This study aims to develop an intelligent method to detect subsurface anomalies for prediction and mitigation of geologic risks. We trained a convolutional neural network (CNN) model using the seismic data from 6 field cases of regular pipes and irregular cavities and a sliding time window technique to augment the datasets, and tested the CNN model using another 2 field cases. We derived probability distribution as an indicator of anomaly existence and validated high-value and low-value probability distributions corresponding to underground openings and surrounding disturbed zones, respectively. We found that the characteristics of subsurface anomalies determines the effective seismic features and influences the CNN model performance. Finally, we applied the CNN model to investigate a circular-bored tunnel and surrounding disturbed zones. Our results demonstrate that the CNN model is fast and accurate to detect the horizontal locations of subsurface anomalies, while the accuracy of vertical locations depends on the estimation of P-wave velocity. The intelligent method has the potential to identify hidden risks at early stages and to mitigate subsequent geologic hazards.