Active-learning based cascade classification of multitemporal images for updating land-cover maps

This paper presents a novel active-learning (AL) technique in the context of the cascade classification of multitemporal remote-sensing images for updating land-cover maps. The proposed AL technique is based on the selection of unlabeled samples that have maximum uncertainty on their labels assigned by cascade classification, and explicitly exploits temporal correlation between multitemporal images. Uncertainty of samples is assessed by conditional entropy that is defined on the basis of class-conditional independence assumption in time domain. The proposed conditional entropy based AL method for cascade classification technique is compared with a marginal entropy based AL technique adopted in the context of single-date image classification. Experimental results obtained on two multispectral and multitemporal data sets show the effectiveness of the proposed technique.