INTERGEO Agenda - english

This research focuses on the use of airborne global navigation satellite system reflectometry (GNSS-R) phase and excess Doppler characteristics for remote sensing and surface classification applications. The objective is to exploit the coherence properties of reflected GNSS signals in order to differentiate various reflective surfaces from their scattering behavior along the satellite traces. For this purpose, a statistical phase model based on the in-phase and quadrature components of the reflected GNSS-R signal is developed based on the geometry of bistatic radars and the principles of signal propagation. The phase behavior is modeled within a von Mises distribution framework, which captures the transition between coherent and incoherent regimes. In parallel, excess Doppler spectrum analysis is introduced using three quantitative criteria: peak amplitude of the dominant Doppler component, the number of significant spectral peaks, and the spectral dispersion across peaks. These criteria were used as input features for the classification experiments. A synthetic GNSS-R signal generation framework is proposed, relying on the multi-path reflection model. The generation methodology relies on large-scale parameter extraction, performed to retain the fundamental characteristics of surfaces during the experimental flight campaign. This allows the synthetic signals to maintain the properties of real GNSS-R measurements. This dataset is used to evaluate the discriminative capability of the proposed Doppler-based criteria under controlled signal conditions. The classification experiments are carried out across several surface types, such as land, sea, sand, and inland water bodies. The study evaluates the performance of supervised learning techniques using both engineered spectral features and raw signal representations. In addition, deep neural networks are assessed based on both synthetic and real GNSS-R signals to evaluate their ability to discriminate scattering patterns without feature engineering. All of the models achieved more than 98% accuracy when tested on synthetic test sets, therefore validating their learning ability. The crucial validation performed on real airborne data showed that XGBoost and 1D-CNN achieved 90% classification accuracy even though they had only been trained on synthetic signals. Random Forest and MLP had also been found to show high performance with 88% and 87% accuracy, respectively. These results validated both the physical realism in the generation of the synthetic data and the resilience of the classification paradigm.