Hyperspectral image segmentation using spatial-spectral graphs

Abstract

Spectral graph theory has proven to be a useful tool in the analysis of high-dimensional data sets. Recall that, mathematically, a graph is a collection of objects (nodes) and connections between them (edges); a weighted graph additionally assigns numerical values (weights) to the edges. Graphs are represented by their adjacency whose elements are the weights between the nodes. Spectral graph theory uses the eigendecomposition of the adjacency matrix (or, more generally, the Laplacian of the graph) to derive information about the underlying graph. In this paper, we develop a spectral method based on the 'normalized cuts' algorithm to segment hyperspectral image data (HSI). In particular, we model an image as a weighted graph whose nodes are the image pixels, and edges defined as connecting spatial neighbors; the edge weights are given by a weighted combination of the spatial and spectral distances between nodes. We then use the Laplacian of the graph to recursively segment the image. The advantages of our approach are that, first, the graph structure naturally incorporates both the spatial and spectral information present in HSI; also, by using only spatial neighbors, the adjacency matrix is highly sparse; as a result, it is possible to apply our technique to much larger images than previous techniques. In the paper, we present the details of our algorithm, and include experimental results from a variety of hyperspectral images.

Instance influence estimation for hyperspectral target signature characterization using extended functions of multiple instances

“…To understand the correlation between influence of superpixel and proposed approaches, the Gulfport dataset was over-segmented into about seven thousand superpixels having similar sizes using a Normalized Cut method. 9 Then, the rest of experiment was very similar with single point influence experiment with one difference which is instead of computing the target proportion of all points in each superpixel, we relied on the largest target proportion in each superpixel as the surrogate influence metric. This process is outlined in Alg.…”

A graph Laplacian regularization for hyperspectral data unmixing

“…Finally, k-nearest neighbors and thresholding are commonly used in order to set to zero small weights in W [22]. The authors of [5,20,21] propose different strategies for defining an affinity matrix that takes into account both the spatial and the spectral information of a pixel.…”

Schroedinger Eigenmaps with nondiagonal potentials for spatial-spectral clustering of hyperspectral imagery

“…Recent research 7-9 has shown that due to spatial correlations in hyperspectral imagery (especially in high resolution hyperspectral imagery), spatial information should be included, or fused, with the spectral information in order to more adequately represent the properties of the image data in the lower-dimensional space. Incorporating spatial information has been approached from multiple fronts: modifying the structure of the graph, 7, 8 modifying the edge weights, 9 or fusing spatial and spectral Laplacian matrices and/or their generalized eigenvectors. 8 We propose a different generalization of the LE algorithm for dimensionality reduction of hyperspectral imagery in a manner that fuses spatial and spectral information.…”