Ground Penetrating Radar (GPR) is widely used in the fine detection of shallow structures such as urban road disease detection and archaeology, etc. Small-scale underground geological bodies such as cavities and cracks are usually the most concerned detection objects. In general, the energy of hyperbolic diffracted waves generated by small-scale underground geological bodies in the measured GPR profile is weaker than the linear reflected waves generated by the subsurface layered interface, and is easily mixed or masked by the reflected waves with stronger energy, which seriously interferes with the accurate identification and interpretation of small-scale targets. To this end, this paper proposes a multiple singular spectrum analysis method based on k-means clustering algorithm to separate the reflected and diffracted waves in the GPR profile. Then, the k-means clustering algorithm is used to cluster the all singular spectrum, and the singular values in the singular spectrum are divided into k classes according to the similarity, and the several types of singular values representing the reflected wave and the noise are set to zero, and some of the singular values representing the diffracted wave are retained to reconstruct the GPR diffraction wave field. Numerical experiments show that the multiple singular spectrum analysis method based on k-means clustering algorithm can efficiently separate the reflected wave from the diffracted wave. Compared with the reverse time migration of the unseparated GPR data, the imaging resolution of the small-scale geological body is higher and easier to be identified in the separated diffraction wave reverse time migration, which provides a feasible and effective method for the high-precision interpretation of small-scale targets.