Labeled Samples Expansion-based Band Selection For Hyperspectral Image

In recent years, hyperspectral remote sensing technology is developing rapidly and is widely applied in the many fields. We can realize accurately recognition of targets by using the rich information of hyperspectral image. But the high-dimensional and highly redundant information are huge challenges to the subsequent data processing. Therefore, how to retain the rich information of hyperspectral data with reduced dimensionality is one of the important technical issues in hyperspectral image processing.Band selection is one of the most used methods to reduce the dimensionality of hyperspectral image. Band selection selects an optimal band subset from original bands, which can retain the physical meaning and spectral characteristics when compared with feature extraction. Band selection is often divided into unsupervised based algorithms, semi-supervised based algorithms, and supervised based algorithms depending on the number of labeled samples. Though supervised band selection can improve the performance of band selection using the label information, the labeled samples are difficult to obtain which limits the applications of supervised band selection.For the problem that the applications of supervised band selection are limited by the number of labeled samples, we propose two methods to extend the applications of supervised band selection. The main contents of this thesis are as follows:1. Pixels clustering-based band selection. Firstly, we segment hyperspectral image into small blocks and pick one representative pixel from every block. Then these representative pixels are clustered and we use support vector machine to optimize the clustering result further. Finally, we take the cluster categories as pseudo labels, i.e., we get the pseudo labeled samples. Pseudo labeled samples are used to replace the real labeled samples, and supervised band selection is used to select bands. This method converts unlabeled samples to pseudo labeled samples, so it can extend the range of applications of supervised band selection.2. Small sample expansion-based band selection. In order to solve the problem that labeled samples is scarce, this method improves band selection performance by increasing the number of labeled samples. In this method, we use two ways to increase labeled samples. One is based on the property that adjacent pixels often share the same labels, we can add these labeled samples’ neighbors into the labeled samples set. The other way is linear composition, we use the labeled samples which have the same class to synthesize some new samples. By adding these new labeled samples, the amount of labeled samples is increased, then better band selection effect can be reached.The above two methods proposed in this thesis both can expand the labeled samples set. By using expanded labeled samples set, supervised band selection can have a wider range of applications and achieve better performance. We conduct some simulations with real hyperspectral image to demonstrate the effectiveness of the two proposed methods.