Abstract: Tactical route correction for low-altitude logistics and public low-altitude route networks must balance interpretability, online latency, adaptive congestion avoidance, and safety fallback. Static shortest-path methods cannot perceive segment capacity variation and local congestion. Incremental search methods such as dynamic Dijkstra, D* Lite and LPA* respond to edge-weight changes but still rely mainly on deterministic global or incremental replanning, whereas end-to-end reinforcement learning suffers from limited interpretability and difficult safety certification. This paper models the low-altitude route network as a three-dimensional directed graph with time-varying non-negative weights that incorporate free-flow travel time, capacity utilization, altitude-transition cost, risk exposure and temporary restrictions. An event-triggered hybrid Dijkstra-DQN framework is then developed. Dijkstra provides an interpretable global backbone route, while a DQN agent is activated only within a bounded local subgraph when a rolling monitoring window detects excessive congestion or abrupt edge-weight changes. Action masking, anchor-node reconnection, local-step limits and dynamic-Dijkstra fallback are embedded to keep the learning policy bounded, auditable and recoverable. Controlled simulations using synthetic urban scenarios and simulated-realistic data show that the proposed method reduces average flight time and congestion exposure under local disruptions while maintaining lower online latency than repeated global replanning. The findings should be interpreted as mechanism-level evidence under controlled simulation settings rather than proof of operational effectiveness in real-world complex airspace.