Adaptive Satellite Selection via Deep Reinforcement Learning for Dynamic Emergency Scenarios

Disasters often lead to severe disruptions of terrestrial backbone networks, resulting in information isolation in affected areas. In such situations, very small aperture terminals (VSATs) can be rapidly deployed to establish terrestrial–satellite communication links, enabling the transmission of large volumes of situational awareness data, such as high-definition videos and three-dimensional maps. However, the high mobility of low earth orbit (LEO) satellites necessitates frequent satellite switching to maintain service continuity. This challenge is further exacerbated in emergency scenarios, where the timely and reliable transmission of post-disaster data is of critical importance, while unpredictable disturbances can severely degrade the stability of ground–satellite backhaul links through abrupt channel variations and resource fluctuations. Consequently, robust and efficient satellite selection is crucial for sustaining reliable and high-throughput satellite backhaul. To address these challenges, this paper proposes a disturbance-aware deep reinforcement learning (DRL) framework for adaptive satellite selection, aiming to support robust and high-capacity terrestrial–satellite backhaul for VSATs in disaster scenarios. Specifically, the proposed approach jointly accounts for link quality, handover (HO) frequency, and connection failures in the decision-making process. Moreover, the robustness and stability of the DRL model is improved by incorporating the stochastic disturbances encountered in LEO satellite networks into the training stage. These disturbances comprise weather-induced carrier-to-noise ratio (CNR) degradation and reductions in available satellite bandwidth caused by user contention. Simulation results demonstrate that the proposed approach achieves robust and stable performance under dynamic conditions, outperforming existing methods.