A Training-Free GPR Target Classification Method With Unsupervised Clustering Model

In recent years, deep learning methods have been playing an important role in extracting features from ground penetrating radar (GPR) images to detect underground targets rapidly. However, the shortage of GPR training data and the limited transferability of supervised deep learning models bring difficulties in generalizing the applications. To overcome the training-data dependence and enhance the model transferability, we propose an unsupervised GPR target clustering method based on the contrastive language-image pretraining (UGTC-CLIP) model to achieve a rapid target classification. The experimental results indicate that the proposed method can effectively classify different GPR targets with an accuracy of 0.88 without fine-tuning, which is comparable to the performance of the trained vision transformers (ViTs) model. Furthermore, the remarkable transferability of the model enables rapid adaptation to new samples or different geological conditions of GPR civil detection scenarios. This proposed unsupervised classification method is expected to be an efficient solution for real-time GPR data interpretation in transportation infrastructure inspections.