Hyperspectral Image Classification Based On Recursive Network

In the last 20 years, hyperspectral remote sensing technology has become an advanced technology with rapidly advanced development, which can detect and classify objects on the earth and has been widely applied in many fields including the management of the environment, precision agriculture, social security, and military defense as well as mineralogy. Classification is a very important application for hyperspectral remote sensing technology, and feature learning is a very important method for improving classification performance. However, feature learning for hyperspectral images is faced with some problems that consist of the difficulty of obtaining label pixels and also the high dimension of pixels. Therefore, solving the problems becomes a topic of feature learning.Recently, deep learning, as an important method for feature learning, has been widely applied in many fields. However, in the field of hyperspectral image, deep learning has been rarely applied. In the thesis, we proposed a method based on recursive network to learn feature in an unsupervised manner, which can overcome the difficulty of obtaining the label for hyperspectral image. Finally the learned features can be used to for classification. Our main contributions are as follows.1. Deep learning methods can learn the intrinsic features for data and achieve excellent features. In the paper, we proposed a classification method using spatial-spectral information based on a recursive autoencoder network belonging to deep learning for hyperspectral image. For the hyperspectral image, each pixel can build a neighboring window of N ′N, the similarity weights are computed according to the similarity between the central pixel and its neighboring pixels. In the neighboring window, each pixel is given the corresponding similarity weight as the input of recursive autoencoder network, then, the new features can be learned by the network. Finally, the high-level features are used for classification.2. Since some pixels locate in the near boundary between land-cover classes, the class labels of these pixels may be different to the ones of their neighboring pixels, which can affect the classification accuracy. And the number of parameters of the recursive autoencoder network is large to cause high complexity. Therefore, we proposed a classification method based on recursive locality preserving projection(LPP) network for hyperspectral image. First, in the N ′N neighboring window, we selected 2K(K <N) pixels which are more similar with the central pixel, and other unselected neighboring pixels are abandoned. Then each neighboring window consists of K pixels. The neighboring window including K pixels can learn features through the recursive LPP network, and the learned features stand for the center pixel in the neighboring window. Finally, the learned features are used for classification.3. For hyperspectral image, the pixels are of high dimension and there exists redundancy between bands. We proposed band selection method based on the recursive LPP network for hyperspectral image. First, we use K-means algorithm on all bands to form P groups in which each group includes relevant bands. Then each group of bands is used to learn feature via the recursive LPP network. Finally, the most relevant band with the learned feature is selected in each group. Therefore, we can select P representative bands and band selection is accomplished. Experiments show that the band selection method based recursive LPP network for hyperspectral image is feasible and effective.