In this paper a combination of the multi-agent paradigm and a very well known clustering technique is used for unsupervised classification of subsoil characteristics working on a collection of ground penetrating radar (GPR) survey files. The main objective is to assess the feasibility of extracting features and patterns from radargrams. By optimizing both the field work and the interpretation of the raw images our target is to obtain visualizations that are automatic, fast, and reliable so to suitably assess the characteristics of the prospected areas and extract relevant information. The architecture of the system may be split into three interrelated processes: (a) pre-processing, (b) hierarchical agglomerative clustering, and (c) retrieval and visualization. The proposed system shows the viability of arranging GPR data from survey files into clusters, thus reducing the amount of information to be dealt with, while preserving its reliability. The system also helps characterize subsoil properties in a very natural and fast way, favors GPR files interpretation by non-highly qualified personnel, and does not require any assumptions about subsoil parameters. A powerful tool to analyze underground components in water supply systems is thus generated that acts in a non-destructive way and supports decision-making in water supply management.