A genetic algorithm based cost-sensitive active learning technique for classification of remote sensing images

This paper proposes a novel cost-sensitive active learning (CSAL) technique for the classification of remote sensing images with Support Vector Machines. The proposed technique assumes that the labeling cost of samples during ground survey depends on both the samples accessibility and the traveling time to the considered locations. Thus, it is not equal for the samples on the ground. Accordingly, the proposed method aims at selecting the most informative (the most uncertain and diverse) as well as cost-efficient samples at each iteration of the active learning process. This is accomplished according to three steps. In the first step the most uncertain unlabeled samples are selected by using the multiclass-level uncertainty technique. In the second step, the small (and important) portion of the image, in which the highest density of the most informative samples exists, is selected to effectively limit the study area. The objective of restricting the study area to a small portion of the image is to reduce the traveling time for labeling the samples. This is achieved on the basis of a novel clustering based approach. In the third step, uncertain samples that are diverse and cost-efficient are selected from the small portion of the image chosen at the second step. The selection of cost-efficient diverse samples is achieved on the basis of a genetic algorithm. Thanks to the second and third steps of the proposed CSAL technique, the cost of sample labeling is significantly reduced, while obtaining accurate classification maps. Experimental results show the success of the proposed CSAL method compared to the most promising literature active learning methods.