Hyperspectral Image Classification Based On Deep Convolutional Neural Networks

Hyperspectral image has be used for agriculture,forestry,and mining et al because of its abundant spectral information,by the developing of hyperspectral imaging and increasing demand,as one of the most important research fields in hyperspectral research,hyperspectral image classification has getting more and more attention.However,when increasing spectral resolution bringing more abundant information,challenges have also been brought,the methods that used in nature image,panchromatic image and multispectral image classification no longer suiting for classification of hyperspectral image,moreover,little samples,high dimensionality and correlation and nonlinear problems has been more and more serious.So how to extract more abundant features from this little samples,and how to aggregate this features to getting more discriminatory features has been problems to be solved.To solve this two problems,this thesis has tried to combine deep convolutional neural network,which has getting a great success in computer vision and other fields,with nonsubsampled contourlet transform,along with spectral information and local spatial information of pixels in hyperspectral image.More specifically:1.The thesis introduced characters in hyperspectral image and problems by this characters bringing,then induced existing methods in classification and pointed out their drawbacks.Nonsubsampled contourlet transform has many advantages like multi-scale,multi-direction and anisotropic,so this thesis tried used this transform for features extraction,first of all,this method used principal component analysis to reducing the dimensionality of hyperspectral image to be classified,then employed nonsubsampled contourlet transform for each dimension and used a sliding window to getting patches,along with discarding background pixels,then used deep convolutional neural network to aggregating these features and extracting more discriminatory features,then fed these discriminatory features to Softmax classifier for classifying.This thesis discussed different dimension reduce methods,different ratio to be reduced,size of patches,hyperparameters et al for the classification results.2.Every pixel in hyperspectral image has very abundant information,this information is redundancy and nonlinear,so this thesis combined nonsubsampled contourlet transform features with nonlinear spectral features,combined nonsubsampled contourlet transform features with spatial information of every pixel,and combined nonsubsampled contourlet transform features with spectral features and spatial information for hyperspectral image classification.The spectral features was got by using kernel local fisher discriminate analysis,the spatial information was got by sliding windows.Kernel local fisher discriminate analysis can decorrelation and minimize intra-class distance and maximize inter-class distance,and can extract local nonlinear features,so it is good choice for nonlinear features extraction.The combined features were then fed into deep convolutional neural network to aggregating these features and extracting more discriminatory features,then fed these discriminatory features to Softmax classifier for classifying.This fusion is beneficial to hyperspectral image classification by experiments.