Research On Dimensionality Reduction And Spectral-Spatial Classification Methods For Hyperspectral Image

Hyperspectral Image(HSI)has become an indispensable data carrier in the process of understanding and transforming the world,thanks to its fine spectral resolution and good ground resolution.HSI usually has dozens or even hundreds of bands and has a ground resolution of several meters or even sub-meter levels.Generally speaking,higher spectral resolution can enhance the ability of classifier to distinguish complex ground objects.Higher ground resolution makes it possible to detect the complex and changeable ground objects in more detail.However,with the increase of spectral resolution and ground resolution,the law of data distribution of various ground objects in feature space becomes extremely complex.This brings great difficulties to the accurate identification of ground objects by using hyperspectral remote sensing technology.Meanwhile,the high-dimensional features of HSI not only easily lead to the occurrence of Hughes phenomena,but also increase the computational cost of classification.The big data features of HSI require us to propose fast and accurate classification methods to facilitate the application of remote sensing technology in practical problems.The high-dimensional feature and big data feature of HSI make it more difficult to obtain the distribution of ground objects by traditional visual interpretation methods.As a result,classifying HSI by using artificial intelligence has been one of the popular topics in the field of remote sensing in recent years.Existing research results show that satisfactory classification result cannot be obtained by using classical human intelligence methods directly.Therefore,it is still worthy of further study how to reduce the dimensionality of HSI while ensuring the classification accuracy,and how to develop fast and accurate spectral-spatial classification methods by effectively fusing the spectral and spatial information.To address the problems mentioned-above,the following works have been investigated in this dissertation.(1)Based on fuzzy c-mean algorithm,subspace decomposition technology,maximum entropy principle,gray wolf optimization algorithm,an unsupervised band selection methods is proposed to deal with the problem of high-dimensional HSI.The proposed dimensionality reduction method can select informative band subsets from the original bands,which effectively reduces the redundancy between the selected bands and avoids the occurrence of Hughes phenomena.In addition,the computing time of the classification algorithm is reduced while ensuring the classification performance.Experimental results on three standard test datasets demonstrate the effectiveness of the proposed method.(2)To reduce the influence of noisy pixels on the classification results,an effective spectral-spatial classification framework is proposed based on superpixel and discontinuity preserving relaxation.In the proposed classification framework,the discontinuity preserving relaxation and the spatial information of pixels are used to smooth the noisy pixels effectively in preprocessing.In post-processing,the application of superpixel can better improve the classification results.Moreover,a popular superpixel segmentation method,simple linear iterative clustering(SLIC)is improved.The improved SLIC algorithm is parameter-free and can be directly applied to segment HSI with arbitrary dimensions into superpixel.Experimental and comparative results confirm that the proposed method is superior to the other popular spectral-spatial classifiers for the same labeled ratio.(3)Although superpixel segmentation provides powerful tools for HSI classification,classifying HSI at superpixel level remains a challenging problem,due to the characteristics of adaptive size and shape of superpixels.Also,big data feature of HSI will undoubtedly increase the computation time of the classification.In order to partially tackle with this problem,a novel superpixel-level semi-supervised spectral-spatial classification method is introduced in this dissertation.The similarity between two superpixels is first defined based on the local average pseudo-nearest neighbor.Then superpixel-level classification is realized by k-nearest neighbor method.Finally,experimental results show that the proposed superpixel-level classifier performs better than the other several representative spectral-spatial classification methods.Three standard hyperspectral data sets,Indian Pines?Pavia University and Salinas,are used in our experiments.These three datasets are widely used to test the performance of HSI classification algorithms.It can be found by browsing the webpage.