@article{article_1783547, title={Cross-Domain Transfer Learning for LEO Satellite Constellations: A Similarity-Based Multi-Source Approach}, journal={Politeknik Dergisi}, pages={1–1}, year={2025}, DOI={10.2339/politeknik.1783547}, author={Osmanca, Mustafa Serdar}, keywords={Transfer Öğrenimi, LEO Uyduları, Alan Uyarlama, Yörünge Tahmini, Uydu Konstellasyonları}, abstract={The rapid expansion of Low Earth Orbit (LEO) satellite constellations such as Starlink, OneWeb, and Iridium has created new opportunities for global connectivity while introducing major challenges in orbit prediction, traffic management, and resource allocation. Traditional orbit propagation models (e.g., SGP-4) and physics-informed approaches often fail to meet accuracy requirements due to atmospheric drag, space weather, and orbital heterogeneity. Although machine learning (ML) techniques show strong potential for improving prediction accuracy, their dependence on large, high-quality datasets limits their applicability to new constellations. This paper presents a similarity-based multi-source transfer learning (MSTL) framework that leverages orbital similarities across heterogeneous constellations to enable accurate orbital period prediction with minimal target data. Unlike conventional physics-informed feature engineering, which can degrade performance by up to 461%, our method employs a minimalist feature set (altitude, inclination, and eccentricity) directly extracted from Two-Line Element (TLE) data. Through similarity-driven source selection and filtered multi-source knowledge integration, the proposed framework reduces prediction error by 88.2% (RMSE = 0.045 min, R² = 0.9972) using only 25 labeled samples from the target constellation. The findings show that domain-aware similarity filtering outperforms complex feature engineering, challenging conventional assumptions about transfer learning in physics-based domains. This work offers a scalable, efficient, and practical solution for emerging LEO operators, enabling rapid model development without extensive data collection.}, publisher={Gazi Üniversitesi}