ABSTRACT
Land cover maps are usually obtained from remote sensing (RS) images by using automatic supervised classi†cation techniques, which require a set of labeled samples for training the classi†cation algorithm. However, the accuracy of the thematic maps that can be obtained with these techniques strongly depends on the quality and quantity of the available training samples, whose collection is costly and time consuming. Accordingly, in real classi†cation problems, the available training samples are often not enough for an adequate learning of the classi†er. A possible approach to address this problem is to exploit unlabeled samples in the learning of the classi†cation algorithm according to a semisupervised classi†cation procedure. The semisupervised approach has been widely investigated
15.1 Introduction ..........................................................................................................................303 15.2 Background on Active Learning...........................................................................................305
