In remote sensing image classification, it is commonly assumed that the distribution of the classes is stable over the entire image. This way, training pixels labeled by photointerpretation are assumed to be representative of the whole image. However, differences in distribution of the classes throughout the image make this assumption weak and a model built on a single area may be suboptimal when applied to the rest of the image. In this paper, we investigate the use of active learning to correct the shifts that may appear when training and test data do not come from the same distribution. Experiments are carried out on a VHR remote sensing classification scenario showing that active learning can effectively learn the covariance shift and provide robust solutions.