Data Analysis
Remotely sensed ocean waves from European and Chinese space missions have significantly supplemented the insufficient coverage of traditional wave observations such as buoys. The accuracy of the wave remote sensing is not only important for the wave forecast, but also critical to the air-sea interactions, which impact significantly weather and climate projections. Therefore, it is of great importance to improve the accuracy of retrieved wave products before their usage. The assimilation technique is known by its efficiency to improve the forecast accuracy by using the observations. Beside of good quality observations, a proper method of assimilation is also needed to make the best use of the observations, which will greatly impact the assimilation effects. Deep learning, which is based on artificial neural networks, has proved its efficiency and effectiveness in computer vision, speech recognition and many fields. It is innovative to apply this powerful method into both improvement and assimilation of wave products retrieved from several sensors such as SAR, wave spectrometer and altimeters available on European and Chinese satellites. Based on the satellite missions of Sentinel-1A/B SAR (Synthetic Aperture Radar), ENVISAT ASAR, Sentinel-3A/B altimetry, HY-2A/B altimetry, CFOSAT wave spectrometer SWIM (Chinese-French Oceanic Satellite, Surface Waves Investigation and Monitoring), this proposal will focus on:1) Finding the key factors which affect the accuracy through the detailed assessments of wave observations from each mission. This is critical to decide which parameter should go into the deep learning network. 2) Designing and building Deep Neural Networks (DNN) to improve the accuracy of altimetry missions (delay Doppler altimetry mode) by doing related researches and experiments. The structure and hyper-parameter will be carefully tuned and optimized to obtain a robust correction model. 3) Developing a model to correct the spectra from SAR and SWIM missions based on deep learning. Processing the wave directional spectrum as an image, techniques such as CNN (Convolutional Neural Network) will be applied in the correction. 4) Bringing out and verifying a new method of assimilation by combining techniques from deep learning, including those used for the quality control procedure before the assimilation and the method of data merging.5) Reprocessing the data from multi-missions using the deep learning correction. By using the corrected wave data in the assimilation, a global wave reanalysis from the model MFWAM will be produced and will highlight the use of the new deep learning assimilation method.6) Providing a better estimate of Modulation Transfer Function (MTF) of SAR and SWIM by using deep learning technique from model wave spectra. To demonstrate the feasibility of these objectives above, a series of exploratory experiments have been done by our teams. According to the comparisons between the corrections and buoy observations, the DNN has removed 77% of bias and improve 22% of SI (Scatter Index) for the SWH (Significant Wave Height) from Sentinel-3A. The DNN also removes 98% of bias and 26% of SI for the SWH of CFOSAT SWIM Nadir observation. The directional spectra from CFOSAT SWIM can be also improved by the combination of CNN and DNN. The deep learning can total decrease 45% of bias and lower the SI by 35%. Therefore, the feasibility experiment results have proved that deep learning is very promising when applied in the improvement of wave products retrieved from different sensors such as SWIM, SAR, and altimeters. So it should now be extended to other European and Chinese satellite missions. Furthermore, the study by using the deep learning in assimilation system will also be investigated. This proposal is funded by National Key R&D Program of China on Global high-quality marine data assimilation research and operational application and other relevant programs.
