Dynamic low-altitude airspace partitioning and management strategy optimization based on graph neural networks and spatial cognitive constraints

The rapid expansion of urban air mobility operations demands adaptive airspace management approaches that transcend traditional static sectorization. This paper proposes an integrated framework for dynamic low-altitude airspace partitioning and management strategy optimization by fusing graph neural networks with spatial cognitive science. Urban low-altitude airspace is modeled as an attributed weighted directed graph encoding spatial adjacency, traffic flow coupling, and environmental constraints. A spatial cognitive constraint system is developed, quantifying boundary discriminability, shape complexity, and hierarchical cognitive load as differentiable optimization terms. A cognition-enhanced graph attention network architecture with spatiotemporal feature aggregation is designed to generate end-to-end partition assignments, with an explicit cognitive attention modulation mechanism that steers message-passing toward perceptually coherent neighborhoods. A reinforcement learning module, formalized as a Markov decision process with clearly defined state, action, and reward spaces, fine-tunes the partition and management strategy outputs through proximal policy optimization. The joint training scheme co-optimizes sector boundaries with capacity allocation, priority sequencing, and conflict alert policies through multi-task learning with curriculum scheduling. Experimental results on both synthetic simulation data and real-world ADS-B trajectory data from the OpenSky Network demonstrate that the proposed method achieves an airspace utilization rate of 81.2 ± 1.4% and reduces the flight conflict rate to 3.1 ± 0.5 per 100 flight-hours, representing a 46.6% improvement over vanilla graph attention baselines and outperforming advanced spatiotemporal GNN baselines including STGCN and DCRNN. Statistical significance is confirmed via Wilcoxon signed-rank tests (p < 0.01). A small-scale human-in-the-loop study with eight certified air traffic management researchers points to a roughly 37% reduction in operator decision response time, a signal we read as preliminary rather than conclusive given the modest sample. Scalability tests indicate that model inference alone stays feasible up to 10,000-node airspace graphs within the reconfiguration window, though we are careful to note that the complete operational pipeline has not yet been integrated end-to-end.