TY  - JOUR
AB  - In sixth-generation communications, non-terrestrial networks provide scalable solutions to bridge connectivity gaps, particularly in areas where terrestrial infrastructure is lacking. Low Earth orbit (LEO) satellites are central to this effort due to their reduced signal propagation delays, which enable real-time, mission-critical communications, especially during emergencies when ground infrastructure is compromised. In such dynamic environments, accurate channel prediction is essential to maintaining consistent performance under varying conditions. However, conventional channel prediction methods that rely on uplink-to-downlink reciprocity, common in terrestrial time-division duplexing systems, are ineffective in LEO scenarios due to rapid channel variations, Doppler shifts, and channel aging. These challenges are intensified during unpredictable environmental conditions, such as natural disasters. To address this, we propose a hybrid offline-online learning framework for real-time downlink channel prediction based on uplink data in LEO systems. This approach enables an adaptive response to dynamic communication conditions. Additionally, a split learning strategy distributes computational load between satellites and user equipment, avoiding overloading LEO platforms. Simulation results demonstrate that the proposed model achieves lower normalised mean square error across various LEO configurations. The results validate the effectiveness of the proposed model under dynamic 6G channel conditions.
AU  - Weththasinghe, K
AU  - Ngo, QT
AU  - He, Y
AU  - Jayawickrama, B
DA  - 2026/01/01
DO  - 10.1109/TGCN.2026.3680486
EP  - 1
JO  - IEEE Transactions on Green Communications and Networking
PB  - Institute of Electrical and Electronics Engineers (IEEE)
PY  - 2026/01/01
SP  - 1
TI  - Split Learning-Based Channel Prediction for 6G-Enabled LEO Satellite Systems
VL  - PP
Y1  - 2026/01/01
Y2  - 2026/05/17
ER  -