Application Of Deep Learning In Hyperspectral Image Classification

Hyperspectral image classification is the key research content in the field of remote sensing image processing.Image feature extraction plays an important role in classification,while the high dimensionality and nonlinearity of hyperspectral images make it difficult to extract features.In addition,marking hyperspectral images requires a lot of manpower and material resources,and limited image labels tend to cause over-fitting of the algorithm,which reduces the classification accuracy.In this paper,two solutions are proposed for the above two problems.The specific research work is as follows:(1)In order to solve the problem that high-dimensionality and nonlinearity of hyperspectral data are difficult to extract joint feature of image spectrum,this paper proposes an improved spectral clustering algorithm called hypergraph algorithm.The algorithm constructs two types of hypergraphs,one for extracting spectral features of hyperspectral images and the other for extracting spatial features of hyperspectral images.(2)Combining the hypergraph algorithm with the Support Vector Machine(SVM)algorithm to form the G-SVM(Graph SVM)algorithm,achieving 95.53%,96.42%,and 96.17% accuracy of the experimental classification in the Indian Pines,Pavia University,and Salinas data sets respectively.The results show that the hypergraph algorithm can effectively extract the spectral features and spatial features of hyperspectral images,and verify the effectiveness of G-SVM algorithm in hyperspectral image classification.(3)For the hyperspectral image data label is limited,it is easy to cause the algorithm to over-fitting.This paper constructs a novel Deep Convolutional Neural Network(CNN).The model`s over-fitting problem is solved by using Dropout and L2 regularization methods;it is combined with the hypergraph algorithm to form the G-CNN(Graph CNN)algorithm;and different numbers of training samples are used for comparison experiments.The results show that the classification accuracy of the algorithm is 96% when the number of training samples is 200,the number of training samples is reduced to 50,and the classification accuracy of 90% is also achieved.Compared with other algorithms,the G-CNN algorithm has a small interference with the number of training samples,high classification accuracy,and strong ability to solve over-fitting problems.It has certain application value in hyperspectral image classification.