Hyperspectral Classification With Small Samples Based On Sparse Representation And Semi-supervised Active Learning

The number of spectral bands of hyperspectral data is dozen tens to hundreds, which provide rich and valuable information to discriminate various materials. However, classification of hyperspectral data is challenged by several problems, such as high dimension of data, limitation of labeled samples and high costs to label and identify the samples. In order to solve these problems, many scholars have been putting forward more and more new methods. Unsupervised classification methods do not need any labeled samples, but their effects of classification are not good enough. The traditional supervised methods relatively achieved satisfactory classification accuracy, but they require a large number of labeled samples. In addition, in order to avoid Hughes phenomenon, many approaches to tackle the problem with dimension reduction, which will result in the valuable information lost.Sparse decomposition algorithm can be implemented to classify high dimensional data without dimension reduction. This paper focused on the problem of hyperspectral data classification based sparse representation framework and the research fruits can be summarized as follows.1. In this paper, we attempt to find relationship between the classification accuracy and the number of labeled samples in the classification of hyperspectral data with small sample based on sparse representation. Although several researchers have made some achievements to improve the classification accuracy of hyperspectral data with small samples, so far, there are not any research reports on the area of the relationship between the number of labeled samples and classification accuracy. In this paper, we proved that classification accuracy of hyperspectral data is positively correlated with the number of labeled samples from both theoretic derivation and empirical experiments. The research provides a basis for the classification of hyperspectral data on theory and experiment.2. To solve this problem of hyperspectral data classification, a novel framework which is based on sparse representation and semi-supervised active learning is proposed in this paper. This framework constructs the base classifier with sparse representation algorithm and combines active learning and semi-supervised learning. On the one hand, semi-supervised learning increases the number of labeled samples by assigning pseudo-labels to unlabeled data. On the other hand, active learning labels little “key” samples by human beings to drop the error rate of pseudo-labels. To improve the accuracy of pseudo-labels, the framework removes the pseudo-labels with wrong labels by using pseudo-labels confirmation process. We proved the effectiveness of the proposed algorithm based on our framework by the empirical experiments.Meanwhile, the framework combines active learning and semi-supervised learning based on sparse representation algorithm is proposed to realize high classification accuracy of hyperspectral data with low proportion of labeled sample.