Hyperspectral Image Classification Based On Convolutional Neural Networks

Since hyperspectral remote sensing technique has risen steadily, remote sensing data dimension has increased year by year, so that there is a challenge to the classification problem, the mainstream task of remote sensing data analysis. In the face of such characteristics as multiple dimensions, correlation, nonlinearity, large amount of data, how to apply the successful algorithm in the issue of hyperspectral data classification becomes an important problem in hyperspectral remote sensing data analysis. In addition, with the promotion of space testing technology, the spatial resolution of hyperspectral images getting higher and higher, leading to a high correlation between pixels, which all makes classifying hyperspectral data with spatial information possible. This thesis intends to use convolution neural networks, which is successful in image recognition and language testing, in hyperspectral classification with the advantage of its simple construction, able to deal with nonlinear, parallel computation.This thesis studies the convolution neural networks(CNN) of deep learning theoretical framework, and discusses the depth model structure of CNN, as well as researches the supporting platform of it, then verifies the platform by nature image sample set, and analyzes the possibility of the method that uses this model to classify hyperspectral image.Based on CNN study and spectral information classification basis, this thesis proposes two methods of transforming the spectral information to the image: one transforming spectral information to grey-scale map, which classifies data using the texture information between multispectral learned by CNN model; and another one transforming spectral information to oscillogram, which classifies data using the fluctuation information between multispectral learned by CNN model. Experiments show that CNN classification model can provide better performance in some data set. It demonstrates that classifying hyperspectral image with CNN is feasible.Besides, in order to make full use of the spatial information, based on the grey-scale map and the oscillogram with CNN, this thesis proposes two methods to classify hyperspectral data by using spatial-spectral information. Experiments show that the classification results provided by CNN model have a noticeable improvement after making use of spatial information. As well as, experiments demonstrates the spatial information has significantly meaning for hyperspectral data classification, so that it can improve the utilization rate of the rich information of remote sensing.