Deep Learning in Spaceborne GNSS Reflectometry: Correcting Precipitation Effects on Wind Speed Products

Abstract

Deep learning techniques have shown the capability in GNSS reflectometry (GNSS-R) for retrieving geographical parameters based on GNSS-R observations. Recent studies have proved that such data-driven approaches can significantly improve the quality of ocean surface wind speed products retrieved from delay-Doppler Maps. However, based on the theoretical knowledge, several known error sources are associated with bias in the deep learning model estimations. Rain splashing on the ocean affects the surface roughness of the ocean, altering the scattering pattern of GNSS signals and consequently bringing in considerable bias in wind speed estimations. Correction of such bias is challenging because of its nonlinear dependence on different environmental and technical parameters. Deep learning has the potential to learn such trends from corresponding environmental parameters and correct the associated biases. Therefore, we investigate how deep learning-based data fusion using precipitation data can correct the rain effect and improve wind speed estimations. Our proposed fusion model outperforms both the baseline model and the operational Minimum Variance Estimator (MVE) method on unseen dataset. The root mean square error (RMSE) of our fusion model is 3.3% better than the baseline model and 30% better than the MVE method. For samples affected by rain, our fusion model also shows superior performance compared to the baseline model. Specifically, the retrieval RMSE of the fusion model is improved by 1.9% overall, with a 3.6% improvement in the low wind speed range (<4>16 m/s).