Clustering approach for crack detection in GPR data: Influence of infilling material on detectability

Ground-Penetrating Radar (GPR) is a non-destructive method widely used for detecting cracks and other subsurface defects in concrete pavements. The effectiveness of GPR in detecting cracks can be significantly influenced by the infilling material present in the cracks. This work presents an approach based on clustering algorithms, to automatically highlight cracks in concrete slabs. In particular, the crack detection problem is cast as an anomaly detection scenario, where the clustering algorithm is used to discern between anomalous and non-anomalous A-Scans in a radargram, the former considered a potential sign of a crack. In order to cluster the time series of the A-Scans, the usage of a K-Spectral Centroid (KSC) algorithm is proposed, though a comparison with the popular K-means algorithm is also provided. Different crack sizes and infilling materials were simulated to analyze their influence on detectability. Understanding the properties of these materials and their impact on GPR signal reflection, absorption, and transmission is crucial for accurate crack detection. The results have demonstrated that infillings showing large dielectric contrast with concrete have enhanced crack detection. Moreover, GPR data collection was conducted in both perpendicular and parallel broadside orientations. The results obtained with the perpendicular broadside orientation are equivalent for both KSC and K-means, whereas for the parallel broadside orientation KSC outperforms K-means. According to these results, clustering algorithms can be successfully applied to the problem of crack detection from GPR data, thus contributing valuable information for decision-making in maintenance.