Research On Ground Feature Classification Of Hyperspectral Remote Sensing Image Based On Deep Learning

With the development of imaging spectrometer technology,the acquisition of hyperspectral remote sensing image becomes easier and easier.The classification of hyperspectral remote sensing image is a research hotspot of hyperspectral remote sensing technology.Compared with multispectral image,hyperspectral image has abundant spectral information,which provides great help for classification task.In recent years,deep learning has made great achievements in image processing and has been applied in hyperspectral remote sensing image classification.In this paper,aiming at the problems existing in hyperspectral remote sensing image classification,the application of deep learning in hyperspectral remote sensing image classification is studied in depth.The main work is as follows:Firstly,in view of the large number of bands,strong correlation between bands and redundancy of hyperspectral remote sensing images,the data preprocessing of the hyperspectral remote sensing image data sets Indian Pines and Pavia University is carried out.In the process of preprocessing,the principal component analysis method is used to reduce the remote sensing data to B dimension,which can retain the components that have a great impact on classification,eliminate redundant information,simplify the calculation process,and save the calculation resources.Secondly,two 3D convolution neural networks Hierarchy1 and Hierarchy2 are constructed by adding convolution layer and pooling layer to deepen the depth of the network.These two networks combine the spatial and spectral features of the image,so as to minimize the impact of using only a single spatial or spectral feature on the classification accuracy as much as possible.In order to make better use of the spatial information,we take the pixels to be classified as the center,and the data cube of S*S*B size is divided as the input of the network with the zero-padding method.After using softmax to classify,the best value of S is selected according to the classification result.Thirdly,in view of the fact that the current network is easy to misclassify the edge parts of two adjacent types of ground objects,a multi-scale 3D-2D deep convolutional neural network(MS-3DNet)based on multi-scale fusion is proposed,which integrates the classification results of Hybrid SN(Hybrid Spectral Convolutional Neural Network),Hierarchy1 and Hierarchy2 at a decision level.The experimental results show that the classification accuracy of MS-3DNet network is significantly improved compared with the three networks,and the situation that the edge parts of adjacent two kinds of ground objects are easy to be misclassified in space has been significantly improved.Finally,in view of the difficulty of labeling hyperspectral remote sensing image data and the high cost of labeling,the classification effect of MS-3DNet network on small sample data set is verified.The experimental results show that the proposed MS-3DNet network greatly shortens the training time on the premise of ensuring the overall classification accuracy on the small sample labeled data set.