Hyperspectral Image Classification Based On Convolutional Neural Network

Hyperspectral images classification is an important research direction of remote sensing technology.Hyperspectral images obtain richer information by extracting images of objects in multiple spectra,which is helpful for objects classification and detection.Hyperspectral images have the characteristics of large number of channels,strong correlation between adjacent channels,and heavy information redundancy.Therefore,the process of feature extraction from hyperspectral images is more complicated than the process in RGB images.For now,the methods of feature extraction from hyperspectral images mainly from the following two aspects: Extracting local space information in the same channel,extracting the spectral information at the same spatial position in different channels.However,those feature extraction methods require lots of prior knowledge,it is difficult to make full use of the space and spectral information of hyperspectral images at the same time,and the methods have poor generalization ability.Convolution neural network has strong ability for feature extraction,and the features extracted by convolution neural network are translation,rotation,scale invariant.The convolution neural network can effectively extract the features characteristics of hyperspectral images.Convolution neural networks use random gradient descent for training,with no need for too much priori knowledge.Give the inputs and labels,the network can do the feature extraction and classification automatically.In this paper,we study the structural of the hyperspectral image,get 3-D image blocks from hyperspectral image,and the blocks are then input into convolution neural network.After several convolution and pooling layers,the input block are classified with predicted labels.Our research work includes these three points:Firstly,a convolutional neural network for hyperspectral image classification is designedaccording to the structure of hyperspectral images.The hyperspectral image is a multi-channel image,and each channel represents images of objects under a particular spectrum.The number of hyperspectral image channels is large,in agreement with the nature of the convolutional neural network that the CNN requires a large amount of training data.Therefore,the convolution neural network is suitable for dealing with hyperspectral image.In this paper,we design a convolutional neural network with multi-scale feature extraction layer,feature fusion layer and feature dimension reduction layer.The network extracts the hyperspectral image space and spectral features at the same time,and obtains excellent classification performance.This paper analyzes the influence of network structure on classification performance,does some experiments to show the relationship among the network depth,classification accuracy and the time cost for running.Then,the convolution network is combined with the traditional support vector machine to solve the problem that the fully-connected layer of the convolution neural network lacks nonlinear structure.The convolution,pooling and relu layers of the convolution neural network have strong feature extraction ability,but the fully-connected layer has only a simple linear structure,which may decrease the classification accuracy.The support vector machine is a classifier which maximizes the classification interval.When dealing with linear indivisible problems,the feature is mapped into high dimensional space,so that the linear indivisible problem can be transformed into a linearly separable problem.To make full use of the advantages of the feature extraction ability in convolution network and the ability of dealing with nonlinear problems in support vector machine,this paper designs the algorithm combine the convolution network with the support vector machine.The algorithm firstly trains the convolution neural network.After the network training is completed,the training samples are input intto the convolution network to get the output of the fully-connected layer,and then the support vector machine is trained using the output of the fully-connected layer and the given labels.The combination of convolution network and support vector machine further enhances the classification accuracy of hyperspectral images.Finally,an adaptive spatial window method according to image content is designed.When using a convolutional neural network to classify a pixel of a hyperspectral image,it is necessary to extract the spatial window centered on the pixel,and then send the window into the network to get the predicted label.The size of window should be based on the content of the image,that is,we should use large windows in the area where the image content is simple to obtain a higher signal to noise ratio,and use small windows in areas where the image content is complex and the category are mixed together.In this paper,an algorithm of adaptive window is designed,we use convolution network to get the classification confidence.Some rules are set to select the spatial window with higher classification confidence adaptively,and the final classification result is obtained by putting the selected window into convolution neural network.