Research On High Efficient Classification And Unmixing Algorithms For Hyperspectral Remote Sensing Imagery

Hyperspectral remote sensing technology,rising in the 1980 s,is an advanced technology incorporating the theory of spectroscopy and imaging technology.Hyperspectral imaging,consisting of hundreds of spectral bands,can provide abundant and sophisticated observed data for human society.As a research hotspot,identification and analysis of ground objects is an integral part of hyperspectral image processing,which can be solved mainly via hyperspectral classification and spectral unmixing.Hyperspectral classification is a kind of pixel-level processing technology,through labeling and recognizing the categories of the pixels,while spectral unmixing is a sub-pixel level processing technology,through decomposing a mixed pixel into pure spectral signatures weighted by the corresponding fractional abundances.Hyperspectral image with high spectral resolution,which has an extraordinary advantage of merging image with spectrum,can offer a wealth of detailed information on land cover for hyperspectral classification and spectral unmixing.However,the accompanying challenges and difficulties occurred as follows:(1)Hyperspectral imaging is limited by the spatial resolution of the sensors,and influenced by the natural environmental factors,such as light,air,cloud thickness and other factors.Thus,the same object may exhibit different spectral signatures and the different objects may exhibit the same spectrum.Both phenomena enhance the difficulty of hyperspectral classification and spectral unmixing.(2)Hyperspectral imagery with high dimensions can lead the “Hughes” phenomenon,so the hyperspectral classification performance increases at first and then decreases sharply.(3)Large amounts of hyperspectral data bring a great amount of computation for hyperspectral classification and unmixing.To solve these problems of classification and unmixing in hyperspectral imagery,this dissertation studies hyperspectral classification algorithms based on artificial neural network and unmixing algorithms based on sparse regression.The main contributions in this dissertation are summarized as follows:(1)Research on optimizing extreme learning machine(ELM)for hyperspectral image classificationDue to the large amounts of hyperspectral data,the high computational complexity and longsample training time,this paper pioneers applying the extreme learning machine method into hyperspectral image classification,and proposes a highly effective hyperspectral image classification algorithm based on an optimizing extreme learning machine.This paper explores an empirical linear relationship based on a certain mathematical theory between the number of training samples and the number of hidden neurons.To decrease the calculation cost deriving from large amount of training samples,such an empirical linear relationship can be easily extended from small training sets to large training sets.Meanwhile,we realize the kernel version of extreme learning machine algorithm via using the radial basis kernel function and the empirical linear relationship also can be extended into the kernel version of ELM(KELM).Compared with the classical support vector machine and the kernel support vector machine methods with good classification performance,the proposed method not only improves the classification accuracy but also improves the sample training speed greatly.(2)Research on an efficient radial basis function(RBF)neural network for hyperspectral remote sensing image classificationTo solve the problems of too many adjustable parameters in neural networks and low classification accuracy of the traditional hyperspectral classification algorithms based on neural networks,here we propose an efficient radial basis function neural network for hyperspectral classification.This paper studies and designs the network structure with no parameters needs to be manually tuned.By incorporation a simple spatial averaging filter without additional computational cost,the spatial information and spectral information can be incorporated effectively.Thus,the classification accuracy can be further improved.By using the matrix decomposition lemma,the large matrix can be decomposed into several small matrices.The inverse of large matrix can be obtained by the combination of the inverse of small matrices,therefore,the parallel processing method is implemented.Compared with the kernel support vector machine method,the sample training speed of the proposed method increases slightly and the classification accuracy improves greatly.(3)Research on spectra-weighted total variation regularization based sparse hyperspectral unmixingThe existing representation of spatial information is too simple to represent the true distribution of the observed pixels in traditional hyperspectral unmixing algorithms.Thispaper proposes a spectra-weighted total variation regularization based sparse hyperspectral unmixing algorithm.This method studies and analyzes the relationship between spatial information and spectral information.A novel total variation regularization using different norms of vector,in which the difference of fractional abundances is weighted by the difference of the according observed pixel spectra,is proposed to elegantly model the spatial information.The different sparse unmixing models utilizing the four different regularizers can form four novel different sparse unmixing methods.Then for applying into the subsequent sparse unmixing model,the optimal regularizer is analyzed and determined.Compared with the traditional sparse unmixing algorithms,the proposed method can improve the unmixing performance by 2 to 16 d B.(4)Research on regional collaborative sparse unmixing of hyperspectral imageryDue to the fact that both the pixel-based sparse unmixing and image-based collaborative sparse unmixing algorithms ignore the mutual interferences between pixels in heterogeneous region and interactions between internal pixels in homogeneous region.This paper presents a region-based collaborative sparse unmixing algorithm for hyperspectral imagery.This algorithm studies the characteristic between the spectral domain and spatial domain,considers the advantages from spatial and spectral information,analyzes the relationship between the segmentation and unmixing performance,and verifies the influences between the different segmentations methods and the hyperspectral unmixing performance.In this paper,the pixels in the same homogeneous region are demixed in the same way via collaborative sparse unmixing method.Compared with the pixel-based sparse unmixing and image-based collaborative sparse unmixing methods,the proposed method can improve the unmixing performance by 4 to 7d B.Furthermore,the proposed method can support the knowledge of theory and practice for the combination between classification and unmixing methods.