Options
2017
Doctoral Thesis
Title
A contribution to active infrared laser spectroscopy for remote substance detection
Abstract
Stand-off detection of hazardous substances has been subject to extensiveresearch in the past years and - though several approaches have been proposed- remains to be an ambious goal. Among the suggested methods, laser based measurement techniques have emerged as themost promising. In this work we present a stand-off long-waveinfrared (LWIR) spectroscopysystem for remote detec on of hazardous substances. The principle isbased upon wavelength selec veillumina on using External Cavity Quantum Cascade Lasers (EC-QCLs), that are tunable inthe LWIR wavelengthrange from 7.5mm to 10 mm. This spectral range is part of the molecularfingerprint region for many chemical compounds, including explosive substanceslike PETN, RDX, TNT and AN. In the imaging variant of the hyperspectral data acquisition system, the diffusely backsca ered light is collected by a high performance Mercury CadmiumTelluride (MCT) camera. Using synchronous tuning of the illumina onwavelength, a hyperspectral image iscreated, where each pixel vector comprises the backscaering spectrum of a specific location in the scene. Every pixel vector isregarded as to be a linear superpositon of few contributing spectra (endmembers) with unknown mixture weights. These endmembers are usually the spectra of the spectrally distinct materials in the scene andpossibly one or several target substances ofinterest. This is mathemacally described by the Linear Mixture Model(LMM), on which many exisisting hyperspectral target detection algorithms are based on. In this work, we present details of the hyperspectral imaging sensor that touch both optimization of illuminationas well as data acquisiquion andanalysis.We analyze a variety of well known target detection algorithms for the task of substance detection in the acquired hyperspectral image data. The considered target detection algorithms include fast and intuitive full-pixel detection algorithms like the NCC, Matched Filter (MF) and CEM, as well as well known and powerfulsub-pixel detection algorithms like the ACE and AMSD. The latter utilizes the structured LMM variant, that requires explicit knowledgeof the contributing background materials in the scene. As these are in general unknown the corresponding endmembers have to be extracted from the available hyperspectral observation data using an endmember extraction algorithm. In this work, we present an endmember extraction algorithm, tailored to the boundary conditions, induced by the measurement technique. An additional parameter that is required for data analysis is the number of spectrally distinct materials in the scene, that essentially comprises the required model order. We present a model order estimation method based on the Minimum Description Length (MDL) principle by adaption of a method proposed by Wax and Kailath [63] for estimation of the number of signal sources in complex radar observations. Whereas the method is developed in the context of hyperspectral image analysis, it can be readily applied to signal source estimation in any real-valued multi-bansed observations that comprises a linear superposition of independent sources. Whereas the proposed MDL model order estimation method competes well or even outperforms comparable model order estimation algorithms like the NSP and SML on artificial data that was explicitly generated using the LMM, it significantly overesestimates the required model order when applied to real-world measurement data. Based on observations of the corresponding covariance matrix eigenvalue-distributions, this behavior is attributed to correlated noise, most likely due to remaining speckle. The latter is caused by the coherent nature of the illumination source and significantly reduced- though not fully suppressed - using a multi-look approach. The proposed background endmember extraction algorithm shows however, to be robust against model-overestimation. Based on the estimate of the background spectra, the AMSD algorithm is successfully applied for detection of all of the mentioned explosive substance residues on various substrates. This is demonstrated, both for hyperspectral image measurement results obtained by a short-range variant of the hyperspectral image sensor over 1.4 m, as well as with an extended range setup, operational up to distances of ≈ 20 m. Finally, we present an extension of the backscattering spectroscopy method to a real-time measurement device, based on a rapid wave length scanning EC-QCL. The latter swipes the full spectral emission range of the QCL chip within 1 ms which enables a spectral acquisition rate of 1 kHz. The diffusely backscattered light is in this measurement setup collected by a single-element MCT detector generating single spectra, rather than full hyperspectral images. We present an experimental setup capable of fast data acquisition and show, that the developed target detection algorithms are capable of real-time detection in the observed hyperspectral data.
Thesis Note
Zugl.: Karlsruhe, Inst. für Technologie (KIT), Diss., 2017
Author(s)
File(s)
Rights
CC BY-SA 4.0: Creative Commons Attribution-ShareAlike
Language
English