Validation of high resolution SAR winds fields obtained by Deep Learning

A Deep Learning methodology based on ResNet, recently developed to retrieve the wind direction exclusively from the SAR images at 500 m of resolution, produces wind fields with unprecedented spatial details. As a consequence, the classical validation method comparing the SAR-derived with model and in-situ winds is not sufficient because of the natural lack of small scale structures provided by the models and the limited spatial coverage of the in-situ data. This paper proposes a complementary approach to the classical validation, estimating the spatial gradients of SAR-derived wind direction θ and speed
$U$
and verifying their compatibility with the typical values obtained from experimental wind time series. Hence getting a consistency test of the spatial information of the SAR-derived wind fields obtained with the ResNet methodology. This analysis on five Sentinel-1 images over the northern Adriatic Sea shows a good compatibility of the local spatial variations of the ResNet SAR-derived wind fields with those derived from experimental time series. These results, together with the statistical agreement with model and in-situ data sets, enforce the reliability of the wind maps obtained with the ResNet methodology, which describe real features of the wind fields.