New Methods of Spectral-Density Based Graph Construction and Their Application to Hyperspectral Image Analysis

摘要

The past decade has seen the emergence of many hyperspectral image (HSI) analysis algorithms based on graph theory and derived manifold-coordinates. Yet, despite the growing number of algorithms, there has been limited study of the graphs constructed from spectral data themselves. Which graphs are appropriate for various HSI analyses---and why? This research aims to begin addressing these questions as the performance of graph-based techniques is inextricably tied to the graphical model constructed from the spectral data. We begin with a literature review providing a survey of spectral graph construction techniques currently used by the hyperspectral community, starting with simple constructs demonstrating basic concepts and then incrementally adding components to derive more complex approaches. Throughout this development, we discuss algorithm advantages and disadvantages for different types of hyperspectral analysis. A focus is provided on techniques influenced by spectral density through which the concept of community structure arises. Through the use of simulated and real HSI data, we demonstrate density-based edge allocation produces more uniform nearest neighbor lists than non-density based techniques through increasing the number of intracluster edges, facilitating higher k-nearest neighbor (k-NN) classification performance. Imposing the common mutuality constraint to symmetrify adjacency matrices is demonstrated to be beneficial in most circumstances, especially in rural (less cluttered) scenes. Many complex adaptive edge-reweighting techniques are shown to slightly degrade nearest-neighbor list characteristics. Analysis suggests this condition is possibly attributable to the validity of characterizing spectral density by a single variable representing data scale for each pixel. Additionally, it is shown that imposing mutuality hurts the performance of adaptive edge-allocation techniques or any technique that aims to assign a low number of edges (<10) to any pixel. A simple k bias addresses this problem.;Many of the adaptive edge-reweighting techniques are based on the concept of codensity, so we explore codensity properties as they relate to density-based edge reweighting. We find that codensity may not be the best estimator of local scale due to variations in cluster density, so we introduce and compare two inherently density-weighted graph construction techniques from the data mining literature: shared nearest neighbors (SNN) and mutual proximity (MP). MP and SNN are not reliant upon a codensity measure, hence are not susceptible to its shortcomings. Neither has been used for hyperspectral analyses, so this presents the first study of these techniques on HSI data. We demonstrate MP and SNN can offer better performance, but in general none of the reweighting techniques improve the quality of these spectral graphs in our neighborhood structure tests. As such, these complex adaptive edge-reweighting techniques may need to be modified to increase their effectiveness.;During this investigation, we probe deeper into properties of high-dimensional data and introduce the concept of concentration of measure (CoM)---the degradation in the efficacy of many common distance measures with increasing dimensionality---as it relates to spectral graph construction. CoM exists in pairwise distances between HSI pixels, but not to the degree experienced in random data of the same extrinsic dimension; a characteristic we demonstrate is due to the rich correlation and cluster structure present in HSI data. CoM can lead to hubness---a condition wherein some nodes have short distances (high similarities) to an exceptionally large number of nodes. We study hub presence in 49 HSI datasets of varying resolutions, altitudes, and spectral bands to demonstrate hubness effects are negligible in a k-NN classification example (generalized counting scenarios), but we note its impact on methods that use edge weights to derive manifold coordinates or splitting clusters based on spectral graph theory requires more investigation.;Many of these new graph-related quantities can be exploited to demonstrate new techniques for HSI classification and anomaly detection. We present an initial exploration into this relatively new and exciting field based on an enhanced Schroedinger Eigenmap classification example and compare results to the current state-of-the-art approach. We produce equivalent results, but demonstrate different types of misclassifications, opening the door to combine the best of both approaches to achieve truly superior performance. A separate less mature hubness-assisted anomaly detector (HAAD) is also presented.