Ilmu Komputer & AI editorial
Open AccessOA2026
Learning the LoS Skyline from LEO Satellite Observations for Proactive Handover
A map-free, binary-classification approach to reconstructing local obstruction skylines and predicting line-of-sight termination in non-terrestrial networks
Marius Corici; Manar Zaboub; Fabian Eichhorn; Hauke Buhr· 2026· DOI 10.48550/arXiv.2608.00125
The core problem
In non-terrestrial network (NTN) deployments, local obstructions such as buildings, terrain, or vegetation can block line-of-sight (LoS) satellite links before the satellite reaches the geometric elevation mask. This causes abrupt and unplanned handovers, degrading service continuity and reliability. Traditional mitigation strategies rely on 3D building maps, sky cameras, or additional environmental sensing, which are costly, difficult to scale, and often unavailable. The authors address this limitation by proposing a map-free method for learning the local LoS skyline, defined as the obstruction elevation over azimuth, directly from binary availability labels derived from passive satellite signal observations at the terminal. The core research question is: can the skyline be reconstructed and future LoS termination events be predicted using only the terminal's own signal observations and satellite ephemeris, without external environmental data? The paper formulates the problem as a binary classification task in the azimuth-elevation space, where the skyline is extracted as the decision boundary of the learned blockage probability surface.
Innovation
The authors evaluate the proposed estimators against empirical bracketing, a baseline method that likely uses simple thresholding or interpolation of observed blockage events. The results show that both the GP classifier and the MLP with circular azimuth encoding and Monte Carlo Dropout improve the skyline reconstruction compared with empirical bracketing. While specific quantitative metrics (e.g., accuracy, precision, recall, or skyline error) are not detailed in the abstract, the improvement is consistent across the evaluated scenarios. The learned estimators successfully capture the local obstruction profile from binary availability labels alone, demonstrating the feasibility of a map-free approach. Furthermore, the integration with EphemerisWindow enables proactive handover preparation by predicting future LoS termination events before the serving link is lost. This proactive capability is a key outcome, as it allows the terminal to initiate handover procedures in advance, reducing the risk of abrupt link failures.
In non-terrestrial network (NTN) deployments, local obstructions such as buildings, terrain, or vegetation can block line-of-sight (LoS) satellite links before the satellite reaches the geometric elevation mask. This causes abrupt and unplanned handovers, degrading service continuity and reliability. Traditional mitigation strategies rely on 3D building maps, sky cameras, or additional environmental sensing, which are costly, difficult to scale, and often unavailable. The authors address this limitation by proposing a map-free method for learning the local LoS skyline, defined as the obstruction elevation over azimuth, directly from binary availability labels derived from passive satellite signal observations at the terminal. The core research question is: can the skyline be reconstructed and future LoS termination events be predicted using only the terminal's own signal observations and satellite ephemeris, without external environmental data? The paper formulates the problem as a binary classification task in the azimuth-elevation space, where the skyline is extracted as the decision boundary of the learned blockage probability surface.
The problem is formalized as follows. Let denote azimuth and denote elevation. For each observation, a binary label indicates whether the satellite signal is available () or blocked (). The goal is to learn a blockage probability function and extract the skyline as the decision boundary where .
Why it matters
The paper's main contribution is a map-free, data-driven method for learning the local LoS skyline and predicting LoS termination events in NTN deployments. By formulating the problem as binary classification in azimuth-elevation space, the approach leverages passive satellite signal observations that are readily available at the terminal, avoiding the need for costly or unscalable environmental sensing. The two estimators offer complementary strengths: the GP classifier provides principled uncertainty quantification, while the MLP with circular azimuth encoding and Monte Carlo Dropout offers a scalable, neural alternative with uncertainty indicators. The combination with EphemerisWindow bridges the gap between static skyline learning and dynamic trajectory prediction, enabling proactive handover. Potential limitations include the reliance on sufficient binary observations to train the models, the assumption of a static skyline over the learning period, and the need for accurate satellite ephemeris. Future work could explore online learning, multi-terminal collaboration, and integration with standard handover protocols. Overall, the method represents a practical step toward autonomous, resilient NTN operations without reliance on external maps or sensors.
Who should read this
CS practitioners and researchers
Opening member content…