Michael J. Mendenhall

Relevance-Based Feature Extraction for Hyperspectral Images

Hyperspectral imagery affords researchers all discriminating details needed for fine delineation of many material classes. This delineation is essential for scientific research ranging from geologic to environmental impact studies. In a data mining scenario, one cannot blindly discard information because it can destroy discovery potential. In a supervised classification scenario, however, the preselection of classes presents one with an opportunity to extract a reduced set of meaningful features without degrading classification performance. Given the complex correlations found in hyperspectral data and the potentially large number of classes, meaningful feature extraction is a difficult task. We turn to the recent neural paradigm of generalized relevance learning vector quantization (GRLVQ) [B. Hammer and T. Villmann, Neural Networks, vol. 15, pp. 1059-1068, 2002], which is based on, and substantially extends, learning vector quantization (LVQ) [T. Kohonen, Self-Organizing Maps, Berlin, Germany: Springer-Verlag, 2001] by learning relevant input dimensions while incorporating classification accuracy in the cost function. By addressing deficiencies in GRLVQ, we produce an improved version, GRLVQI, which is an effective analysis tool for high-dimensional data such as remotely sensed hyperspectral data. With an independent classifier, we show that the spectral features deemed relevant by our improved GRLVQI result in a better classification for a predefined set of surface materials than using all available spectral channels.

Application of wavelet-based RF fingerprinting to enhance wireless network security

This work continues a trend of developments aimed at exploiting the physical layer of the open systems interconnection (OSI) model to enhance wireless network security. The goal is to augment activity occurring across other OSI layers and provide improved safeguards against unauthorized access. Relative to intrusion detection and anti-spoofing, this paper provides details for a proof-of-concept investigation involving “air monitor” applications where physical equipment constraints are not overly restrictive. In this case, RF fingerprinting is emerging as a viable security measure for providing device-specific identification (manufacturer, model, and/or serial number). RF fingerprint features can be extracted from various regions of collected bursts, the detection of which has been extensively researched. Given reliable burst detection, the near-term challenge is to find robust fingerprint features to improve device distinguishability. This is addressed here using wavelet domain (WD) RF fingerprinting based on dual-tree complex wavelet transform (DT-CWT) features extracted from the non-transient preamble response of OFDM-based 802.11a signals. Intra-manufacturer classification performance is evaluated using four like-model Cisco devices with dissimilar serial numbers. WD fingerprinting effectiveness is demonstrated using Fisher-based multiple discriminant analysis (MDA) with maximum likelihood (ML) classification. The effects of varying channel SNR, burst detection error and dissimilar SNRs for MDA/ML training and classification are considered. Relative to time domain (TD) RF fingerprinting, WD fingerprinting with DT-CWT features emerged as the superior alternative for all scenarios at SNRs below 20 dB while achieving performance gains of up to 8 dB at 80% classification accuracy.

Improving Intra-Cellular Security Using Air Monitoring with RF Fingerprints

Improved intra-cellular security is addressed using device-specific RF fingerprints to mitigate malicious network activity that can occur through unauthorized use of digital identities. In air monitoring applications where physical equipment constraints are not overly restrictive, RF fingerprinting remains a viable option for providing regional intra-cellular security for systems such as cellular telephone and last mile WiMax networks. Proof-of-concept results are provided for GSM signals given they are readily available in most areas. Recent RF fingerprinting work has demonstrated average device classification accuracies (serial number identification) of 92% using OFDM-based 802.11a preamble responses at SNR = 6 dB. The goal here was to determine if similar performance could be achieved using RF fingerprints extracted from near-transient and midamble regions of GSM signals. This was done using instantaneous phase responses from each region to form RF statistical fingerprints that are subsequently classified using Fisher-based MDA/ML processing. Considering all GSM device permutations from four different manufacturers, near-transient RF fingerprinting provided nearly 13% improvement in classification performance when compared with midamble RF fingerprinting and achieved average classification performance consistent with the 802.11a benchmark of 92% correct classification at SNR = 6 dB.

Detection of Human Skin in Near Infrared Hyperspectral Imagery

One of the difficulties in search and rescue missions is finding a small target, such as a person, in a large cluttered area. Airborne hyperspectral cameras are now being deployed to aid in this SAR mission. Motivated by the successes of such systems, we define a hyperspectral model of human skin in the visible and near infrared regions of the spectra so we can exploit knowledge gained during the modeling process to aid in human skin detection. Based on observations of the skin model results, an efficient and robust skin detection algorithm using channels in the near infrared region of the spectra is developed. Our algorithm is denoted the Normalized Difference Skin Index, motivated by the Normalized Difference Vegetation Index used in the literature for detecting vegetation in hyperspectral imagery. We demonstrate the capabilities of our skin detection methodology to detect skin amongst objects known to cause false detections for methodologies using three channel color data.

Feature Aided Tracking with Hyperspectral Imagery

Target tracking in an urban environment presents a wealth of ambiguous tracking scenarios that cause a kinematic-only tracker to fail. Partial or full occlusions in areas of tall buildings are particularly problematic as there is often no way to correctly identify the target with only kinematic information. Feature aided tracking attempts to resolve problems with a kinematic-only tracker by extracting features from the data. In the case of panchromatic video, the features are often histograms, the same is true for color video data. In the case where tracks are uniquely different colors, more typical feature aided trackers may perform well. However, a typical urban setting has similar size, shape, and color tracks, and more typical feature aided trackers have no hopes in resolving many of the ambiguities we face. We present a novel feature aided tracking algorithm combining two-sensor modes: panchromatic video data and hyperspectral imagery. The hyperspectral data is used to provide a unique fingerprint for each target of interest where that fingerprint is the set of features used in our feature aided tracker. Results indicate an impressive 19% gain in correct track ID with our hyperspectral feature aided tracker compared to the baseline performance with a kinematic-only tracker.

Human skin detection in the visible and near infrared

Skin detection is a well-studied area in color imagery and is useful in a number of scenarios to include search and rescue and computer vision. Most approaches focus on color imagery due to cost and availability. Many of the visible-based approaches do well at detecting skin (above 90%) but they tend to have relatively high false-alarm rates (8%-15%). This article presents a novel feature space for skin detection in visible and near infrared portions of the electromagnetic spectrum. The features are derived from known spectral absorption of skin constituents to include hemoglobin, melanin, and water as well as scattering properties of the dermis. Fitting a Gaussian mixture to skin and background distributions and using a likelihood ratio test detector, the features presented here show dominating performance when comparing receiver-operating characteristic curves (ROCs) and statistically significant improvement when comparing equal error rates and area under the ROC (AUC). A detection/false-alarm probability of 98.6%/1.1% is achieved for the averaged equal error rate (EER). EER values for the proposed feature space show a 5.6%-11.2% increase in detection probability with a 6.0%-11.6% decrease in false-alarm probability compared to well performing color-based features. The AUC shows a 0.034-0.173 increase in total area under the curve compared to well performing color-based features.

Persistent hyperspectral adaptive multi-modal feature-aided tracking

An architecture and implementation is presented regarding persistent, hyperspectral, adaptive, multi-modal, feature-aided tracking within the urban context. A novel remote-sensing imager has been designed which employs a micro-mirror array at the focal plane for per-pixel adaptation. A suite of end-to-end synthetic experiments have been conducted, which include high-fidelity moving-target urban vignettes, DIRSIG hyperspectral rendering, and full image-chain treatment of the prototype adaptive sensor. Corresponding algorithm development has focused on: motion segmentation, spectral feature modeling, classification, fused kinematic/spectral association, and adaptive sensor feedback/control.

Feature-aided tracking via synthetic hyperspectral imagery

Hyperspectral imaging (HSI) feature-aided tracking (FAT) is an emerging area of research, employing HSI instruments and exploitation techniques with the goal to track moving objects within challenging environments and across frequent ambiguities. A series of studies have been conducted to demonstrate HSI-FAT with contemporary and novel HSI instruments. Synthesized HSI data have been the key enabler to this effort. Capabilities have been evaluated with synthetic models of low-cost, off-the-shelf sensors such as a video-rate liquid crystal tunable filter, as well as sophisticated emerging sensor concepts such as microelectromechanical-adapted systems. A suite of end-to-end synthetic experiments have been conducted, which include high-fidelity moving-target urban vignettes, synthetic hyperspectral rendering, and full image-chain treatment of the various sensor models. Corresponding algorithm development has focused on motion segmentation, spectral feature modeling, classification, fused kinematic/spectral association, and adaptive sensor feedback/ control.

Application of wavelet denoising to improve OFDM‐based signal detection and classification

The developmental emphasis on improving wireless access security through various OSI PHY layer mechanisms continues. This work investigates the exploitation of RF waveform features that are inherently unique to specific devices and that may be used for reliable device classification (manufacturer, model, or serial number). Emission classification is addressed here through detection, location, extraction, and exploitation of RF [fingerprints] to provide device-specific identification. The most critical step in this process is burst detection which occurs prior to fingerprint extraction and classification. Previous variance trajectory (VT) work provided sensitivity analysis for burst detection capability and highlighted the need for more robust processing at lower signal-to-noise ratio (SNR). The work presented here introduces a dual-tree complex wavelet transform (DT-ℂWT) denoising process to augment and improve VT detection capability. The new methods performance is evaluated using the instantaneous amplitude responses of experimentally collected 802.11a OFDM signals at various SNRs. The impact of detection error on signal classification performance is then illustrated using extracted RF fingerprints and multiple discriminant analysis (MDA) with maximum likelihood (ML) classification. Relative to previous approaches, the DT-ℂWT augmented process emerges as a better alternative at lower SNR and yields performance that is 34% closer (on average) to [perfect] burst location estimation performance. Copyright

Extension of the Linear Chromodynamics Model for Spectral Change Detection in the Presence of Residual Spatial Misregistration

A generalized likelihood ratio test (GLRT) statistic for spectral change detection based on the linear chromodynamics model is extended to accommodate unknown residual misregistration between imagery described by a prior probability density function for the spatial misregistration. Using a normal prior distribution leads to a fourth-order polynomial that can be numerically minimized over the unknown misregistration parameters. A more computationally efficient closed-form solution is developed based on a quadratic approximation and provides comparable results to the numerical minimization for the investigated test cases while running 30 times faster. The results applying the method to hyperspectral imagery indicate up to an order of magnitude reduction in false alarms at the same detection rate relative to baseline change detection methods for synthetically misregistered test data particularly in image regions containing edges and fine spatial features. Sensitivity to model parameters is assessed, and the method is compared with a previously published misregistration compensation approach yielding comparable results. Although the GLRT approach appears to exhibit comparable change detection performance, it offers the possibility of tailoring the algorithm to a priori knowledge of expected misregistration errors or to compensate structured misregistration as would occur due to parallax errors due to perspective variations (e.g., image parallax).