Application Research Of Deep Learning In Hyperspectral Image Classification

Hyperspectral remote sensing technology has received more and more attention in recent years.This is because hyperspectral image contains spectral information,spatial information and radiant energy of a large number of features,so the identification of ground objects can be realized and classified through the rational use of such information.However,with the increasing dimension of remote sensing data,the problems of large amount of image data,data redundancy and spectral correlation need to be solved.How to extract the deep features of data often determines the quality of classification results.For the feature extraction of hyperspectral images,scholars have proposed a number of targeted algorithms,such as linear dimension reduction based on principal component analysis,nonlinear dimensionality reduction based on popular learning,and support vector machines.However,the features extracted by these methods are often limited to a relatively shallow feature level,so the classification results are difficult to improve.With the development of deep learning technology,hyperspectral image feature extraction based on deep neural network is a research hotspot in the field of machine learning in recent years.In general,network parameters are trained to be optimal by constructing deep neural networks suitable for hyperspectral images and fine-tuning them.Deep neural networks can often extract features that are more abstract and easier to classify,so that high-accuracy classification targets can be achieved.Combining the spectral spatial information of hyperspectral imagery,this thesis studies the theoretical knowledge of two deep learning methods,convolutional neural network and residual dense network,and its application in hyperspectral imagery.The main research contents are as follows:(1)Feature extraction and classification based on spatial edge joint information of hyperspectral image and deep convolutional neural network.In this method,the data amplification and denoising processing of the hyperspectral image are first performed,then the spatial information is extracted by the minimum noise processing method,and the edge information is extracted by the bilateral filter,and then the two kinds of information are combined to generate a new,preprocessed hyperspectral image.Through the training test of the preprocessed hyperspectral image,some specific improvements are made in the traditional convolutional neural network model,including changing the size of the sampling window,adding the Dropout layer,and changing the number of layers by fine-tuning.Convolution kernel size and other parameters,and finally use the convolutional neural network model to achieve spatial and edge feature extraction of hyperspectral images.Experimental results on two public datasets,Salinas and Indian Pines,show that the classification ability of the model is better than some traditional algorithms and other models,thus verifying the effectiveness of the method.(2)A hyperspectral image classification model based on residual dense network is proposed.According to the characteristics of the residual network and the densely connected network,the residual network and the densely connected network are combined to form a residual-dense network,and on this basis,a three-dimensional residual dense network is derived.The network can reuse the characteristics of each layer and train deep networks.In the experiment,an initial three-dimensional residual dense network model is established based on the high dimensionality and large amount of data of the hyperspectral image.A hyperspectral image cut into a block form is then input thereto.Through continuous optimization of the model through experiments,the optimal hyperspectral image classification model is finally determined,so as to effectively extract the spectral and spatial features of hyperspectral images.The results of the final experiment on two data sets,Indian Pines and PaviaU,demonstrate that this method is indeed effective in improving the classification accuracy of hyperspectral images.In summary,the deep learning-based method proposed in this thesis can deeply mine the spectral and spatial feature information of hyperspectral images.Therefore,compared with the traditional classification algorithms based on shallow neural networks and manual computational features,the classification performance has been greatly improved.