The Airborne Maneuvering Network (AMN) is becoming an emerging field due to its wide-area coverage and localized service enhancement characteristics, in which Low-altitude Unmanned Aerial Vehicles (LUAVs) interact directly with ground-based devices after receiving commands from upper layers. However, the exponential increase in communication devices has led to a severe scarcity of spectrum for LUAVs. Furthermore, LUAVs communications are highly susceptible to interception by Eavesdroppers (Eves) due to the open characteristic of the wireless environment. Therefore, a secure spectrum sharing at LUAVs layer in AMN is studied. Moreover, to address the issue that the dynamic and heterogeneous characteristic of wireless environments presents significant challenges for resource allocation, a Digital Cousin based Q-learning (DCQ) method is proposed. Specifically, the original Probability Transition Matrix (PTM) obtained from sampling in the environment is transformed using the co-link method to obtain multiple virtual environments. Multiple agents are trained in parallel in multiple environments and the training results are fused to obtain the final Q function to output the policy of the original environment. The simulation results demonstrate that the proposed scheme can achieve more robust policies and faster convergence compared to conventional Deep Reinforcement Learning (DRL) methods.