An Attempt at Detecting Reflection Images from Buried Objects in Ground Penetrating Radar Data using Anomaly Detection Methods with Convolutional Auto Encoder (CAE)

Ground Penetrating Radar (GPR) plays a pivotal role in road maintenance by efficiently detecting underground voids and buried objects. However, manual interpretation of the data collected, characterized by unique hyperbolic reflections from buried objects, is laborious. Machine learning methods, particularly Convolutional Neural Networks (CNN), have been actively proposed for detecting and interpreting these objects, yet they typically rely on supervised learning and extensive labelled data, which is burdensome.

In our approach, we apply Convolutional AutoEncoder (CAE) used in anomaly detection to learn only the background data in GPR exploration. This method identifies parts where buried objects may exist, reducing the need for creating labelled data and enhancing onsite flexibility by learning area-specific background noise.

We’ve successfully detected reflective anomalies from two-dimensional images obtained during subsurface exploration by learning parts without abnormalities. The CAE model, based on a VGG16-based SegNet, operates by dimensionally compressing the input image to match the output image. Anomalies become apparent when input images containing unlearned reflections aren’t restored fully, thereby indicating their location. The method holds promise for more efficient anomaly detection in GPR data.