The Application Of Deep Learning In Dimensionality Reduction And Classification Of Hyperspectral Image

With the development of remote sensing technology,the new remote sensing sensors can collect the image with continuous spectral domain and spatial domain.The traditional remote sensing image classification framework only uses the spectral information to classify,ignoring the influence of spatial information on classification.Convolutional neural networks have unique advantages.The image does not need too much early treatment can directly enter the network,from the training data of implicit learning,to avoid the process of data reconstruction in feature extraction and classification.Its unique inter layer connection and the close connection of spatial information make it suitable for classification and recognition in image processing.In this thesis,on the premise of taking full account of the characteristics of hyperspectral images,a convolution neural network which has achieved remarkable results in the field of image classification is proposed.The main content of this paper:(1)Using the design idea of Le Net-5 network framework,we design a convolution neural network framework for hyperspectral image classification.This thesis mainly studies the network frame number,the number of neurons in layer convolution,sampling layer number of neurons and the number of neurons in the output layer,achieves the effective transformation of hyperspectral data classification to image classification.(2)The spatial neighborhood information of each pixel is used as the input sample of the convolution neural network framework to explore the validity of the designed framework for hyperspectral image classification.(3)The design of the activation function ReLU in the framework of the study,to achieve the purpose of reducing the gradient dispersion,improve network efficiency and classification accuracy.Compared with the gradient descent method,the mini-batch stochastic gradient descent method can greatly improve the efficiency of the framework.The use of these strategies can help to extract classification features and improve classification results.(4)Using the framework designed in this thesis,the simulation experiment of The University of Pavia data set is carried out to verify the feasibility,and compare with thetraditional k nearest neighbor,BP neural network and SVM classification method.The simulation results show that the classification accuracy is higher than that of other classification methods,which is 97.57%.