Unmanned aerial vehicles (UAVs) face fundamental limitations in multi-scale obstacle navigation: onboard perception constraints (e.g., restricted sensing range and measurement noise) impede real-time large-scale obstacle avoidance, while existing methods predominantly focus on local planning in narrow environments. To overcome this gap, we propose Multi-Scale Switching Navigation (MSS-Nav), a unified multi-scale navigation framework integrating semantic decision-making with hierarchical motion planning. Our approach synergizes: (1) obstacle-oriented bounding boxes (OBB) for critical path extraction enabling early detours around large-scale obstacles, and (2) B-spline trajectory optimization for agile small-scale avoidance in cluttered regions. Crucially, we leverage vision-language models(VLM) for zero-shot scene understanding to dynamically switch between these complementary strategies. Experimental results demonstrate that MSS-Nav reduces computational latency by 82.4% and memory usage by 66.7% versus traditional hierarchical planners. In simulation with over 50 obstacles across diverse multi-scale environments, the framework achieved a 100% navigation success rate and real-time operation 16 Hz on an embedded Jetson Xavier NX. This work establishes a foundation for robust UAV navigation in complex urban environments (e.g., delivery in high-rise districts), potentially enabling safe high-speed operations in cluttered multi-scale spaces.