Modélisation des zones de réception de l’AIS enenvironnement maritime par réseaux de neurones en graphes
Soutenue par Ambroise Renaud, à l’Université PSL le 28/11/2025. Dirigée par Aldo Napoli. Très bel accueil et discussions, en lien avec la thèse de Maryam. Rapporteur de thèse très agréable à rédiger. [theses.fr] [rapporteur & président du jury]
Maritime situational awareness, essential for safety at sea, relies on the Automatic Identification System (AIS), whose reliability is conditioned by radio propagation. This propagation depends on environmental factors, whose fluctuations lead to heterogeneous reception zones. The limitations of classical propagation modeling approaches in maritime environments, when faced with the complexity and variability of radio propagation phenomena, motivate this thesis to propose an alternative methodology.This thesis proposes a graph-based, data-driven methodology to model and predict AIS (Automatic Identification System) reception areas from environmental and geographic inputs. The developed framework constructs spatial propagation graphs on hourly gridded data and employs a deep architecture combining GraphSAGE, LSTM-based Jumping Knowledge and a multi-aggregation readout layer. Trained on real-world AIS data enriched with terrain and meteorological features, the model captures both local and long-range reception patterns. It outperforms classical baselines, including ITU-R P.2001 and XGBoost, in F1-score and accuracy. This work illustrates the value of deep learning for maritime situational awareness and highlights the potential of data-driven approaches in modeling propagation phenomena.