Research On Hyperspectral Image Classification Based On Deep Learning

Hyperspectral remote sensing image is a kind of remote sensing image, which is mainly formed by dealing with the electromagnetic waves of different wavelengths, and it was composed of tens to hundreds of bands. It is rich in spectral information and is applied to the ground environment monitoring, crop detection, geological survey and military reconnaissance. Among them, the classification of hyperspectral images is the most critical and basic technology. However, with the increase of the amount of hyperspectral image data, the spectral information will produce a lot of redundancy, and the phenomenon of "isomorphism" and "foreign matter spectrum" also increased. This has put forward a hug challenge to hyperspectral image classification. Therefore, it is necessary to study the highly efficient hyperspectral image classification algorithm for the wider application of hyperspectral images.In this paper, we analyzed the spectral information and spatial structure information of hyperspectral images, and proposed the hyperspectral image classification algorithms based on PCA Network and convolution neural network. The main work includes:(1) Analyzing the spectral characteristics and the spatial structure of the hyperspectral image, which provided the theoretical basis for the hyperspectral image classification algorithm.(2) Based on the unique characteristics of hyperspectral remote sensing image, a feature extraction and classification algorithm based on PCA Network for hyperspectral remote sensing image is proposed. This method uses the constructed PCA Network and Gaussian SVM multi - classifier to extract spectral features and classify respectively. At the same time, this method combined with the band selection and threshold determination method to enhance the classification of robustness, greatly improving the classification accuracy.(3) Based on the analysis of structural characteristics of the convolution neural network, we realize classification of hyperspectral image, and analyze the feasibility of the deep learning network in the hyperspectral image classification.In conclusion, this thesis we proposed a classification algorithm for hyperspectral remote sensing images based on deep learning network (PCA Network and Convolution Neural Network). The experimental results show that the classification method can excavate the implicit deep spectral characteristics and improve the classification accuracy to a large extent with combining spatial structure information.