In order to improve the spatial resolution of remote sensing images, the nonparametric Bayesian dictionary learning model for natural images super-resolution reconstruction is introduced into the field of remote sensing image processing. Based on nonparametric Bayesian and classified texture patches, an improved method of the single remote sensing image super-resolution reconstruction is proposed. The method uses the Beta-Bernoulli process for dictionary learning, and establishes the probability distribution models of dictionary elements and parameters. The Gibbs sampling is used to calculate the posterior distribution. Finally, the image block is divided into two types: smooth block and non-smooth block during reconstruction. The non-smooth block reconstructs the high resolution remote sensing image by using the posterior distribution of the high-resolution dictionary and the sparse coefficients of the low-resolution image blocks. While the smooth block only uses the bicubic convolution method to reconstruct. Furthermore, different from the shortage of traditional algorithm that needs to set a large dimension dictionary in advance to ensure a higher reconstruction precision, a smaller dimension dictionary is obtained by non-parametrical deviation of dictionary dimension in this paper, which reduces the calculation. The results show that the proposed algorithm outperforms traditional approaches both in visual and quantitative evaluation indexes whether the test image is noisy, and the reconstruction speed is faster.