Active Learning Algorithms For The Classification Of Hyperspectral Sea Ice Images

Sea ice is one of the most critical marine disasters, especially in the high latitude regions. Compared with the conventional remote sensing techniques, the most obvious feature of the hyperspectral monitoring sea ice is high spectral resolution. In addition, hyperspectral image contains continuous spectrum information and abundant spatial information, so it can accurately distinguish between seawater and the different types of sea ice.The classification of hyperspectral sea ice image often applies the supervised classification techniques, and often the typical classification method is based on support vector machine classification. Because of the limitations of environment and conditions, the measured data used in the sea ice detection are very rare. However, marking samples requires a lot of time and cost. But if only a small amount of labeled samples are used to train classifier model, the classifier’s generalization performance is not good. This is a waste for unlabeled samples. Although unlabeled samples have no specific label, their distribution may reveal the internal structure of the processed data, which can improve the classifier’s generalization ability. Therefore, in the case of a small number of labeled samples, how to mark as few samples as possible artificially and obtain better classification performance become the key issues of sea ice detection.In this paper, the active learning method is improved and the semi-supervised learning is introduced, which is applied to the classification of hyperspectral image. The main work is as follows:1) We introduce the basic principle and the process of the classification of hyperspectral image. In addition, we focus on the support vector machine algorithm based on supervised classification. We elaborate the support vector machine algorithm followed from linear to nonlinear; from the second category to the multi-class classification. And finally we implement the SVM Multi-class Classification.2) We propose a novel investigated active learning algorithm based on the evaluation of two criteria: uncertainty and diversity. The algorithm was tested the classification performance by hyperspectral sea ice image. Compared to the different active learning methods and random sampling, the experimental results show that the BvSB + ECBD method can provide more informative samples and use far fewer training samples to achieve higher accuracies on the basis of SVM. Finally,we carry out an analysis of the sensibility of our proposed BvSB + ECBD method with different number of initial training samples. The experiment result indicates that the different initial training samples don’t provide a large benefit in the BvSB + ECBD method. Furthermore, we can also observe that, when using more initial training samples, convergence is easily achieved.3) We propose an active learning algorithm combined with semi-supervised learning(BvSB + ECBD + TSVM). When only a small number of labeled sample, the method can make full use of a lot of the information of unlabeled samples to establish sea ice classifier. In the semi-supervised learning, the process of the selected semi-labeled samples introduced the active learning thought. Finally, we combine active learning and semi-supervised learning based on support vector machine classifier, which can further reduce the cost. The algorithm was tested using hyperspectral sea ice image. Compared to the traditional passive techniques, we found that BvSB + ECBD + TSVM algorithm can use less labeled samples to achieve higher classification accuracy. In addition, our proposed method can effectively overcome the wrong assessment of the unlabeled samples and greatly improved the classifier’s generalization ability.