Objectives In the context of the rapidly developing low-altitude economy, unmanned aerial vehicle (UAV) are widely deployed for tasks such as urban logistics, emergency response, and infrastructure inspection, which impose stringent requirements for continuous and high-precision positioning in complex urban environments. In dense urban canyons, global navigation satellite system (GNSS) signals suffer from severe non-line-of-sight (NLOS) interference due to low flight altitudes and restricted sky visibility. This frequently leads to increased positioning errors and abrupt positioning jumps, threatening the safety and reliability of UAV operations. Although deep neural network-based methods have achieved progress in NLOS detection, two critical challenges remain. First, existing approaches typically require fixed-dimension inputs, necessitating zero-padding for variable numbers of satellite observations, which introduces non-physical pseudo-observations. Second, these models exhibit insufficient generalization across different spatiotemporal scenarios, making it difficult to adapt to the rapidly changing low-altitude environments encountered during UAV flight.

Methods An edge-feature-encoded graph Transformer with sparse adapter (EGTSA) model is proposed to address the generalization challenges of GNSS NLOS detection. To avoid the distortion caused by forced dimension alignment, satellites observed within the same epoch are organized into a sky satellite graph. This allows variable-length GNSS measurements to be naturally represented through graph structures while flexibly characterizing inter-satellite geometric topological relationships. To enhance environmental representation learning, a graph node-edge attention (GNE-Attention) mechanism is introduced. By utilizing edge features to dynamically adjust inter-satellite similarity scores, the model can adaptively extract discriminative environmental features. Furthermore, a sparse adapter is designed to enable satellites to selectively attend to key environmental representations from the GNE-Attention output, thereby improving generalization in dynamic multipath environments.

Results Extensive experiments were conducted using real-world GNSS datasets collected in complex urban canyon environments at different times and locations. The experimental results demonstrate that the proposed EGTSA model achieves NLOS detection accuracies exceeding 93%, representing an improvement of approximately 3% to 20% compared with existing state-of-the-art methods. Furthermore, cross-time and cross-location tests verify the robustness of the method in complex low-altitude urban environments, demonstrating its strong generalization ability.

Conclusions A graph Transformer-based NLOS detection framework is proposed to address the issues of variable-length satellite measurements and poor cross-domain generalization. By modeling satellites as graph nodes and explicitly encoding geometric relationships via GNE-Attention, the proposed method preserves the intrinsic structure of GNSS observations and achieves adaptive environmental feature extraction. The superior detection accuracy and strong generalization performance across different urban environments, times, and locations provides reliable technical support for the stable and safe GNSS positioning of low-altitude UAVs.