Digital modeling of tunnels plays a critical role in the refined management of urban metro infrastructure. Backpack LiDAR systems, due to their portability and operational efficiency, show great potential for tunnel applications. However, the quality of point cloud data collected by such systems is often compromised by various factors during acquisition, leading to issues such as instability, uneven density, and structural distortion. To address these challenges, this study proposes a point cloud quality modeling framework that integrates statistical modeling with interpretable machine learning, and validates it through metro tunnel field experiments. Five key inspection factors were identified based on operational characteristics and tunnel environment. A total of 12 experimental groups were conducted, resulting in 360 point cloud samples. Four quality metrics were extracted and modeled using the Skew-Normal distribution to capture their statistical characteristics. A CatBoost regression model was then constructed to predict the distribution parameters, and the SHAP method was employed for global and local interpretability, revealing the causal pathways between inspection conditions and quality responses. The results indicate significant differences in how quality dimensions respond to inspection variables. Speed and scan density emerged as dominant factors across multiple metrics, while interaction terms had particularly strong effects on structure-related indicators. The proposed framework provides a quantifiable foundation for understanding point cloud quality, optimizing data acquisition strategies, which demonstrates strong engineering applicability and scalability.