Research On Deep Learning Based Hyperspectral Image Feature Learning

Feature extraction usually refers to the process of extracting useful information from raw data and plays an important role in hyperspectral image processing because it can improve the value of subsequent data processing.In recent years,deep learning has made remarkable achievements in the fields of image classification,speech recognition and natural language processing."Feature learning" can start from the original image of the image and automatically discover the hidden patterns in the image through a specific neural network model to learn effective features,which has achieved unprecedented development in recent years.Inspired by the results obtained by deep learning in image feature learning,this thesis is based on the deep learning theory to study the feature learning of hyperspectral imagery under the support of the National Natural Science Foundation of China,in order to remove redundant information in the image and improve its performance in applications such as object classification and target detection.The main work and innovations of the thesis mainly include the following two aspects.Firstly,the convolutional neural network(CNN)structure currently designed for feature extraction inevitably use the samples from the target hyperspectral scene to train the convolutional neural network parameters.To overcome this limitation,this thesis proposes a sensor-specific spatial-spectral feature learning based on convolutional neural network.The concept of feature extraction,transference and fine-tuning to different scenes acquired by the same sensor are described respectively;on this basis,spatial context and spectral features are integrated to construct a five-layer convolutional neural network for hyperspectral images(C-CNN).Experimental results from several benchmark hyperspectral datasets for feature learning and hyperspectral image/sensor classification show that the proposed C-CNN is superior to the most advanced CNN-based classification method,and its corresponding feature learning CNN(FL-CNN)can be very effectively extract sensor-specific spatial spectral features for hyperspectral applications in both supervised and unsupervised modes.Secondly,for convolutional neural networks,a large number of labeled samples are usually needed to learn the effective features under the classification task,and a large number of labeled samples are difficult to obtain in hyperspectral remote sensing images.This thesis develops a three-dimensional convolution autoencoder(3D-CAE).An unsupervised spatial-spectral feature learning strategy is proposed for hyperspectral images.This strategy is designed to reconstruct the input image and train all the parameters involved in the network without the need for labeled training samples.In the proposed 3D-CAE,the self-encoder(AE)is constructed by using only 3D element operations to store structural information in hyperspectral images,such as three-dimensional convolution,3D pooling,3D batch normalization,etc.The flat operation(e.g.,the fully connected layer)is excluded from the network,which makes the proposed 3D-CAE more efficient for learning the spatial-spectral features.The experiments are systematically organized,and the experimental results of several benchmark hyperspectral datasets show that the proposed 3D-CAE is very effective in extracting the spatial-spectral joint features,and it is superior not only to the traditional unsupervised feature extraction algorithm in classification applications,but also to many supervised special extraction algorithms.In this thesis,the application of convolutional neural networks in hyperspectral image feature learning is explored.The proposed feature learning strategy can not only improve its performance in feature classification,target detection,etc.,but also the strategy is not limited to labeled samples,which is important for hyperspectral images.