A novel graph matching based approach for land-cover classification of multi-temporal images

We address the problem of automatic land-cover map updating of multi-temporal and multi-spectral remotely sensed images in this paper. Given a pair of images acquired on the same geographical area at two distinct time instants, it is assumed here that the training data are available for one of the acquisitions, which is known as the source domain image. The task is to classify the other image (target domain image) for which no reliable reference map is available. It is further assumed that both the images share the same set of land-cover classes though the statistical properties of the classes may change considerably over time. Under these assumptions, a novel graph matching technique is proposed to approximate the class labels of the target domain data which is initially clustered. Given that the samples from different classes may overlap in the spectral domain, which a clustering algorithm fails to detect properly, a partially supervised Maximum Likelihood (ML) classifier coupled with the Expectation Maximization (EM) based parameter re-training scheme is used iteratively to further jointly update the class-conditional densities and the prior probabilities of the classes in the target domain image. Experimental results obtained on multi-spectral datasets confirm the the effectiveness of the proposed method.