive comparisons.