Abstract: Detecting a potential scintillation event is precondition for any following countermeasures in GNSS applications. Performance of machine learning (ML) methods to identify scintillation events is analyzed for various scenarios with measurements from different periods, areas and observing systems. It shows that signal spectrum characterizes scintillation variation fundamentally. Accuracy of different methods based on ML is generally better than 90% from test data, with the best one 98.5% and the least one 91.3%. It is also found less spectral coefficients with a lower cut-off frequency for model training contributes to better performance, indicating the descending trend exists around Fresnel frequency is the essential one to distinguish a potential scintillation event. This give evidence that precise GNSS routine observations with sampling rate of 1 Hz might be used for scintillation recognition. When a set of parameters is adopted to fit spectrum feature and then used for ML training, better performance can even be expected.