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

In recent years,with the continuous development of hyperspectral remote sensing technology,the dimension of remote sensing data has also increased,bringing new challenges to the classification of hyperspectral data.Then,in the face of the high dimensionality,strong correlation,nonlinearity and large data volume of hyperspectral data,how to extract and classify the information efficiently becomes an important issue in the field of hyperspectral remote sensing.In view of the difficulties in the classification of hyperspectral remote sensing images by three-dimensional convolutional neural network models,this thesis has carried out in-depth research.A complete and effective classification scheme for hyperspectral remote sensing images is proposed.Firstly,First,virtual samples will be generated from the original samples of hyperspectral data to alleviate data imbalance problems;Then,the 3D convolutional neural network model is introduced to extract the 3D spatial spectrum features of the hyperspectral data cube,and incorporate a framework of dense residual connections to build deeper networks,and more expressive spatial spectral features are extracted to improve the accuracy of image classification;Finally,cross-validation is used to enhance the generalization ability of the model and combine the conditional random field to optimize the classification results.The main contents of this paper are as follows:(1)For data imbalance problems in hyperspectral remote sensing image data,Virtual samples are introduced to alleviate data imbalance by mixing raw samples and virtual samples to increase the number of hyperspectral image data samples.(2)For the problem of hyperspectral remote sensing image 3D cube data,the 3D convolutional neural network model is introduced into the hyperspectral remote sensing image classification,making full use of the optical spectrum features of hyperspectral remote sensing images enhances the classification accuracy of the algorithm,breaking through the limitations of traditional algorithm descriptions and insufficient use of spatial information.(3)The problem of network degradation caused by the increase of the number of network layers,that is,the growth rate of the accuracy of classification identification does not increase and decrease with the increase of the number of network layers,A deep dense residual connection framework is introduced,and the DR-3D-CNN algorithm is proposed to effectively mitigate the degradation of the network.(4)Using cross-validation methods to enhance model generalization capabilities,and combine the conditional random field to optimize the classification results.The effectiveness and advancement of the proposed algorithm are verified by experimental comparison,and the classification accuracy of hyperspectral remote sensing images is effectively improved.