Low-level wind shear poses a significant hazard to aviation, especially at airports located on high plateaus and surrounded by complex terrain. In this study, we present a comprehensive analysis integrating Doppler Lidar and radiosonde measurements collected at the Xining Caojiapu Airport, situated on the northeastern Tibetan Plateau, during June 2022. The results indicate a remarkably high frequency of severe wind shear events (|Δv| ≥ 6 m/s), with an overall occurrence rate of 34% during the observation period. These events are predominantly confined to two distinct atmospheric layers: just above the surface and near the top of the convective boundary layer. The diurnal cycle of wind shear is closely associated with boundary-layer dynamics, exhibiting sharp increases after sunrise and pronounced peaks around midday, coinciding with enhanced turbulent mixing and surface heating. Case analyses further reveal that the most intense shear episodes occur at strong thermal inversions, where momentum decoupling produces thin, critical interfaces conducive to turbulence generation. In contrast, well-mixed convective conditions result in more distributed but persistent shear throughout the lower atmosphere. Diagnostic profiles of atmospheric stratification and dynamic instability, characterized by the Brunt–Väisälä frequency and Richardson number, elucidate the intricate interplay between thermal structure and vertical wind gradients. Collectively, these findings provide a robust quantitative basis for improving wind shear risk assessments and early warning systems at airports in mountainous regions, while offering new insights into the complex interactions between turbulence and atmospheric stratification.

Lidar technology is widely used in atmospheric detection, but its performance is often degraded by various noise sources such as solar background radiation and dark current noise. This paper proposes a novel lidar signal denoising method, to our knowledge, UNNP-TV, which combines untrained neural network priors (UNNP) with total variation (TV) regularization to directly reconstruct clean signals from noisy lidar measurements. Unlike traditional denoising approaches that rely solely on prior knowledge of the signal or pre-trained models, UNNP-TV exploits the inherent structural priors of neural networks, enabling effective noise suppression without explicit training. The integration of UNNP and TV regularization mitigates overfitting, stabilizes the denoising process, and enhances signal restoration efficiency. Experimental results demonstrate that UNNP-TV significantly improves the signal-to-noise ratio (SNR) while preserving signal integrity, even in highly noisy environments. Furthermore, the UNNP-TV method is applied to real lidar signals for denoising, and the denoising results show that the method can be adapted to noise suppression in a variety of detection environments and improve the accuracy of aerosol extinction coefficient inversion.