Project context

Wind waves are surface waves that derive their energy and their geophysical properties from wind blowing over the sea surface. However, the local wind forcing itself is not sufficient to define the local wave properties. While the swell component can be rather well imaged and then estimated by SAR instrument thanks to its quasi-linear properties, the wind-sea part presents more complex non-linearities and are only partially or totally removed from the SAR spectral signature. SAR-derived measurements are therefore difficult to use given the partial and complex spectral coverage of the wind waves. Nevertheless, some initiatives have shown the possibility to derive a more exhaustive description of the ocean state, estimating the significant wave height, comparable to altimeters [1], [2], [3]. Yet, wind-sea component impacts the SAR spectral measurements and some SAR-derived parameters such as the sea surface wind, the azimuth cut-off and the wave age. Usually, with SAR observations, geophysical features retrieval applications are developed to make specific decisions based on predefined rules derived from human experience, and on an understanding of the phenomena that modulate these observations. However, while this mastery of SAR observation improves with time and feedback, the definition of theoretical models to accurately process the geophysical variables derived from SAR would require massive investment and is becoming increasingly complicate.

After raising and confirming this finding, CLS as member of Sentinel-1 MPC was developed a new method to derive the wind sea Hs from L2 SAR features using a deep-learning (DL) approach. One of the advantages of such approach is that the classical statistical approaches require a lot of computational power on learning a very complex model while machine learning algorithms are more data intensive and only in the learning phase; their serving service is so fast. We note also that their results are too static to cope with dynamically changing service environments. Moreover, the greatest added value of machine learning algorithms is their ability to improve over time: Machine learning technology typically improves efficiency and accuracy through ever-increasing amounts of processed data. The proposed method has been tested on data from Sentinel-1A Wave Mode (WV) acquisition for both incidences angle 23.8° (WV1) and 36.8° (WV2) collocated with global numerical wave spectra given by WAVEWATCH III (WW3) over the period from July 2021 to August 2022. We used a Deep Neural Networks (DNN) to relate the input geophysical waves parameters derived from SAR observations and two targets :

1- Hs Wind Sea derived from delineated wind sea spectrum from W3 wave spectrum
2- Phs0 : the Hs Wind Sea partition given by WW3

[1] Li, X. M., S. Lehner, and M. X. He. “Ocean Wave Measurements Based on Satellite Synthetic Aperture Radar (SAR) and Numerical Wave Model (WAM) Data – Extreme Sea State and Cross Sea Analysis.” International Journal of Remote Sensing 29, no. 21 (2008): 6403–16. https://doi.org/10.1080/01431160802175546.

[2] Pleskachevsky, Andrey, Sven Jacobsen, Björn Tings, and Egbert Schwarz. “Estimation of Sea State from Sentinel-1 Synthetic Aperture Radar Imagery for Maritime Situation Awareness.” International Journal of Remote Sensing 40, no. 11 (June 3, 2019): 4104–42. https://doi.org/10.1080/01431161.2018.1558377.

[3] Quach, Brandon, Yannik Glaser, Justin Edward Stopa, Alexis Aurélien Mouche, and Peter Sadowski. “Deep Learning for Predicting Significant Wave Height From Synthetic Aperture Radar.” IEEE Transactions on Geoscience and Remote Sensing, 2020. doi: 10.1109/TGRS.2020.3003839

New Models Perfomances

1- Hs wind Sea from delineated Wind sea spectrum

For WV1

For WV2

)

2- Phs0

For WV1

For WV2

Install of s1_hswind_predictor

mkvirtualenv -p python3.9 $name_of_your_virtualenv
git clone git@github.com:s1tools/sar_hswind_predictor.git
git checkout develop
pip install -U setuptools
pip install -e s1_hswind_predictor

Test installation

cd s1_hswind_predictor/test
bin/python3.9 test_prediction.py

launch inference

1- Refer to hswind_inference/cmdline.py to have a clear view of the needed arguments

2- The algorithm request a file text that list the L2 OCN OSW products (netcdf files) to be inferred (required)

3- The algorithm needed AUX_ML2 to load the models and the normalizers (can be defined in the config file or in the cmdline). The AUX_ML2 can be founded in : https://sar-mpc.eu/

4- some other accessory arguments could be filled to define the format in which the results will be saved, ...

/!\ the default configuration is available in config /hswind_inference/config.py

python run_inference.py --list_l2_nc path/to/file.txt