Denoising plays a fundamental role in ground penetrating radar (GPR) data processing and determines the effect of anomaly extraction, inversion imaging, and other subsequent processing. In recent years, the sparse dictionary representation method k-singular value decomposition (K-SVD) based on K-means, which can adaptively change the basis function according to the data, has become a hotspot in the field of image denoising and data reconstruction. Nevertheless, the SVD is a time-consumed calculation, especially unacceptable in multidimensional problems, we introduce a dictionary learning method based on the sequential generalized K-means (SGK), where the dictionary atoms are updated by the arithmetic average of several training signals instead of a great deal of SVD calculation in K-SVD. We establish a 3-D road simulation model and conducts finite difference time domain forward numerical simulation to acquire 3-D GPR data. Through three sets of experiments on 3-D numerical examples and 3-D field data, the resultsshow that both dictionary learning algorithms can successfully remove random noise from GPR data even at a lower input signal-to-noise ratio. The clutter interference in the random medium forward data can be effectively eliminated simultaneously, and both denoising methods exhibit promising applications in 3-D field data. However, the SGK method solves the serious problem of computational efficiency to a certain extent. The computational acceleration ratio of SGK remains consistently above 7.5 times that of the K-SVD algorithm in multi-group experiments, with only a marginal decline in denoising performance.