Spatial-Spectral Graph Embedding For Dimensionality Reduction Of Hyperspectral Data

Due to the characteristics of large amount data,non-linear,expensive to labelled samples,small sample size and high dimension of hyperspectral remote sensing data,the issue for dimensionality reduction of hyperspectral remote sensing data is investigated by using statistical learning,machine learning and image processing interdisciplinary theory and method.Based on image processing,graph embedding dimensionality reduction and representation theories,the main contributions are as follows.1.Incremental graph embedding based on spatial-spectral neighbors for hyperspectral image classification.Recently,graph embedding based algorithms have been developed in feature extraction for hyperspectral image.The key step for graph embedding algorithms is the construction of graph.The commonly used method is to manually choose nearest neighbors and then compute edge weights using the spectral feature.However,the adjacency graph is inappropriate due to the negligence of spatial information.What is more,the construction of graph only takes training samples or 6)nearest neighbors into account,which may lead to unsuitable graph representation.In this work,we propose an incremental graph embedding algorithm to construct a spatialspectral neighbor graph for hyperspectral image classification.The incremental graph embedding can spread the discrimination information contained in training samples to their neighbors until each testing sample has a pseudo label.The spatial affinity weights between unlabeled data points and their labeled neighbors are calculated according to the construction strategies of spatial-spectral neighbor graph.The pseudo label of the unlabeled data point is determined based on the maximum spatial affinity weights.Moreover,three weight strategies are designed for those samples nearby the decision boundary to improve the separability of different classes.In addition,the window size of spatial neighbors is able to be adjusted adaptively according to whether labeled data points in spatial neighbors exit.2.Spatial-spectral neighbor graph for dimensionality reduction of hyperspectral image classification.Recently,graph embedding-based algorithms have drawn increasing attention for dimensionality reduction of hyperspectral image classification.Graph construction is a critical step for those dimensionality reduction algorithms.Pairwise similarity graph is generally employed to reflect the geometric structure in the original data.However,it ignores the similarity of neighboring pixels.In order to further improve the classification performance,both spectral and spatial-contextual information should be taken into account in hyperspectral image classification.In this paper,a novel spatial-spectral neighbor graph is proposed for DR of incremental graph embedding classification,which consists of the following four steps.First,a superpixel-based segmentation algorithm is adopted to divide hyperspectral image into many superpixels.Second,a novel distance metric is utilized to reflect the similarity of two spectral pixels in each superpixel.In the third step,a spatial-spectral neighbor graph is constructed according to the above distance metric.At last,support vector machine with a composite kernel is adopted to classify the dimensionality-reduced hyperspectral image.3.Superpixel maximum noise fraction for dimensionality reduction of hyperspectral data.A superpixel maximum noise fraction algorithm is proposed to overcome the problem of a unified projection is employed for dimensionality reduction of hyperspectral data.Firstly,image segmentation algorithm is adopted to divide the hyperspectral image into many regions and extract contextual information of each pixel in hyperspectral image.Secondly,maximum noise fraction is employed to obtain a local projection and reduce the dimension of each pixel in each divided region and samples feature is obtained in low dimension subspace.Thirdly,feature representation is obtained by merging sample feature in each low dimension subspace for the total hyperspectral data.Finally,maximum voting mechanism decision fusion strategy is used to determine the final label of testing sample.4.Graph regularized adaptive joint collaborative representation for hyperspectral image classification.A graph regularized adaptive joint collaborative representation algorithm is proposed to overcome low classification accuracy problem caused by insufficient utilization of spatial information for hyperspectral image classification.Firstly,the bilateral filter is adopted to extract spatial information for hyperspectral image to fully explore the spatial information of each pixel.Secondly,a graph regularized term is introduced into the objective function of joint collaborative representation to maintain the local manifold structure of hyperspectral image data.On one hand,the image segmentation is used to adjust the shape of the spatial neighborhood;on the other hand,an adaptive space-spectral feature fusion strategy is proposed by assigning different weights to the spatial neighbors of the central pixels.Finally,the label of each testing sample is determined by utilizing the least errors criterion.Experimental results on Indian Pines,Pavia University,KSC and Salinas datasets demonstrated that the proposed algorithms can achieve higher overall accuracy and Kappa coefficient and obtain smooth classification map by employing hyperspectral spatial and spectral domain feature.