Based on the GPR A-scan data， in order to further implement rapid intelligent detection of highway diseases， first of all， through data collection， sampling， data pre-processing and expert interpretation， road disease datasets with labels were established. A comparative analysis on different diseases and its degrees of severity was carried out to fully explore the characteristics of underground diseases. Based on the dimensions of time and frequency domain， the energy， variance， kurtosis and log power spectrum of A-scan were selected as the features to research the distribution of various road diseases. Finally， a state-of-art classification named Extreme Gradient Boosting algorithm （XGBoost， Extreme Gradient Boosting） was introduced to train and classify the data. The results show that the XGBoost classification algorithm achieves the accuracy of more than 90% for voids， looseness， cracks recognition.