Applications of deep-learning on TRITON data: results and findings

This study investigates the impact of sea surface roughness, influenced by both wind and wave characteristics, on GNSS-R observations using TRITON data. Traditional wind speed inversion algorithms often simplify Delay-Doppler Maps (DDM) to represent sea surface roughness caused solely by wind. However, this research highlights the critical role that wave-induced roughness plays in the scattering of microwave signals. To analyze these effects, the study employs the ResNet18 deep convolutional neural network, which excels at handling the complex features present in DDM data. The model integrates parameters such as wave conditions and GNSS-R incident angles to extract relevant features, enhancing wind speed prediction accuracy. The research uses TRITON data along with ECMWF wind speed and sea surface parameters for model training and evaluation. The findings indicate substantial improvements in wind speed prediction accuracy when accounting for both wind- and wave-induced roughness. This comprehensive approach reduces prediction errors and provides more reliable data for applications such as weather forecasting and climate modeling. These results underscore the potential of deep learning to integrate detailed sea surface characteristics into GNSS-R observations, offering significant advancements in predictive accuracy and operational applications.