Low-altitude networks, which support communication among ground stations and aerial nodes such as UAVs, require real-time monitoring to ensure reliable operations under dynamic conditions. The highly mobile nature of UAVs, coupled with frequent topology changes and limited onboard resources, presents unique challenges for network monitoring. In-band Network Telemetry (INT) provides a promising approach to monitor low-altitude networks by embedding telemetry metadata directly into data packets, enabling fine-grained visibility into network behavior. However, traditional INT schemes introduce substantial bandwidth overhead, a critical issue for resource-constrained low-altitude links. Furthermore, existing optimization approaches typically rely on static configurations, which cannot effectively handle the highly dynamic and volatile conditions we observe in low-altitude network environments. We present MGAT (Multi-Granularity Adaptive INT), a lightweight telemetry mechanism for low-altitude networks that achieves efficient monitoring through decoupling control decisions from data collection operations. Our data plane design exploits the spatiotemporal characteristics of different telemetry data types through three specialized collection schemes: flow-level aggregation is used for statistical metrics, static attribute sampling handles time-invariant properties, and packet-level extremum tracking captures critical outliers. The control plane combines QoS-aware flow classification with load-adaptive sampling, allowing the system to dynamically adjust monitoring intensity according to real-time network utilization. We evaluate MGAT through experiments and find that it reduces overhead by 98.7% compared to traditional INT while maintaining high detection accuracy (F1-score of 0.965). When network load varies dynamically, MGAT demonstrates 8.3% higher accuracy than static baseline methods and consumes 11.8% less bandwidth, confirming its ability to adapt effectively to the complex dynamics of low-altitude networks.