Deep learning in spaceborne GNSS-R for ocean remote sensing: First insights from the AI4GNSSR project

Abstract

GNSS Reflectometry (GNSS-R) has emerged as a novel remote sensing technique for monitoring geophysical parameters. GNSS signals reflected from the Earth’s surface are tracked and measured by low-mass receivers onboard small satellites, providing abundant information about the target with higher sampling frequency and special coverages. The main observable of GNSS-R is Delay-Doppler Maps (DDMs), which map signal power at a range of delay and Doppler frequency shifts. The conventional retrieval algorithms rely on the parametric regression approaches inverting observables derived from the DDMs to the ocean wind speed products. Thus, GNSS-R has become a new technique for ocean wind retrieval and hurricane monitoring. With the large datasets of cost-effective GNSS-R measurements available, the AI4GNSSR project (Artificial Intelligence for GNSS Reflectometry: Novel Remote Sensing of Ocean and Atmosphere) was proposed to implement Artificial Intelligence for characterizing geophysical parameters and investigating new applications and approaches for the GNSS-R technique. In this study, A global ocean wind speed dataset is created by processing the observables of NASA’s Cyclone GNSS (CyGNSS) mission. The primary implementations of AI algorithms have shown great potential in improving the quality of the existing wind speed products. The deep learning model based on convolutional layers and fully connected layers processes the input CyGNSS measurements and directly extracts features from bistatic radar cross section (BRCS) DDMs. This model achieves an overall RMSE of 1.31 m/s compared with the ERA5 reanalysis data on an unseen dataset and leads to an improvement of 28% in comparison to the operational retrieval algorithm. Moreover, we found that data fusion with ancillary precipitation data is able to correct the rain effects, especially for high wind speed. For wind speeds larger than 16 m/s, our data fusion model outperforms the operational retrieval algorithm by 40%. For further validation of the model performance under extreme weather conditions, a case study of Hurricane Laura in August 2020 will be presented and discussed after a brief introduction to our models.