Hyperspectral Image Classification Based On Deep Learning

Hyperspectral data contain rich spatial and spectral information,Therefore,it has been widely used in agriculture,environmental science,ground observation and so on.In these applications,the classification of hyperspectral image plays a crucial role.The traditional machine learning classification method only extracts spectral features or shallow spatial features,thus losing some valuable deep feature information,Resulting in unsatisfactory classification results.With the rapid development of computer technology in recent years,deep learning has played a positive role in the classification of hyperspectral images.Deep learning can obtain more abstract and deep feature information,but how to effectively utilize the spatial features of hyperspectral images has become a hot topic in this field.In this thesis,the method based on multi-channel features is used to classify hyperspectral images.The main contents of this thesis are as follows:First:we use the principal component analysis to reduce the dimensionality of the hyperspectral data,and use the multi-channel automatic encoder to extract the features of the hyperspectral image.That is,a spectral feature and two spatial features of different neighborhood sizes are extracted respectively,and then the extracted features are combined as an input of the logistic regression classifier to obtain the final classification result.Finally,the classification results are compared with single-channel and dual-channel automatic encoders to prove the effectiveness of multi-channel features in hyperspectral image classification.Second:in order to extract more abundant and abstract spatial spectral information,we propose a multi-channel convolutional neural network classification framework.The one-dimensional convolutional neural network,two-dimensional convolutional neural network and three-dimensional convolutional neural network are used to simultaneously extract the features of the hyperspectral image,and then the extracted feature information is combined.Finally,the final classification result is obtained by using the logistic regression classifier.Through comparative experiments,it can be found that the multi-channel convolutional neural network has a strong learning ability and can better utilize the spatial feature information.Third:because hyperspectral images have more nonlinear structures and more dimensions.In order to eliminate the redundancy of spectral information caused by the large number of dimensions,we first use a long and short memory neural network to preprocess the hyperspectral image.The pre-processed images are classified by multi-channel convolution neural network.The effectiveness of this pre-processing method can be found through comparative experiments.Fourth:we also use the Generative Adversarial Networks to classify hyperspectral images,both the generation model and the discriminant model are in the form of convolutional neural networks.The advantages and disadvantages of this method can be found through experiments.At the same time,at the end of this thesis,the future research directions of hyperspectral image classification are also determined by summing up.