Adaptive Convolutional Neural Network Based Hyperspectral Image Classification

Hyperspectral images(HSIs)contain fine-grained spectral information and rich spatial information,which enable objects that cannot be identified in remote sensing with wide bands to be identified in hyperspectral images.Therefore,hyperspectral images are widely used in mineral exploration,agricultural planning,address exploration,disaster warning,etc.However,hyperspectral has some characteristics,such as high dimension and close correction of hyperspectral bands,plenty of hyperspectral data and small amount of manually tagged data,which enhances the difficulty of hperspectral image classification and recognition.At present,convolutional neural networks(CNN),as the representative method of deep learning,are widely used in HSI classification.Nevertheless,the hyper-parameters(the size and number of convolutional kernels)and structure of CNN are often designed mannually,which are not consistent with the complexity of the problem and influence the accuracy of HSI classification.In addition,the training process of network needs a lot of time,because the network contains many parameters(especially,when the network model has deep structure).Aiming at solving above problems,this thesis focuses on the adaptive determination of hyper-parameters and structures of CNN,which can improve the accuracy and timeliness of HSI classification tasks.The principal contents and innovation points of this thesis are as follows:(1)In the existing CNN,the number of convolutional kernels is often determined empirically,which is difficult to match with the tasks and affects the classification accuracy in application.To solve this problem,this thesis proposes a HSI classification method which adaptively determine the number for CNN.Firstly,an adaptive method for determining the number of convolutional kernels based on clustering is proposed,which can adaptively determine the number of convolutional kernels from the selected image blocks;Aiming at the inaccurate of determining the convolutional kernel through the method,a convolutional kernel generation algorithm based on the joint measurement of image block similarity and density is proposed to improve the accuracy of generating convolutional kernel.Experimental results on the Indian Pines and Pavia University datasets show that the proposed method has higher classification accuracy than the CNN which empirically determines the number of convolutional kernels.(2)The convolutional kernel size of the existing CNN is often determined manually,which is difficult to adapt to the data complexity and reduces the classification accuracy.To solve this problem,this thesis proposes a hyperspectral image classification method based on CNN,which adaptively determines the convolutional kernel size.Firstly,an adaptive determination method for convolutional kernel size based on double weighted distance measure is proposed,which can automatically determine convolutional kernel size.Through the trained CNN,hyperspectral images are classified with improved classification accuracy.Experiments on different public datasets of hyperspectral image show that the proposed method is superior to the kernel size empirically chosen CNN.(3)The number of convolutional layers set by the existing experience does not match the complexity of the problem,which leads to the decrease of classification performance.To solve this problem,we propose a hyperspectral image classification method through CNN with the adptive network structure.Based on the genetic algorithm,the method optimizes the number of convolutional layers and the connection mode of CNN,which forms a network structure consistenting with the complexity of the problem.Thus,it improves the classification accuracy of the network.The classification results on Indian pines and Pavia university datasets show that compared with the CNN method with the empirically set number of convolutional layers(connection mode),the proposed method has achieved higher classification accuracy on hyperspectral images.(4)In the process of CNN training,the gradient disappears easily,which affects the convergence speed of the network.To solve this problem,this thesis proposes a hyperspectral image classification method based on convolutional neural network with neuron activation constraints,which accelerates the convergence of the network.We propose a method of input constraints of network node functions,which can better solve the problem of "gradient disappearance" in error back propagation better and improves the convergence speed of the network.Experimental results on different hyperspectral image public datasets show that the proposed method has faster training speed of convolutional neural network.Experimental results on different non-hyperspectral classification datasets verify the effectiveness of the proposed method in different network models.