In this paper, an approach to reducing the effects of registration noise in unsupervised change detection is proposed. The approach is formulated in the framework of the change vector analysis (CVA) technique. It is composed of two main phases. The first phase aims at estimating in an adaptive way (given the specific pair of images considered) the registration-noise distribution in the magnitude-direction domain of the difference vectors. The second phase exploits the estimated distribution to define an effective decision strategy to be applied to the difference image. Such a strategy allows one to perform change detection by significantly reducing the effects of registration noise. Experimental results obtained on simulated and real multitemporal datasets confirm the effectiveness of the proposed approach.