To address the delayed coverage decisions caused by dynamic ground-user demands and the conflict between self-organized network stability and coverage flexibility in infrastructure-limited environments, this paper proposes a dynamic network coverage method for heterogeneous multi-UAV systems based on dual-topology control. The method constructs a detection topology and a coverage topology for different functions. The detection topology supports network state collection and representation and maintains full connectivity, while the coverage topology provides regular network coverage services without enforcing full connectivity. On this basis, the method integrates long short-term memory networks and graph neural networks to represent network states. The LSTM module predicts future ground-user positions to improve the foresight of coverage decisions, and the graph neural network constructs a heterogeneous graph based on the detection topology and designs type-specific message-passing mechanisms for heterogeneous UAVs. Furthermore, the method establishes a heterogeneous reinforcement learning model under dual-topology control and designs differentiated reward functions for high-altitude and low-altitude UAVs. High-altitude UAVs focus on improving coverage-topology stability and maintaining full connectivity of the detection topology, whereas low-altitude UAVs focus on improving coverage quality. Experimental results show that, compared with the best results among existing baseline algorithms, the proposed method improves the coverage rate, network connectivity rate, and user throughput by approximately 23%, 22%, and 58%, respectively. These results demonstrate that the method effectively enhances coverage performance, connectivity stability, and throughput capability in dynamic heterogeneous UAV networks.