A NOVEL APPROACH TO TARGETED LAND-COVER CLASSIFICATION OF REMOTE-SENSING IMAGES

In several real-world applications the objective of land-cover classification is actually limited to map one or few specific "targeted" land-cover classes over a certain area. In such cases, ground truth is generally available for the only land-cover classes of interest, which limits (or hinders) the possibility of successfully employing standard supervised approaches that require an exhaustive ground truth for all the land-cover classes characterizing the investigated area. In this paper, we present a novel technique capable of addressing this challenging issue by exploiting the only ground truth available for the only land-cover classes of interest. In particular, the proposed method exploits the expectation-maximization (EM) algorithm and an iterative labeling strategy based on Markov random fields (MRF) accounting for spatial correlation. Experimental results confirmed the effectiveness and the reliability of the proposed technique.