Classification Of Hyperspectral Remote Sensing Images By Convolutional Neural Network Based On Data Augmentation

Remote sensing technology is a detection technology developed in the 1960s.It is a comprehensive technology to detect and identify ground scenes.The emergence of hyperspectral remote sensing images has greatly promoted the development of remote sensing technology.Hyperspectral remote sensing images can not only represent the spatial characteristics of objects,such as structure,shape and position relation,but also contain spectral information representing the specific physical materials of objects.The outstanding advantage of "combining spectral and spatial information" makes hyperspectral remote sensing images play an important role in geological survey,agricultural monitoring,city planning and other fields.Hyperspectral remote sensing image classification,as a key technology to solve the above problems,has always been a subject of great concern and wide application prospect.By virtue of its implicit learning and depth feature extraction,convolutional neural network has already had a great influence on the classification of hyperspectral remote sensing images.However,the convolution neural network needs to input a large number of labeled samples in the training process,and the small sample problem of hyperspectral remote sensing image greatly limits the accuracy of information interpretation.In order to solve the above problems,based on the existing processing technology,this paper begins from the realization of comprehensive data augmentation of hyperspectral remote sensing images,and carries out the data augmentation at the level of data quantity and data quality respectively.The main research contents are as follows:Firstly,in order to solve the problem of shortage of labeled samples in hyperspectral remote sensing images and high cost of manual labeling,pixel block pairing method is adopted to increase the number and diversity of samples.In this method,pixel blocks are constructed with the pixel points of training as the center,and then the pixel blocks of the same kind and the pixel blocks of different kinds are paired,and finally the paired results are input into the convolutional neural network for feature extraction and model construction.The experiment is carried out based on several open data sets of hyperspectral remote sensing images,and the detailed experimental results strongly confirmed the superiority of the proposed method and fully demonstrate the great potential of the paired method in the field of data quantity augmentation.Secondly,In order to solve the problems of poor quality of labeled samples of hyperspectral remote sensing images,high spectral resolution but low spatial resolution,and extremely easy to be interfered by complex weather conditions such as fog in the process of data collection,this paper rejects the single method of detection and analysis using hyperspectral data,hyperspectral images(HSI)and light detection and ranging(LiDAR)images are fused using principal component analysis(PCA)to achieve data cooperation and data complementarity of multi-source information.In addition,the spatial and spectral features of the fused images are extracted by the dual-tunnel convolutional neural network model to complete the classification task.In this experiment,MUUFL Gulfport is selected as the standard data set,and the proposed method is compared with many excellent research methods at home and abroad.The results show the necessity of data quality augmentation and the superiority of the proposed method.