Hyperspectral Data Classification Based On Deep Belief Network And Ensemble Learning

Hyperspectral imaging has been a novel technology in the remote sense field, which expands people’s ability to obtain the information of earth surface greatly. Especially, the classification for surface objects receives a great deal of attention for hyperspectral data processing. However, the characteristics of hyperspectral data(HSD), such as the combination of spatial and spectral information and high dimensional structure, bring lots of classification problems. In the paper, deep learning and ensemble learning is used to the classification of HSD, aimed to solve the above problems. In the paper, we extract feature in view of the characteristics of HSD with spectral and spatial changes, then realize feature learning and classification based on Deep Belief Network(DBN). After the analyze of the ensemble learning method for data classification, a complete method for the classification of HSD is got by combining the two, and effectively improves the classification results. The research is carried out as the following:Firstly, beginning with the principles and features of hyperspectral imaging, the paper studies the characteristics of hyperspectral remote sensing with spectral and spatial changes. In terms of qualitative and quantitative measures, the paper illustrates that HSD shows diversity in the view of spectrum and space. As a result, HSD has varied characteristics and the extraction of changeless characteristics are important for the classification of HSD.Secondly, the theory of deep learning is introduced, including Restricted Boltzmann Machine(RBM), DBN and corresponding deep classification framework. In detail, the introduction includes the framework and learning process of RBM network, stacked framework of DBN, and establishment and training rules of classification structure. The study has provided a theoretical basis for the followings.Thirdly, DBN has been applied to the classification of HSD. RBM is used to extracte feature of HSD to validate that deep learning is suitable for HSD and is able to extract the changeless characteristics. Based on this, corresponding to the problems caused by the combination of spatial and spectral information and high dimensionality, the paper proposes classification of HSD based on DBN, with three different aspects: spectral, spatial and spectral-spatial information. Experiments show that the proposed method achieves better performance than support vector machine which is a classic classifier and is an effective classification method.Finally, classification of HSD based on DBN and ensemble learning is introduced. Aimed to solve the restriction of classification accuracy caused by small sample size, the method based on DBN and ensemble learning is proposed. The method combines DBN which has excellent classification performance with ensemble learning. Frameworks based on random spectral features selection and random spectral-spatial features selection are founded respectively. Experiments validate that the ensemble methods achieve higher overall accuracy than single DBN classifier.