Dimensionality Reduction And Classification Of Hyperspectral Imagery Based On Superpixelwise Kernel Principal Component Analysis

Dimensionality reduction is an important preprocessing step in hyperspectral image applications,especially when the number of samples is small.In this paper,a dimensionality reduction method called superpixelwise kernel principal component analysis(Super KPCA)is proposed.By performing kernel principal component analysis(KPCA)on each homogeneous region(superpixel)obtained by image segmentation algorithm,Super KPCA can use superpixels to mine spatial information,while using kernel method to deal with the nonlinear features that are widely present in hyperspectral images.To perform SuperKPCA,an appropriate image segmentation algorithm is required.Entropy rate superpixel segmentation(ERS)and simple linear iterative clustering(SLIC)are two widely used algorithms in graph-based and gradient-based image segmentation algorithms,respectively.Therefore,in this paper,the differences in the effects of performing SuperKPCA using these two image segmentation algorithms are compared in detail.The input images of ERS and SLIC are grayscale or RGB images,however,a hyperspectral image usually contains hundreds of bands.Therefore,for these two image segmentation algorithms,when the first principal components obtained by principal component analysis(PCA),KPCA and minimum noise fraction(MNF)transformation are used as the fundamental images for segmentation,respectively,the differences in the dimensionality reduction results of performing SuperKPCA are compared in detail(these different first principal components contain different main information of the hyperspectral image).Both ERS and SLIC tend to form a corresponding number of superpixels of similar size according to the input number of superpixels.To capture objects of different sizes in the real word more accurately,multiscale segmentation-based SuperKPCA(MSuperKPCA)is proposed.It first segments the hyperspectral image multiple times based on different number of superpixels(different segmentation scale),then performs SuperKPCA on each segmentation scale,and finally fuses the classification results after multiscale dimensionality reduction using majority voting strategy.Since the first principal components obtained by PCA,KPCA,and MNF contain different main information of hyperspectral images,their multiscale dimensionality reduction and classification results will also be different.Therefore,a method to improve the classification performance by fusing multiscale classification results obtained from different fundamental images is proposed,and it is named 3-MSuperKPCA.By validating the proposed methods on three public datasets: Indian Pines,Pavia University,and Salinas,the following conclusions are obtained:(1)Since superpixels can contain both spatial and spectral information of hyperspectral images,the ability of KPCA to process nonlinear features can be greatly improved by performing KPCA on each superpixel.For these three public datasets,the classification accuracy obtained by SuperKPCA can be 13.63% to 74.36% higher than the standard KPCA.(2)It is important to choose an appropriate fundamental image for segmentation.When the first principal components obtained by PCA,KPCA,and MNF are used for segmentation,respectively,the obtained dimensionality reduction results are different,and for different hyperspectral images,the optimal fundamental image for segmentation is also different.Moreover,when performing ERS based on the optimal fundamental image for segmentation,for Indian Pines,Pavia University and Salinas images: compared with SuperPCA(it uses the first principal component obtained by PCA as the fundamental image for segmentation,and performs PCA on each superpixel),the classification accuracy obtained by SuperKPCA can be improved by 0.06%-0.74%,3.88%-4.37% and 0.39%-4.85%;using multiscale segmentation can further improve classification performance,and compared with MSuperPCA(multiscale segmentation-based SuperPCA),the classification accuracy obtained by MSuperKPCA can be improved by 0.66%-3.04%,1.26%-3.20% and 0.29%-2.54%.This shows that the proposed dimensionality reduction method based on Super KPCA can have a better dimensionality reduction effect than the dimensionality reduction method based on SuperPCA.Moreover,since Pavia University has the most complex texture features and more nonlinear features,SuperKPCA and MSuper KPCA,which can better deal with nonlinear features,have the most significant improvement in the classification accuracy of this dataset.Since the boundary adherence of the result obtained by SLIC is not as good as ERS,the dimensionality reduction effect obtained by SLIC is also worse than ERS.In the experiments of this paper,ERS is a more suitable choice for obtaining superpixels.(3)Since different first principal components contain different main information of hyperspectral images,the classification results can usually be further improved by fusing the multiscale segmentation results obtained based on different first principal components,and this improvement is more significant when the classification accuracy before fusion is relatively low.For Indian Pines,Pavia University,and Salinas images,compared with the performance of performing MSuperKPCA based on the optimal fundamental image for segmentation: when ERS is used for segmentation,except for the classification accuracy of Pavia University when the number of training samples in each class is 20 pixels,the classification accuracy obtained by 3-MSuperKPCA can be increased by 0.54%-2.68%,0.12%-1.10%,and 0.01%-0.08%,respectively;when SLIC is used for segmentation,the classification accuracy obtained by 3-MSuperKPCA can be improved by 0.50%-3.79%,0.63%-5.49%,and 0.15%-1.64%,respectively.