2810 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 50, NO. 7, JULY 2012 A Novel Domain Adaptation Bayesian Classifier for Updating Land-Cover Maps With Class Differences in Source and Target Domains Kanchan Bahirat, Francesca Bovolo, Member, IEEE, Lorenzo Bruzzone, Fellow, IEEE, and Subhasis Chaudhuri, Fellow, IEEE Abstract—This paper addresses the problem of land-cover map updating by classification of multitemporal remote-sensing images in the context of domain adaptation (DA). The basic assumptions behind the proposed approach are twofold. The first one is that training data (ground reference information) are available for one of the considered multitemporal acquisitions (source domain) whereas they are not for the other (target domain). The second one is that multitemporal acquisitions (i.e., target and source domains) may be characterized by different sets of classes. Unlike other approaches available in the literature, the proposed DA Bayesian classifier based on maximum a posteriori decision rule (DA-MAP) automatically identifies whether there exist differences between the set of classes in the target and source domains and properly handles these differences in the updating process. The proposed method was tested in different scenarios of increasing complexity related to multitemporal image classification. Exper- imental results on medium-resolution and very high resolution multitemporal remote-sensing data sets confirm the effectiveness and the reliability of the proposed DA-MAP classifier. Index Terms—Bayesian classifier, domain adaptation (DA), land-cover map updating, maximum a posteriori (MAP) classifier, multitemporal image classification, partially supervised learning, partially unsupervised learning, remote sensing. I. I NTRODUCTION T HE OBJECTIVE of domain adaptation (DA) tech- niques (also known as transfer learning or partially supervised/unsupervised learning) is to take advantage of the available knowledge on a given source domain in order to infer a model/classifier suitable for the classification of a related (yet not identical) target domain for which a priori information is not available [1], [2]. These kinds of techniques have proven to be effective in different applications mainly related to text analysis and natural language processing [1]–[4]. Moreover, few successful examples can be found also in remote sensing where DA techniques become useful when there is a need of classifying images acquired on the following: 1) spatially Manuscript received July 14, 2010; revised March 31, 2011 and August 10, 2011; accepted October 16, 2011. Date of publication December 13, 2011; date of current version June 20, 2012. K. Bahirat and S. Chaudhuri are with the Department of Electrical Engineer- ing, Indian Institute of Technology Bombay, Mumbai 400076, India (e-mail: bkanchan@ee.iitb.ac.in; sc@ee.iitb.ac.in). F. Bovolo and L. Bruzzone are with the Department of Information Engineer- ing and Computer Science, University of Trento, 38123 Trento, Italy (e-mail: francesca.bovolo@disi.unitn.it; lorenzo.bruzzone@ing.unitn.it). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2011.2174154 disjoint areas that show similar characteristics or 2) the same geographical area at different times. In this paper, the attention is focused on the analysis of multitemporal remote-sensing images acquired on the same area at different times and the related land-cover map updating [5]–[8]. Due to the periodic and regular acquisition of remote-sensing images on the same geographical area and to the difficulties in collecting the ref- erence data on the ground with the same frequency, it is not possible to apply standard supervised classification algorithms to each available image. In this context, DA techniques have a high importance for the development of monitoring systems aimed at regularly mapping geographical areas of interest. In this specific scenario, the source domain is identified as the first image to be classified for which a training set is assumed to be available, whereas the target domain is associated to each new available image to be classified (acquired on the same geographical area of the source domain) for which a training set is likely to be not available [7]–[12]. In this context, DA should be faced according to semisupervised strategies that take advantage of the training set available for the source domain and the unlabeled samples from the target domain in order to derive a classification rule suitable for the target domain. The few DA methods proposed in the remote-sensing lit- erature are based on the assumption that the set of land- cover classes that models the target domain should be the same as those included in the source domain. In other words, the differences between the two domains are only related to differences in the statistical parameters of land-cover classes due to differences in the acquisition conditions (e.g., differences in the atmospheric conditions at the image acquisition dates, sensor nonlinearities, and different levels of soil moisture). Under this assumption, in [9], a partially unsupervised approach is proposed, which can update the parameters of an already trained parametric maximum-likelihood (ML) classifier on the basis of the distribution of a new image for which no ground reference information is available. In [10], in order to take into account the temporal correlation between images acquired over the same area at different times, the partially unsupervised ML classification approach is reformulated in the framework of the Bayesian rule for cascade classification. The basic idea in both approaches consists in modeling the observed spaces by a mixture of distributions, whose components are estimated through the employment of unlabeled data according to a proper inference applied to training samples of the reference image. In [11] and [12], partially unsupervised classification approaches 0196-2892/$26.00 © 2011 IEEE