The deconvolution of blurred and noisy satellite images is an ill-posed inverse problem, which can be regularized under the Bayesian framework by introducing an appropriate image prior. In this paper, we derive a new image prior based on the state-of-the-art nonlocal means (NLM) denoising approach under Markov random field theory. Inheriting from the NLM, the prior exploits the intrinsic high redundancy of satellite images and is able to encode the image's nonsmooth information. Using this prior, we propose an inhomogeneous deconvolution technique for satellite images, termed nonlocal means-based deconvolution (NLM-D). Moreover, in order to make our NLM-D unsupervised, we apply the L-curve approach to estimate the optimal regularization parameter. Experimentally, NLM-D demonstrates its capacity to preserve the image's nonsmooth structures (such as edges and textures) and outperforms the existing total variation-based and wavelet-based deconvolution methods in terms of both visual quality and signal-to-noise ratio performance.