Change detection can be regarded as a twoclassification problem，which can be realized by classifiers. However, the support vector machine, as a classification method, may lead to some noise for lacking spatial constrains between samples. In order to solve this problem, this paper combines the transductive support machine (TSVM) and the markov random field (MRF) in change detection. Specifically, the TSVM is used to train and classify the samples initially then the MRF regularization is applied to refine the posterior probability by employing the spatial context information. This proposed approach is shown better performance when compared to other classification methods.