GNSS signal quality is severely degraded in challenging urban environments due to NLOS receptions and multipath interference. With the rapid development of urbanisation, the emergence of new construction materials as well as the complex architectures of modern buildings, the physic-based GNSS signal propagation channel modeling has encountered the bottleneck especially for the local effects. That's why the data-driven approach is promising to mitigate GNSS NLOS in a robust way. This paper first analyzes the main limitations of the current existing AI-based GNSS LOS/NLOS classifiers. Accordingly, enriched features are proposed and selected by correlation and predictive power analysis associated with physical interpretations. Finally, eight Machine Learning-based LOS/NLOS classification models are trained and evaluated using the proposed feature on a huge real dataset of around 8 hours collected by a vehicle in different urban environments. The performances show that the Gradient Boosting model has the most balanced performance in terms of prediction accuracy (81.63%) and the prediction time, which is promising for real time implementation.