Draper Laboratory and MRAC have recently completed a comprehensive study to quantitatively evaluate deception
detection performance under different interviewing styles. The interviews were performed while multiple physiological
waveforms were collected from participants to determine how well automated algorithms can detect deception based
upon changes in physiology. We report the results of a multi-factorial experiment with 77 human participants who were
deceptive on specific topics during interviews conducted with one of two styles: a forcing style which relies on more
coercive or confrontational techniques, or a fostering approach, which relies on open-ended interviewing and elements of
a cognitive interview. The interviews were performed in a state-of-the-art facility where multiple sensors simultaneously
collect synchronized physiological measurements, including electrodermal response, relative blood pressure, respiration,
pupil diameter, and ECG. Features extracted from these waveforms during honest and deceptive intervals were then
submitted to a hypothesis test to evaluate their statistical significance. A univariate statistical detection algorithm then
assessed the ability to detect deception for different interview configurations. Our paper will explain the protocol and
experimental design for this study. Our results will be in terms of statistical significances, effect sizes, and ROC curves
and will identify how promising features performed in different interview scenarios.
In this article we present a method for hyperspectral band
selection that yields superior classification results while only
using a subset of the available bands. The approach originates
from a comprehensive physical and mathematical understanding of
the distance metrics used to compare hyperspectral signals, and it
exploits an exact decomposition of a common metric, the Spectral
Angle Mapper (SAM), to select bands which increase the angular
contrast between target classes. Using real spectroradiometer and
sensor data collected by the HYDICE sensor, the technique
significantly improves the discrimination performance for two
spectrally similar classes, while using only a fraction of the
available bands. The approach is extended to a hierarchical
architecture for material identification using spectral libraries
that is shown to outperform the traditional angle-based classifier
which employs all available bands. Consequently, better material
identification performance can be achieved using significantly
fewer bands, thus introducing dramatic benefits for the design and
utilization of spectral libraries.
Spectral unmixing has emerged as a key application arising from the wealth of spectral measurements in hyperspectral processing. Several communities have shown great interest in the decompositional analysis of mixed pixels. Unmixing provides the ability to decompose mixed pixels in terms of distinct, unique substances, and provide a foundation for doing sub-pixel material identification. We undertake this comparison of unmixing algorithm performance with the knowledge that many algorithms exist, and new methods are constantly being explored and tested. Several disciplines are participating in the attempt to perform unmixing, such as geology, geophysics, engineering, and analytical chemistry.
This investigation explores how hyperspectral distance metrics may be used as indicators of relative water depth in a coastal region. Spectral reflectance characteristics of near-shore waters imaged by an airborne hyperspectral sensor are examined. Commonly used hyperspectral distance metrics are applied to the data with the goal of distinguishing the spectra derived from various water depths. To improve the separability of the spectra, this study also examines, for one distance metric, the effect of processing only a subset of spectral bands recorded by the sensor. The concept of selecting a subset of bands extends to improving the performance of algorithms that process hyperspectral data for detection, classification, or estimation. An additional benefit is reducing the dimensionality of the dat and, thereby, the computational load. The key to reaching both of these objectives is to understand and match physical processes to appropriate mathematical metrics performance measures in a comprehensive framework. The overall process is driven both by empirical analysis of hyperspectral data and by mathematical examination of the spectra.
The objective of hyperspectral processing algorithms is to efficiently capitalize on the wealth of information in the scene being imaged. Radiation collected in hundreds of contiguous electromagnetic channels and stored as data in a vector provides insight about the reflective and emissive properties of each pixel in the scene. However, it is not intuitively clear that for common applications such as estimation, classification, and detection that the best performance results from utilizing every measurement in the vector. In fact, it is quite easy to show that for some tasks, more data can degrade performance. In this paper, we explore the role of metrics and best bands algorithms in the context of maximizing the performance of hyperspectral algorithms. Specifically, we first focus on creating an intuitive framework for physical information measured by a sensor. Then, we examine how it is translated into numerical quantities by a distance metric. We discuss how two common distance metrics for hyperspectral signals, the Spectral Angle Mapper (SAM), and the Euclidean Minimum Distance (EMD), quantify the distance between two spectra. Focusing on the SAM metric, we demonstrate, in the context of target detection, how the separability of the two spectra can be increased by retaining only those bands that maximize the metric. Finally, this intuition about the best bands analysis for SAM is extended to the Generalize Likelihood Ratio Test (GLRT) for a practical target/background detection scenario. Results are shown for a scene imaged by the HYDICE sensor demonstrating that the separability of targets and background can be increased by carefully choosing the best bands for the test.
In this paper, we introduce a set of taxonomies that hierarchically organize and specify algorithms associated with hyperspectral unmixing. Our motivation is to collectively organize and relate algorithms in order to assess the current state-of-the-art in the field and to facilitate objective comparisons between methods. The hyperspectral sensing community is populated by investigators with disparate scientific backgrounds and, speaking in their respective languages, efforts in spectral unmixing developed within disparate communities have inevitably led to duplication. We hope our analysis removes this ambiguity and redundancy by using a standard vocabulary, and that the presentation we provide clearly summarizes what has and has not been done. As we shall see, the framework for the taxonomies derives its organization from the fundamental, philosophical assumptions imposed on the problem, rather than the common calculations they perform, or the similar outputs they might yield.
Real-time detection and identification of military and civilian targets from airborne platforms using hyperspectral sensors is of great interest. Relative to multispectral sensing, hyperspectral sensing can increase the detectability of pixel and subpixel size targets by exploiting finer detail in the spectral signatures of targets and natural backgrounds. A multitude of adaptive detection algorithms for resolved or subpixel targets, with known or unknown spectral characterization, in a background with known or unknown statistics, theoretically justified or ad hoc, with low or high computational complexity, have appeared in the literature or have found their way into software packages and end-user systems. The purpose of this paper is threefold. First, we present a unified mathematical treatment of most adaptive matched filter detectors using common notation, and we state clearly the underlying theoretical assumptions. Whenever possible, we express existing ad hoc algorithms as computationally simpler versions of optimal methods. Second, we assess the computational complexity of the various algorithms. Finally, we present a comparative performance analysis of the basic algorithms using theoretically obtained performance characteristics. We focus on algorithms characterized by theoretically desirable properties, practically desired features, or implementation simplicity. Sufficient detail is provided for others to verify and expand this evaluation and framework. A primary goal is to identify best-of-class algorithms for detailed performance evaluation.
Symmetric nonlinear matched filters (SNMFs) involve the transformation of the signal spectrum and the filter transfer function through pointwise nonlinearities before they are multiplied in the transform domain. The resulting system is analogous to a multistage neural network. The experimental and theoretical results discussed indicate that SNMFs
hold considerable potential to achieve a high power of discrimination and resolution and large SNR. The statistical analysis of a particular SNMF in the two-class problem indicates that the performance coefficient of the SNMF is about four times larger than the performance coefficient of the classical matched filter. In terms of resolving closeby signals, there seems to be no limit to the achievable resolution. However, artifacts should be carefully monitored.