CFOSAT mission: Building multiparameter all-angle Geophysical Modulation Function for wind vector retrievals using a neural network approach

Abstract

<p>The Chinese French Ocean Satellite (CFOSAT) is an innovative space mission dedicated to the global observation and monitoring of the ocean sea state and the sea surface vector winds. CFOSAT operates two Ku-band rotating radars: the nadir/near-nadir Ku-band wave scatterometer (SWIM) and the dual-polarization, moderate incidence angle, Ku-band wind scatterometer (SCAT). This unique instrumental configuration provides regular collocated measurements of radar backscatter to retrieve sea surface state parameters, including significant wave height, directional wave spectrum, and wind vector. Observations for different incidence angles have a different sensibility to sea surface parameters: short and long waves, surface currents, surface temperature, etc. Particularly, synchronized backscatter from two different sensors could be mutually analyzed  to improve the quality of sea surface wind retrievals.</p><p>To relate collocated backscatter (σ0) measurements with multiple environmental parameters, a new Geophysical Model Function (GMF) was developed using a neural network approach. The new GMF is aimed to describe geophysical and instrument-specific properties for all satellite onboard sensors in a unified form. The traditional set of GMF variables (wind vector, incidence angle, polarization, ….) was extended with various additional geophysical parameters which can impact the signal properties: significant wave height, sea surface current vector, sea surface temperature, ice concentration, precipitation rate. The learning data set contained CFOSAT nadir/near-nadir and moderate angle measurements together with model data and provided with SWISCA S Level 2 product by IFREMER Wave and Wind Operational Center (IWWOC).</p><p>During the learning process, special attention was addressed to the normalization and uniformization of input values to avoid biasing and model overfit. As well, the numerical learning strategy was adapted to reduce the negative impact of using numerical weather prediction models (NWP) in the backscatter measurements regression task. The derived NN GMF reproduces the main features of NSCAT-4 GMF for moderate incidence angles and TRMM/GPM GMF for near-nadir observations. However, instrument-specific features (antenna gain distortions, noise, etc.) are clearly present as well.</p><p>The high volume of available data enables precise studies on the particular impact of different isolated geophysical variables on the backscattering coefficient value. However, a machine learning strategy should be adapted specifically to reduce possible biasing due to unequally distributed geophysical input variables, high dispersion between numerical models, and observation values. The resulting NN GMF could be considered as the approximation of the Ku-band radar cross-section as a function of a multi-parameter environment. In complimentary to wind/wave inversion tasks, the GMF can serve as a robust platform for rapid signal calibration and re-adjustment during mission exploitation. </p><p>The demonstrated results and model are aimed to extend the existing SCAT data processing with collocated SWIM nadir/near-nadir observations and additional NWP variables, to improve existing nadir/near-nadir measurement interpretation. As well, the proposed approach could be naturally extended for use with other scatterometer missions, e.g. MetOp ASCAT. </p>