A Deep Learning-Based Method for Automatic Detection, Identification and Location of GPR Levee Hidden Hazards

When using GPR for levee hazard detection, traditional data interpretation is a heavy task due to its heavily manual reliance. Even the machine learning has been effective in road hazards detection by combining with the instrumentation and other auxiliary techniques, but the levee hidden hazards detection based on deep learning is still a challenge. In order to reduce human intervention and realize the fast and accurate detection of levee hidden hazards, an automatic detection, identification and location system based on YOLOv5 is developed in our study. The training and test results show that our developed method can not only learn different types of hazard features well, but also has the ability of single-target automatic detection and multi-target automatic identification of levee hidden hazards. To verify the effectiveness, the real GPR data were obtained and preprocessed from a measured levee, and the types and areas of hidden hazards in the levee were automatically identified and located by using the developed method. The results show that single or multiple levee hazard types and locations in the real GPR images can be automatically and accurately detected, identified and located at the same time. Our automatic identification and location system applied in real levee hazards detection further demonstrates the effectiveness and efficiency of the method.