Alan P. Schaum

Application of stochastic mixing models to hyperspectral detection problems

Hyperspectral images are frequently analyzed in terms of the linear mixing model, which assumes that observed pixel radiances are generated by linear combinations of a relatively small number of spectral constituent signatures. The constituents are generally modeled as deterministic points in color space whose locations can in principle be found by exploiting the convex geometry of the mixture simplex. This paper presents an alternative stochastic mixing model (SMM) that associates scene constituents with distinct probability distributions,the parameters of which are estimated from observed data using statistical clustering methods. By defining distributions corresponding to both constituent and mixed pixel classes, the SMM can often be used to generate physically meaningful classification maps of spectrally-heterogeneous scenes. However, the most significant application of the stochastic approach is to hyperspectral target detection based on statistical decision theory concepts. A SMM can provide accurate parametric estimates of the spectral distributions for mixed scenes, thereby improving the performance of hypothesis testing procedures that utilize replacement targets with spectral signature uncertainty. SMM principles and applications are illustrated using hyperspectral imagery collected by the LIFTIRS and HYDICE instruments.

Airborne hyperspectral detection of small changes.

Hyperspectral change detection offers a promising approach to detect objects and features of remotely sensed areas that are too difficult to find in single images, such as slight changes in land cover and the insertion, deletion, or movement of small objects, by exploiting subtle differences in the imagery over time. Methods for performing such change detection, however, must effectively maintain invariance to typically larger image-to-image changes in illumination and environmental conditions, as well as misregistration and viewing differences between image observations, while remaining sensitive to small differences in scene content. Previous research has established predictive algorithms to overcome such natural changes between images, and these approaches have recently been extended to deal with space-varying changes. The challenges to effective change detection, however, are often exacerbated in an airborne imaging geometry because of the limitations in control over flight conditions and geometry, and some of the recent change detection algorithms have not been demonstrated in an airborne setting. We describe the airborne implementation and relative performance of such methods. We specifically attempt to characterize the effects of spatial misregistration on change detection performance, the efficacy of class-conditional predictors in an airborne setting, and extensions to the change detection approach, including physically motivated shadow transition classifiers and matched change filtering based on in-scene atmospheric normalization.

Spectral subspace matched filtering

The linear matched filter has long served as a workhorse algorithm for illustrating the promise of multispectral target detection. However, an accurate description of a targets distribution usually requires expanding the dimensionality of its intrinsic signature subspace beyond what is appropriate for the matched filter. Structured backgrounds also deviate from the matched filter paradigm and are often modeled as clusters. However, spectral clusters usually show evidence of mixing, which corresponds to the presence of different materials within a single pixel. This makes a subspace background model an attractive alternative to clustering. In this paper we present a new method for generating detection algorithms based on joint target/background subspace modeling. We use it first to derive an existing class of GLF detectors, in the process illustrating the nature of the real problems that these solve. Then natural symmetries expected to be characteristic of otherwise unknown target and background distributions are used to generate new algorithms. Currently employed detectors are also interpreted using the new approach, resulting in recommendations for improvements to them.

Continuum fusion: a theory of inference, with applications to hyperspectral detection

A new theoretical framework is created for the class of detection problems traditionally addressed by the generalized likelihood ratio test. Absent prior knowledge that would permit implementation of the optimal detector, a family of optimal detectors is fused according to any one of a group of criteria. Geometrical solutions are presented to several specific problems motivated by hyperspectral signal processing. For the general case, a set of partial differential relations is derived. The generalized likelihood ratio test is shown to be equivalent to one of several flavors of continuum fusion detector.

Subclutter target detection using sequences of thermal infrared multispectral imagery

Multivariate correlation techniques can be used to enhance target contrast in spectroscopic imagery. But in most cases the detectability of dim targets remains limited by residual background clutter. If, however, multiple-time measurements can be made, detection performance can be markedly enhanced by an integrated spectral/temporal technique that exploits the correlated nature of background spectral trajectories. We demonstrate the detection of extreme subpixel objects, such as is required by long-range remote sensing systems. We also show that the time intervals between data collections can be long. The confusing effects of natural background evolution-in temperature distribution or illumination-can be distinguished from anomalous changes. Data collected with longwave infrared point- and imaging-spectrometers have validated the concept.

Advanced algorithms for autonomous hyperspectral change detection

Persistent ISR (intelligence surveillance and reconnaissance) has proven its value as a tactic for national defense. This activity can collect, in particular, information necessary for executing an important concept of operations: wide-area autonomous change detection over long time intervals. Here we describe the remarkable potential of hyperspectral remote sensing systems for enabling such missions, using either visible or thermal infrared wavelengths. First we describe blind change detection, in which no target knowledge is assumed. Targets that have moved can nevertheless be distinguished from naturally occurring background radiometric changes through the use of multivariate statistics informed by simple physics. Detection relies on the ability of hyperspectral algorithms to predict certain conserved properties of background spectral patterns over long time intervals. We also describe a method of mitigating the most worrisome practical engineering difficulty in pixel-level change detection, image misregistration. This has led, in turn, to a method of estimating spectral signature evolution using multiple-scene statistics. Finally, we present a signature-based detection technique that fuses two discrimination mechanisms: use of some prior knowledge of target spectra, and the fact that a change has occurred.

Continuum fusion methods of spectral detection

Abstract. The continuum fusion (CF) methodology produces new classes of multivariate detection algorithms, some of which have been used in spectral applications. CF principles apply to model-based problems in which not all parameter values are known, a common circumstance in hyperspectral operations. We reviewed the principal theoretical and applied CF results devised to date, summarized recent experimental results, and discussed in detail an important class of algorithms that illustrate the design freedom CF affords. Finally, we reviewed the fundamental CF principles as applied to a new category of model parameters only recently considered, involving a distinction in the form of a constraint that is not recognized by older methods.

Joint hyperspectral subspace detection derived from a Bayesian likelihood ratio test

The standard approach to solving detection problems in which clutter and/or target ditributions are modeled with unknown parameter is to apply the generalized likelihood ratio (GLR) test. This procedure automatically gernerates new estimates of the unknown model parameter for each new feature test value. An alternative approach is to estimate prior distribution for the unknown parameters. The associated Bayesian Likelihood Ratio (BLR) test can be used to generate many standard detectors for example, matched filtering or the GLR as special cases. For the particular problem of Joint Subspace Detection (JSD), several such Bayesian problems often lead to the same test as some GLR problem. Formulating such problems can lend insight into what types of background and target distributions are appropriate for a given GLR test. In addition, the added generality afforded by the new approach, in the form a selectable prior distributions, defines a wider exploratory space fro target detection. JSD can, for example, permit the incorporation of general types of experience gleaned from measurement programs. This paper explores these potentialities by applying several Bayesian formulations of the detection problem to hyperspectral data set.

Automatic ship detection from commercial multispectral satellite imagery

Commercial multispectral satellite sensors spend much of their time over the oceans. NRL has demonstrated an automatic processing system for finding ships at sea using commercially available multispectral data. To distinguish ships from whitecaps and clouds, a water/cloud clutter subspace is estimated and a continuum fusion derived anomaly detection algorithm is applied. This provides a maritime awareness capability with an acceptable detection rate while maintaining a low rate of false alarms. The system also provides a confidence metric, which can be used to further limit the false alarm rate.

The affine matched filter

The hyperspectral matched filter (MF) is a popular tool in remote sensing problems for locating objects that extend over several pixels. However, it is ideally suited only for the detection of sub-pixel targets with known mean signature in radiance space. Here we develop an alternative affine matched filter (AMF) that is more appropriate for detecting extended targets, and which accommodates practical uncertainties in target signature knowledge. In particular, AMF is ideal when used in conjunction with another new method we develop, called Virtual Relative Calibration (VRC), for generating a radiance space representation of a laboratory reflectance signature. VRC is related to the QUick Atmospheric Correction(QuAC) method but is much simpler. We also devise and test an extension of AMF that is meant to reduce false alarms caused by dark pixels. This Joint AMF (JAMF) expands the standard statistical model of hyperspectral backgrounds to accommodate variable illumination levels, analogously to how AMF treats target level uncertainty.