<p>In urban canyons, severe multipath and Non-Line-of-Sight (NLOS) reception caused by buildings and foliage severely degrade GNSS positioning reliability. Currently, data-driven machine-learning models that integrate signal-to-noise ratio, elevation angle, and other features, are widely used to identify signal-quality levels in complex environments. However, existing methods primarily rely on three-dimensional (3D) city models or panoramic cameras for NLOS signal labeling, which are prone to mislabeling at obstruction edges. Moreover, line-of-sight (LOS) category contains a large proportion of multipath interfered signals, making conventional stochastic models ineffective. To address these issues, this study proposes a fine-grained GNSS signal classification and stochastic-model optimization framework that operates without external auxiliary information. The method employs IQR and Jenks Natural Breaks to automatically stratify observations according to pseudorange errors, with category-specific optimized stochastic models fitted accordingly. Concurrently, an XGBoost-based machine learning model incorporating multiple features is developed for robust signal classification, with adaptive weighting factors assigned to different signal categories to enhance positioning performance. Experimental results demonstrate that the proposed classification stochastic model achieves better consistency with observation true errors, while significantly improving data utilization efficiency and terminal positioning accuracy. In a low-rise district of Hong Kong, 92.6% of dynamic GNSS epochs achieved 3D positioning errors below 10 m, the 3D RMS error was reduced by 51.0% compared to traditional NLOS exclusion methods, and by 44.0% and 19.3% compared to conventional unified and LOS/NLOS two-class stochastic models, respectively. In high-rise city clusters, the horizontal RMS error decreased from 11.93 to 8.15 m, with a notable reduction in outliers in positioning results.</p>