Contextual-based image classification attempts at considering spatial/temporal information during the learning process in order to make the classification process smarter. Sequential learning techniques are one of the most used ones to perform contextual classification, being based on a two-step classification process, in which the traditional noncontextual learning process is followed by one more step of classification based on an extended feature vector. In this paper, we propose two ensemble-based approaches to make sequential learning techniques less prone to errors, since their effectiveness is strongly dependent on the feature extension process, which ends up adding the wrong predicted label of the neighborhood samples as new features. The proposed approaches are validated in the context of land-cover classification, being their results considerably better than some state-of-the-art techniques in the literature.