Inspired by the idea of co-training algorithm, in this paper we propose a novel semi-supervised learning algorithm, co-Gaussian Process (co-GP), under a Bayesian framework. Image data are characterized in two distinct views, i.e. two disjoint feature sets. A latent function with a GP prior is employed for each view. In learning process of co-GP, knowledge acquired in each view is transferred by probabilistic labels to the other in turns to enhance learning effect. In this manner, proper parameters are estimated in a bootstrap mode and a satisfying performance can be maintained with only small amount of labeled data. The experiments carried out on multitemporal images validate the proposed algorithm.