GPS occultation detection technology, as an advanced atmospheric detection method, has been widely used in numerical weather forecasting, climate and space weather research. One of the problems in occultation detection is that it is easily interfered by the reflected signals on the earth"s surface. Identifying and separating the reflected signals in the occultation detection signal helps to assimilate the occultation data into the numerical weather prediction system, which has important significance. This paper proposes a deep learning model based on improved GoogLenet (Im-GNet) and applies it to COSMIC-2 occultation detection data to identify reflected signals. This article selects the COSMIC-2 occultation data (conPhs file) from January 1 to 9, 2020. After quality control, the radio holography method is used to obtain the spatial spectrum image of the occultation signal, and the Im-GNet deep learning model is trained , The accuracy rate of Im-GNet model test reached 96.4%, which is significantly higher than the result of support vector machine (SVM) method. This paper also analyzes the impact of reflected signals on occultation data. The geographic distribution of occultation events, and the refractivity comparison between the occultation inversion data (atmPrf file) and the NCEP (National Centers for Environmental Prediction) 12 hour forcast files (avnPrf file) shows that the quality of the occultation event data with reflection signals is better and the atmospheric information contained is richer.