Multi-sensor image classification based on active learning

The insufficient number of training samples may often cause relatively low and unsteady accuracies in multi-sensor image classification. It is also difficult to properly deal with multi-source data simply by traditional classifiers. In this paper, we propose a novel active learning classification system to solve these problems. Firstly, the adaptive query by committee (AQBC) strategy could reduce the need of known labeled samples and meanwhile provide more accurate predictions of the actively selected unlabeled samples to further decrease misclassifying rates. In addition, the involved basic classifier based on the optimized Meta-Gaussian distribution could better fuse different types of feature sources. Finally, compared with other traditional methods, the experiment results show that the proposed classification system could improve classification accuracies effectively and make full use of the limited training samples in multi-source data sets.