Due to complex subsurface situation, echo signals surveyed with Ground Penetration Radar (GPR) often contain a lot of clutters, including direct-coupling wave, random noises and multiples. Existence of these clutters submerges measured feature signals of rock structures with GPR, so suppression of them is often essential conduct for rock feature extraction. For extracting rockmass structure features from surveyed GPR data signals, sparse representation (SR) of the signals is an invaluable scheme with a small number of elementary signals from over-complete dictionary.