Research On The Method Of Terrain Classification Based On Hyperspectral Image

In recent years,with the rapid development of remote sensing technology,hyperspectral remote sensing technology has gradually entered the public vision and has become the forefront of development in the field.At the same time,the increasingly mature imaging spectrometer technology has provided a strong guarantee for the quality of hyperspectral images.Hyperspectral images contain hundreds of narrow and continuous bands,and they can provide rich spatial and spectral information.At present,the terrain classification of hyperspectral images is a frontier scientific research topic of hyperspectral image processing technology.Observation of ground material information can serve the fields of national defense security,food safety,geological exploration,environmental protection,urban construction and so on.Simultaneously,because hyperspectral images have the characteristics of large amount of data,high redundancy and strong correlation between bands,it greatly increases the difficulty of classification and recognition.At present,the dimensionality reduction and classification of hyperspectral images have become the key issues in processing hyperspectral images,so research on them has important theoretical significance and application value.Based on the traditional hyperspectral image dimensionality reduction algorithm and classification method,this paper conducts research from the following two aspects.The main research contents are as follows:1.Hyperspectral image dimensionality reduction method based on principal component analysis(PCA)and MKt-SNE.First,under the premise of ensuring that the basic information is not lost,the PCA algorithm is used to achieve one-dimensional reduction.Second,the MKt-SNE algorithm(improved t-SNE algorithm)is used for secondary dimensionality reduction,which can better extract the essential characteristics of hyperspectral image.Among them,the MKt-SNE algorithm overcomes the problems of the traditional t-SNE algorithm affected by the distribution of high-dimensional spatial samples and the correlation interference between variables.When combined with the PCA algorithm,it not only considers the characteristics of coexistence of linear data and nonlinear data in hyperspectral images,but also overcomes the PCA algorithm’s overlapping projection results and linear inseparability,and the MKt-SNE algorithm occupies large memory and long running time.Experimental results show that compared with PCA,t-SNE,MKt-SNE and PCA-t-SNE,the PCA-MKt-SNE algorithm,the PCA-MKt-SNE algorithm has the best dimensionality reduction effect.The DBI index can be reduced by up to 1.1722 and the DI index can be increased by up to 0.1087.At the same time,the dimensionality reduction time of the PCA-MKt-SNE algorithm is as low as 45s,which is reduced by 49s compared with the MKt-SNE algorithm.Therefore,under the premise of ensuring the operation rate,this method effectively improves the classification accuracy and operation rate.2.Hyperspectral image classification method combining multi-layer feature SENet and multi-scale wide residual.First,the PCA-MKt-SNE algorithm is used for dimension reduction.Second,the multi-layer feature SENet network model and multi-scale wide residual network model are trained separately,and the results obtained by classifying the two models are fused using the weighted average method to improve the classification accuracy of hyperspectral images.Among them,on the one hand,the multi-layer feature SENet network model improves the effective classification features and suppresses the invalid classification features,and on the other hand,it overcomes the problem of high-level and middle-layer information being ignored.The multi-scale wide residual network model gets rid of the problem of the low computing resource utilization rate.Compared with the traditional model,the fused classification model achieves complementary performance of the classifier,making the classification method more stable and robust.The experimental results show that the overall classification accuracy of the SE-Inception-Resnet-MSWideResnet(SEIR-MSWR)network model is as high as 99.63%,and the Kappa coefficient is 0.99,which is the best classification effect.Compared with SVM,K nearest neighbor(KNN),wide residual network(WRN)and InceptionV2-Resnet,the overall classification accuracy can be improved by up to 6.85%,and the Kappa coefficient can be increased by up to 0.06.Simultaneously,it can be seen from the classification results that the algorithm has a certain improvement in accuracy.