Superpixel-based change detection for GPR time-lapse slices using fuzzy c-means and the Markov random field method

Abstract

Ageing and complex underground utility infrastructure present a significant challenge for modern society, requiring long-term monitoring and maintenance to prevent economic and social costs associated with infrastructure degradation and failure. In this study, we proposed an unsupervised superpixel-based change-detection method using ground-penetrating radar time-lapse slices combining fuzzy c-means and the Markov random field model to investigate an invisible subsurface change due to buried void using time-series measurements. First, simple linear iterative clustering was applied to the difference image, which was generated using paired time-lapse images after intensity registration to create different scales superpixel maps. Then, fuzzy c-means clustering was used to generate superpixel-based change maps. Finally, the Markov random field model was used to integrate the information of adjacent neighbourhoods in three dimensions to iteratively refine the change map. We designed two underground cavities (one representing shallow local voids and the other representing voids near pipeline networks) to verify the capability and adaptability of the proposed method. The experimental results demonstrate the feasibility of the method, with F1-scores of 0.82, 0.69, 0.69, and 0.65 and kappa coefficients of 0.81, 0.69, 0.68, and 0.64. Our method represents a significant contribution to the field of GPR-based change detection and has the potential to improve the long-term monitoring and maintenance of complex underground utility infrastructure.