Selected Sample Retraining Semi-supervised Learning Method for Aerial Scene Classification

Abstract

The performance of scene classification for remote sensing images based on deep neural networks is limited by the number of labeled data. To alleviate this problem, a variety of methods have been proposed to apply semi-supervised learning to exploit both labeled and unlabeled samples for training classifiers, but most of them still require a certain number of labeled samples considering the complex context relationship and huge spatial differences of remote sensing images. In this paper, we proposed a novel selected sample retraining semi-supervised learning method (S$^2$R) that is simple but works efficiently on scene classification remote sensing. First, we train several models independently, each model is trained for only a few epochs, and use them to label samples in the unlabeled data set. Then, the labeled unlabeled data set is divided into low-noise labeled data set and sub-unlabeled data set through the high probability sample selection method. Finally, the two segmented data sets are combined with the labeled data sets to train a scene classifier based on the semi-supervised learning method. To verify the effectiveness of the proposed method, it is further compared with several state-of-the-art semi-supervised classification approaches. The results demonstrate that our method consistently outperforms the previous methods on the condition of only a few labeled samples over the scene classification for remote sensing images.