High resolution SAR winds by deep learning technique in coastal area and lagoon

We developed a ResNet methodology based on Convolutional Neural Network (Zanchetta and Zecchetto, 2021), able to estimate the wind direction at a spatial resolution of 500 m by 500 m without external information. The ResNet model can derive the wind field even in the absence of wind streaks, in presence of convective turbulence structures, atmospheric lee waves, and ships. It is indicated to extract wind information over small areas, as the example of Venice lagoon. In this work the wind fields have been producet using the directions from ResNet and the scatterometer-based Geophysical Model Function CMOD7 (Stoffelen et al., 2017). The possibility offered by ResNet led us to investigate the characteristics of the strongest winds blowing on the northern Adriatic Sea and Venice Lagoon, Italy. The area of interest is subjected to high spatial and temporal variability of wind, a peculiarity of many coastal areas, making it a very demanding site. The structure of the wind systems inside and outside the lagoon has been studied in terms of spatial variability of speed, direction and vertical velocity wek in the Ekman layer derived by ResNet. The layout of wek exhibits contiguous cells of upward and downward motion elongated orthogonally to the wind direction with periodicity of 5.4 km. This spatial variability seems to be a signature of the atmospheric Ekman pumping, produced by local variations of direction and speed. An example of results from ResNet and OCN is reported in Fig. 1, which shows the ResNet (left panel) and the OCN SAR wind fields over the Venice lagoon: differently from the ESA OCN winds, the unprecedented resolution obtained with the ResNet allows an exhaustive coverage of the Venice lagoon, making possible to investigate the spatial structure of wind fields. For instance, under northeastern storms (Bora), the wind speed increases from northern to southern lagoon by 30% in average, in agreement with a case study carried out on experimental data (Zecchetto et al., 1997). SAR winds derived by ResNet have been compared with the in-situ and ECMWF model data, showing on average, a 9% of underestimation and 7% of overestimation respectively, in the range from 4 ms-1 to 25 ms-1. The overestimation of SAR derived winds with respect to ECMWF confirms the results obtained in the Adriatic basin from comparisons between scatterometer and ECMWF winds (Zecchetto et al., 2015), while the underestimation with respect to the in-situ data conveys the difference between ECMWF and in-situ winds of ~10%. The importance of a correct determination of the wind direction has been tested by comparing the SAR wind fields produced using ResNet and ECMWF wind directions, which may differ locally up to ±30º: these discrepancies may produce local differences of wind speed as large as ±2 ms-1. Detailed analysis of selected cases raised the issue of the lack of data with true spatial resolution of O(1) km and within half hour from the satellite pass time necessary for exhaustive comparisons. References Stoffelen, A., Verspeek, A., Vogelzang, J., Verhoef, A., 2017. The CMOD7 Geophysical Model Function for ASCAT and ERS Wind Retrievals. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 10, 2123-2134, doi:10.1109/JSTARS.2017.2681806 . Zecchetto, S. , G. Umgiesser and M. Brocchini, Hindcast of a Storm Surge Induced by Local Real Wind Fields in the Venice Lagoon, Continental Shelf Research, Vol.17 No.12,1513-1538, 1997 Zecchetto, S., della Valle, A., De Biasio, F., 2015. Mitigation of ECMWF-scatterometer wind biases in view of storm surge applications in the Adriatic Sea. Adv. Space Research 55, 1291-1299. doi:10.1016/j.asr.2014.12.011 . Zanchetta, A. and S. Zecchetto, Wind direction retrieval from Sentinel-1 SAR images using ResNet, Remote Sensing of Environment, 253, 2021 (https://doi.org/10.1016/j.rse.2020.112178)