ALGORITHMS AND PRODUCTS FOR EXTRACTING WEATHER-MARINE FORCING FROM EO DATA IN NEAR COASTAL WATER

Stefano Zecchetto

Institute of Polar Sciences, National Research Council of Italy, Padua (Italy)

Research theme aims

The ASI-CSA has shown the specificity of coastal with respect to the offshore areas with regard to wind field evaluation; also it has evidenced the need to develop appropriate algorithms for wind direction evaluation and exhaustive validations in different coastal areas. The work done in the framework of that project is continued in the CosteLab project over some of the selected areas of the project, with two main objectives: 1) refinement of the techniques developed and used in the ASI-CSA project; 2) validation of the results in the test areas of the CosteLab project.

Materials and Methods

In this project we used the available SAR images over the north Adriatic site because of the availability of in-situ data, i.e. the wind data from the CNR platform Aqua Alta, offshore the Lido inlet and the in-situ reports from the Istituzione Centro Previsioni e Segnalazioni Maree (ICPSM) of the Venice Municipality. Others coastal data are obtained from the Osservatorio Meteorologico Regionale of Friuli Venezia Giulia. To retrieve the wind field from SAR two things are mandatory: a suitable estimate of the wind direction (done by 2D-CWT [1]) and a reliable radar-backscatter versus wind speed Geophysical Model Function (GMT) (for C-band: [2,3] and for X-band: Li et al., 2012 [4]).

Once the relative scale-angle energy density Z(a,θ) is obtained (eq.1, where M(a,θ) is the scale-angle energy density), we have to select the scales a and angles θ relative to the wind structures, with the aim of reconstructing the SAR images with only these scales/angles, in order to make evident the shape and the layout of these structures.

       (Eq. 1)

This is accomplished taking the scale and angles exceeding the value of the 95% quantile of Z(a,θ), with the constraint that their location must form a spatially connected ensemble. The right panel of Figure 1 reports the wavelet spectrum derive from the 2D-CWT analysis of the SAR image reported in left panel, taken on the northern Adriatic Sea under north-easterly wind (Bora) the 24/04/2016.

Figure 1: On the left: Sentinel-1 SAR image of the northern Adriatic Sea under north-easterly wind (Bora) the 24/04/2016. The arrow reports the observed in-situ winds. On the right: the wavelet spectrum Z(a,θ) of the SAR image. The stars indicate the selected scales/angles used for the image reconstruction.

The wind directions derived from the direction of the major axis of elliptical cells, de-aliased according to the ECMWF mean wind direction over the area, allow to compute the wind speed by using the C-Band model CMOD5 [2]. The values of σ0 (radar cross section) used to compute the wind speed though CMOD5 are derived from the mean value of the backscatter inside each detected wind cell. The resulting wind field is shown in Figure 2, with the ECMWF model wind overlapped for comparison.

Figure 2: The wind field resulting from the 2D-CWT analysis of Fig. 1, with the ECMWF model wind and the in-situ reports overlapped for comparison (black arrows, not in scale). The white areas are those with the wind speed < 5 m/s. The in-situ winds have the wind speed reported sideways.

Results and Discussion

In this project we analysed 26 Sentinel-1A and B images over the northern Adriatic Sea from 2016 to 2019 taken under the Bora wind. The bora is a north-east wind classified as a downslope wind [5]. The wealth of information provided by the SAR derived winds cannot be adequately validated in this project. In this context the available in-situ data can be used only for local validation and not for the validation of the wind field extracted over the whole image. Therefore, we can speak about consistency instead of validation. The in-situ data for the consistency checks of the SAR derived winds have been obtained from different sources (see Materials and Method). They have different characteristics, namely different reference heights and averaging periods, however all reduced to the same time sampling of 1 hour. Before discussing the consistency between the SAR, ECMWF and in-situ wind directions, it is instructive to consider the angle difference between the in-situ and ECMWF wind directions. The biggest discrepancies may be due to different factors: the time difference between in-situ and the ECMWF wind fields (however < 90 minutes) as well as the different nature of the data (temporal and spatial averages). We have also to remind that the coastal areas are the locations where the spatial variability of the wind may be large, as shown in Zecchetto and Accadia (2014) [6]. Figure 3A reports the angle differences between in-situ and the SAR wind directions versus the in-situ wind speed. The results show a mean difference close to zero and a mean standard deviation of 27º. The comparison of the SAR with ECMWF wind directions provides essentially the same results (Fig.3B), with a mean bias of 2º and a standard deviation of 22º. They indicate a substantial agreement, even if the ECMWF wind direction distribution spans a wider range of angles.

Figure 3: A) in-situ minus SAR-derived wind direction difference versus the in-situ wind speed. The colors identify the in-situ data sources. ARPAfvg (blue) and Icpsm (green) are coastal data; BuoySlovenianNIB (red) and MedaAbate (black) are offshore buoys; Ptf (pink) is an offshore platform. B) The wind direction distribution of ECMWF (left panel) and SAR derived (right panel) winds.

The directions derived from the 2D-CWT provide reasonable results, but their detailed verification is not practicable with the data available at present. The 2D-CWT methodology developed for SAR direction estimation is known to have problems to retrieve the direction under meteorological situations characterized by large wind vorticity. To overcome this shortcoming, we have developed another methodology based on the Convolutional Neural Network (CNN), with a particular architecture called Residual Networks (ResNets) [10], which  solves this problem. Both 2D-CWT and CNN retrieve the aliased wind direction without the need of external information. An example is shown in figure 4: The 2D-CWT SAR wind reproduces well the north-eastern wind in the northern part of the area but fails in reproducing the probable south-eastern wind in the south part of the area. The CNN, on the contrary, reproduces very well the wind turning, however positioning the front more north than the ECMWF. This is probably due to the relatively coarse (3 hours) time resolution of the ECMWF data.

Figure 4: SAR wind fields interpolated over a 2×2 km grid. Left panel: derived by 2D-CWT. Right panel: derived by CNN. The black arrows (not in scale) represent the ECMWF wind and the in-situ data.

Future perspectives

The work done has shown the necessity of using suitable methodologies to retrieve the wind direction from SAR and hence the wind field at high spatial resolution without the use of external information. The performance of the presented methods does not depend on the SAR resolution and seems also to be insensitive from the SAR radar frequency and incidence angle. Further refinement and testing of both CNN and 2D-CWT methodologies is necessary, in particular in areas of sharp turning winds. Furthermore, it will be necessary to compare the 2D-CWT results with the other methodologies , such as the  Residual Networks (ResNets), which have been applied to Sentinel-1 SAR images over the Mediterranean Sea has shown very good results [10].

 Summarizing, the next steps which should be considered in further research are: 1) refinement and testing of 2D-CWT and CNN/ResNets methodologies; 2) development of a combined use of ResNet and 2D-CWT; 3) application of these methodologies to the COSMO-SkyMed SAR images.

Publications and Presentations

  • Zecchetto, S., WIND DIRECTION FROM SENTINEL-1 SAR IMAGES IN REGIONAL SEAS, International Geoscience and Remote Sensing Symposium 2018, Valencia, Spain, 23-27 July 2018 (978-1-5386-7150-4/18/$31.00).
  • Zecchetto, S., Wind Fields from SAR in Coastal Areas,Primo Workshop Nazionale: La Missione COSMO-SkyMed: Stato dell’Arte, Applicazioni e Prospettive Future, 13-15 November 2017, ASI, Rome, Italy
  • Zecchetto, S. (2018). Wind Direction Extraction from SAR in Coastal Areas. Remote Sensing, 10(2), 261.
  • Zanchetta, A. and S. Zecchetto, Wind direction retrieval from Sentinel-1 SAR images using ResNet,    Remote Sensing of Environment, 2021 (https://doi.org/10.1016/j.rse.2020.112178)
  • Zecchetto, S., Determination and use of the wind field from SAR in coastal areas, International Ocean Vector Wind Science Team Meeting, Barcelona, Spain, April 24-26, 2018

References

  1. Zecchetto, S. (2018). Wind Direction Extraction from SAR in Coastal Areas. Remote Sensing, 10(2), 261.
  2. Hersbach, H., Stoffelen, A., de Haan, S. (2007). An improved C‐band scatterometer ocean geophysical model function: CMOD5. Journal of Geophysical Research: Oceans, 112(C3).
  3. Stoffelen, A., Verspeek, J.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(5), 2123-2134.
  4. Li et al., 2012, Algorithm for sea surface wind retrieval from TerraSAR-X and TanDEM-X data, IEEE TGRS-2012-00997
  5. Atkinson, B. W., Meso-scale atmospheric circulations, Academic Press, London, UK, 1989
  6. Zecchetto, S., & Accadia, C. (2014). Diagnostics of T1279 ECMWF analysis winds in the Mediterranean Basin by comparison with ASCAT 12.5 km winds. Quarterly Journal of the Royal Meteorological Society, 140(685), 2506-2514.
  7. Cun, Y.L., (1989). Generalization and Network Design Strategies. Tech. rep.
  8. Goodfellow, I., Bengio, Y., Courville, A. (2016). Deep Learning. MIT Press.
  9. He, K., Zhang, X., Ren, S., Sun, J. (2016). Identity mappings in deep residual networks. In: European 494 conference on computer vision. Springer. 630–645.
  10. Zanchetta, A. and S. Zecchetto, Wind direction retrieval from Sentinel-1 SAR images using ResNet,    Remote Sensing of Environment, 2021 (https://doi.org/10.1016/j.rse.2020.112178)