CoAuthors

Gilles Guitton (OceanDataLab, France); Fabrice Collard (OceanDataLab, France)

The SWIM instrument is designed to provide directional ocean wave spectral measurements by means of an inversion of the modulation transfer function (MTF), i.e., the relationship between the measured backscatter modulation spectra and the surface ocean waves directional spectra. Current inversion algorithms, however, rely on either geometrical and/or geophysical models, which present limitations that may hinder the accuracy and precision of the inversion process. Recently, data-driven machine learning approaches have appeared as an appealing alternative to model-based approaches, allowing for the leverage of the ever-increasing availability of remote-sensing, in situ and model data to overcome the limitations of current models. Thus, we propose here to explore deep learning as a means to improve on current strategies for the inversion of the MTF. In particular, physically-informed machine learning approaches have shown great promise to exploit existing knowledge and produce better constrained models with enhanced relevance and interpretability. We therefore use classical statistical analysis to better understand the dependencies of the MTF on relevant geophysical parameters and complement the proposed data-driven models. Specifically, the MTF between SWIM cross-section modulation spectra in wavenumber space and WaveWatchIII wave model spectra in frequency space will be discussed. To this end, we exploit both model data, in the form of WaveWatchIII-derived synthetic ocean waves directional spectra, as well as real SWIM observations, alongside with theoretical MTF formulations, to evaluate the potential of deep neural networks to produce a more accurate, data-driven inversion of the MTF.