Digital Twin-Assisted Deep Reinforcement Learning for Anti-Jamming in UAV-Satellite Integrated 6G Networks Under Mobile UAV Jamming

UAV--satellite integrated 6G networks are vulnerable to mobile jamming because of time-varying satellite visibility, aerial mobility, co-channel interference, imperfect channel state information (CSI), and partially observable jammer behavior. This paper proposes DT-HARL, a digital-twin-assisted handover- and jammer-aware reinforcement learning framework for predictive anti-jamming control. DT-HARL couples a synchronized digital twin and jammer-state predictor with link survival probability, handover-risk index, and topology-jamming resilience modeling to construct a predictive graph belief. A feasibility-masked hybrid actor jointly controls satellite association, UAV relay selection, channel assignment, transmit power, and handover, while a trajectory-level conditional value-at-risk (CVaR) critic suppresses persistent high-loss events. Simulation results over ten independent seeds show that DT-HARL achieves 3.56±0.08 Mbps throughput, 17.27±0.26 dB SINR, and a packet delivery ratio of 0.872±0.006. Compared with DT-DRL without trajectory-risk learning, DT-HARL improves throughput by approximately 11.3%, while reducing outage probability from 0.100 to 0.062, recovery time from 2.82 to 1.91 s, and trajectory-level CVaR from 0.161 to 0.102. The ping-pong ratio is further reduced from 0.077 to 0.041, while the topology resilience index increases to 0.826±0.009. Robustness, multi-jammer, ablation, and sensitivity evaluations further indicate consistent advantages under imperfect CSI, delayed telemetry, stronger jammer mobility/power, and heterogeneous attack strategies.