Spatial And Spectral Dimensionality Reduction And Classification By Machine Learning Of Hypespectral Image

Hyperspectral remote sensing imaging has gradually become an indispensable technical means in ground target detection and prediction because of high resolution in the spectral domain and fine structural information in space of hyperspectral remote sensing images(Hyperspectral Images,HSI).However,dozens or even hundreds of high-dimensional feature bands and the complexity of detecting the spatial structure characteristics bring great difficulties to the classification and prediction of ground targets.With the increase of feature band dimensions and the complexity of the spatial structure of features,the performance and accuracy of machine learning classifiers used for remote sensing feature target recognition are severely affected.Therefore,the study of dimensionality reduction methods for hyperspectral images in the spectral and spatial domains is an extremely important research goal.This thesis has carried out comprehensive research work in two aspects: spectral domain feature band dimensionality reduction and band(channel)spatial domain dimensionality reduction.The machine learning classifiers were tested and compared on the three test data sets of Indian Pines,Salinas and Tea Farm.1.Spectral domain band dimensionality reduction mainly studies Recursive Feature Elimination(RFE),Linear Dimensionality Reduction(LDA,PCA and ICA),non-linear dimensionality reduction(LLE),and deep dimensionality reduction methods such as Deep PCA and Auto Encoder(AE).Among them,PCA and AE are more suitable for dimensionality reduction in the spectral domain,the HSI can be effectively reduced 80-224 bands to 8 dimensions,and significantly improve the accuracy of ground object classification.2.The spatial dimensionality reduction of hyperspectral imagery compares the methods of conventional spatial flattening linear dimensionality reduction and AE dimensionality reduction,2D-PCA and PCANet.Spatial dimensionality reduction can compress the spatial neighborhood information with higher dimensions to lower dimensions(2×2 and 3×3),which significantly reduces the time consumption of the classifier,and improves the accuracy of ground object classification.Among them,the 2D-PCA-based boosting tree model reaches and exceeds the accuracy of the deep learning model.3.Aiming at the problem of spatial spectrum feature extraction and dimensionality reduction in fusion,which are commonly used in hyperspectral image classification,the different methods and strategies in the process of feature extraction with Gabor and feature reduction in the spectral domain and the spatial domain are compared.Research has shown that the order and process of feature extraction and feature reduction have a significant impact on the fusion effect of spatial spectrum information.The feature extraction based on the original spectral image and then the dimensionality reduction is better than PCA after the dimensionality reduction in the spectral domain.In terms of feature extraction and dimensionality reduction,the fusion effect of spatial spectrum information is better.The accuracy of the former is 0.83-6.01% higher than that of the latter,indicating that for the classification of hyperspectral image features,the effective extraction and maintenance of spatial information is extremely important.4.The application methods and effects of hyperspectral image dimensionality reduction methods in five lightweight networks such as Shuffle Net,SENet and Ghost Net are comparatively studied.PCA and encoder dimensionality reduction in the spectral domain can reduce the time consumption of deep learning to 3-5% while maintaining classification accuracy.At the same time,the denoising effect of PCA and encoder in the dimensionality reduction process can improve the convergence speed of the deep learning model and enhance the generalization of the model,especially for the model embedded with the attention mechanism,the effect is more significant.