Publication: Teledetección de gases mediante imagen hiperespectral por transformada de Fourier en el infrarrojo medio. Una contribución al problema de la cuantificación y la reducción de la dimensionalidad
Loading...
Identifiers
Publication date
2016
Defense date
2016-10-28
Authors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
A lo largo de las últimas dos décadas, el desarrollo de sistemas de imagen infrarroja ha experimentado una notable evolución. Esta tecnología, aunque originaria del ámbito militar y de la teledetección, actualmente supone un claro ejemplo de transversalidad tras haberse trasladado con éxito a sectores como el industrial, el medioambiental, o la imagen médica, conforme su nivel de competitividad y prestaciones iban en aumento.
Nuevas aplicaciones han demandado dispositivos cada vez más complejos, dotados de mecanismos de selección espectral cada vez más sofisticados, y dando lugar a sistemas de imagen con capacidad para un número reducido de canales. Recientemente, este progreso culmina con los sistemas de imagen hiperespectral, que aúnan las posibilidades que brinda la resolución espacial con una resolución espectral equiparable a la de los espectrorradiómetros clásicos; lo que justifica el nombre de espectrorradiómetro de imagen aplicado también a estos equipos.
El potencial ofrecido por estos últimos desarrollos permite afrontar bajo nuevos enfoques desafíos tradicionalmente complejos, como es el caso de la teledetección de compuestos gaseosos. Muchas aplicaciones de carácter tanto civil (como la monitorización de fuentes de contaminantes, la detección de emisiones fugitivas o el control de procesos industriales) como militar (desde la detección e identificación de agentes químicos a labores de reconocimiento de amenazas) se beneficiarán de los logros alcanzados en la resolución de este problema. En particular, una de las tareas más exigentes es la cuantificación, esto es, la estimación de parámetros físicos relevantes, como la concentración de una determinada especie química o la temperatura a la que se encuentra. Esto se debe fundamentalmente a la baja determinación de los datos observados, donde hay fuertes relaciones de acoplamiento entre magnitudes, lo que obliga al uso de complejos algoritmos para su recuperación.
Esta tesis doctoral aborda el problema de la cuantificación a distancia de gases mediante un espectrorradiómetro de imagen provisto de un interferómetro de Michelson, que le confiere una resolución espectral de 0,25 cm -¹, acoplado a un array de 320 x 256 detectores de InSb con sensibilidad en la región comprendida entre 1,5 y 5,5 μm. Nuestro estudio comienza con el análisis de una aplicación preliminar, la teledetección activa de gases mediante sensores de banda integrada, que nos sirve para introducir herramientas como el modelado de la transferencia radiativa o la simulación de prestaciones radiométricas. El trabajo continúa con el desarrollo de las funciones de procesado que permiten obtener espectros a partir de interferogramas en el sistema de imagen hiperespectral utilizado. A continuación, se propone una metodología que permite la recuperación de la densidad de columna (producto de la concentración y el camino óptico) y de la temperatura mediante un algoritmo de ajuste iterativo, resolviendo la relación de ambigüedad que estas variables mantienen, y que finalmente es validada de forma experimental a partir de espectros de transmitancia.
Nuestro enfoque del problema de la cuantificación trabaja con un espectro de alta resolución para cada píxel de la imagen. Esto supone un volumen de datos sumamente grande, y consecuentemente, unos tiempos de cómputo muy elevados, lo que nos ha llevado a la búsqueda de procedimientos que permitan extraer la dimensionalidad intrínseca de la información contenida en dichos datos. La solución propuesta hace uso del análisis de componentes principales como técnica para generar sub-espacios adaptados a los espectros medidos, aunque con la particularidad de que dichas componentes se construyen a partir de datos teóricos que simulan el escenario observado. De esta forma se consigue no sólo una notable reducción de la dimensionalidad y la multicolinealidad, sino que ésta viene acompañada por una reducción del ruido, y se facilita la recuperación de las variables densidad de columna y temperatura. Esta nueva estrategia de estimación es validada a partir del caso experimental estudiado inicialmente, y se complementa su uso junto a distintas estrategias de generalización, que buscan optimizar la recuperación de datos explorando diferentes alternativas de aproximación de funciones (como el perceptrón multicapa) y de regresión no lineal (mediante redes neuronales de base radial y máquinas de vectores de soporte). Otro de los factores responsables del alto volumen de datos presente en los interferogramas es el sobremuestreo con el que son adquiridos, por lo que aquí se propone un algoritmo que consigue un uso eficiente de las muestras redundantes para reducir el nivel de ruido.
Por último, esta tesis concluye poniendo a prueba los planteamientos expuestos hasta aquí para afrontar un problema metrológico actual, como es la medida de las altas temperaturas (>2400 K) que se producen en determinados procesos de combustión, en el contexto de un proyecto europeo de investigación. Nuestra aportación se basa en el uso de técnicas de espectroscopia de emisión, junto al de una metodología de recuperación de la temperatura basada en el método de ajuste iterativo ya desarrollado, obteniéndose como sub-productos estimaciones de la densidad de columna de las principales especies.
No obstante, este proceso no resulta libre de contratiempos, principalmente por la dificultad que implica observar fenómenos dinámicos con un interferómetro de Michelson, en el que la radiación entrante fluctúa temporalmente durante el recorrido del espejo. Esto provoca una superposición entre la modulación propia del interferómetro y la modulación temporal, que ha sido necesario corregir para evitar la aparición de artefactos en los espectros obtenidos. Finalmente, el análisis de las medidas realizadas mediante el método de Monte Carlo arroja niveles de incertidumbre en la estimación de la temperatura del orden de ± 5K.
During the last two decades, the development of infrared imaging systems has experienced a remarkable evolution. This technology, although primarily originated in military and remote sensing fields, currently has become an example of transversality once it has been succesfully transferred to other sectors, like industry, environment, or healthcare, while both its competitiveness and performance followed a growing trend. New applications have required new and more complex devices, including mechanisms for spectral selection that become increasingly more sophisticated, and give rise to imaging systems able to sense along a reduced number of channels. Recently, this progress has culminated with the development of the so-called hyperspectral imaging systems, equipped with a spectral resolution comparable to that provided by classic spectroradiometers, whilst maintaining all the possibilities that spatial resolution grants. That is the reason why they are often referred to as imaging spectroradiometers. The potential offered by these developments allows to address old challenges through new perspectives, as it is the case of remote sensing of gases. Many applications both civil (as the case of pollutants emissions monitoring, fugitive emissions detection or industrial processes) and military (from detection and identification of chemical agents to threat recognition) will benefit from advances performed in this topic. In particular, one of the most demanding tasks is quantification, that is, the estimation of relevant physical parameters, as the concentration of a chemical compound or its temperature. This is mainly due to the fact that the targeted data are poorly determined, because of coupling relations between magnitudes, and it motivates the use of complex algorithms in the retrieval procedures. This dissertation faces the problem of gas remote quantification through the use of an imaging Fourier transform spectrometer (IFTS), fitted with a Michelson interferometer that provides a maximum spectral resolution of 0,25 cm-¹, and coupled to a 320 x 256 InSb array of detectors operating in the mid-infrared range, from 1,5 to 5,5 μm. Our analysis begins with a preliminar application, the active remote sensing of gases based on single band infrared imaging detectors. This stage will helps us to introduce some important tools, like modelling of radiative transfer functions, and the ability to perform radiometric simulations. Our work continues with the setup of our hyperspectral imaging system, and the development of processing functions that allow to obtain spectra from interferograms. Then, a full methodology has been proposed for both column density (the product of concentration and optical path) and temperature retrieval, through an iterative fitting algorithm that solves the ambiguity between these variables. Finally, it has been experimentally validated making use of transmittance spectra. Our approach to the quantification problem requires of high resolution spectra for each pixel in the image. This implies a huge volume of data, and consequently, very large computing times. This has lead us to the search of procedures for extracting the intrinsic dimensionality of the information contained in data. Our solution is based on the use of principal component analysis to build signal sub-spaces adapted to measured spectra, but with the distinctive feature that they are generated from theoretical datasets that simulate the target scene. Therefore, not only a notable dimensionality and multicolinearity reduction is achieved, but also noise levels are decreased, making easier to retrieve column density and temperature. This methodology is also validated for the experimental setup previously considered, and it is jointly applied with different generalization strategies, that aim to optimize data retrieval exploring different approaches like function approximation (with a multilayer perceptron) and non linear regression techniques (through radial basis neural networks and support vector machines). Another important factor responsible for the great volume of interferograms is oversampling. In that case, an algorithm is proposed to squeeze the redundant samples, so that they are rearranged to reduce noise and promote a more efficient use. This thesis concludes testing all the exposed work by dealing with a metrology topic of current interest: the accurate measurement of high temperatures (>2400 K) in combustion processes, in the context of a European research project. Our approach is based on the use of emission spectroscopy techniques along with a temperature retrieval methodology, that also provides column density estimations of major species. However, this is not a straightforward procedure, mainly due to the difficulties that entails to observe dynamic phenomenons with a Michelson interferometer, in which the collected radiation temporarily fluctuates during mirror’s sweep. This causes a superposition between both interferometer modulation and external modulation, that has been corrected to avoid artifacts in the associated spectra. Finally, the Monte Carlo method applied to measurements reveals an uncertainty in the estimation of temperature of about ±5K.
During the last two decades, the development of infrared imaging systems has experienced a remarkable evolution. This technology, although primarily originated in military and remote sensing fields, currently has become an example of transversality once it has been succesfully transferred to other sectors, like industry, environment, or healthcare, while both its competitiveness and performance followed a growing trend. New applications have required new and more complex devices, including mechanisms for spectral selection that become increasingly more sophisticated, and give rise to imaging systems able to sense along a reduced number of channels. Recently, this progress has culminated with the development of the so-called hyperspectral imaging systems, equipped with a spectral resolution comparable to that provided by classic spectroradiometers, whilst maintaining all the possibilities that spatial resolution grants. That is the reason why they are often referred to as imaging spectroradiometers. The potential offered by these developments allows to address old challenges through new perspectives, as it is the case of remote sensing of gases. Many applications both civil (as the case of pollutants emissions monitoring, fugitive emissions detection or industrial processes) and military (from detection and identification of chemical agents to threat recognition) will benefit from advances performed in this topic. In particular, one of the most demanding tasks is quantification, that is, the estimation of relevant physical parameters, as the concentration of a chemical compound or its temperature. This is mainly due to the fact that the targeted data are poorly determined, because of coupling relations between magnitudes, and it motivates the use of complex algorithms in the retrieval procedures. This dissertation faces the problem of gas remote quantification through the use of an imaging Fourier transform spectrometer (IFTS), fitted with a Michelson interferometer that provides a maximum spectral resolution of 0,25 cm-¹, and coupled to a 320 x 256 InSb array of detectors operating in the mid-infrared range, from 1,5 to 5,5 μm. Our analysis begins with a preliminar application, the active remote sensing of gases based on single band infrared imaging detectors. This stage will helps us to introduce some important tools, like modelling of radiative transfer functions, and the ability to perform radiometric simulations. Our work continues with the setup of our hyperspectral imaging system, and the development of processing functions that allow to obtain spectra from interferograms. Then, a full methodology has been proposed for both column density (the product of concentration and optical path) and temperature retrieval, through an iterative fitting algorithm that solves the ambiguity between these variables. Finally, it has been experimentally validated making use of transmittance spectra. Our approach to the quantification problem requires of high resolution spectra for each pixel in the image. This implies a huge volume of data, and consequently, very large computing times. This has lead us to the search of procedures for extracting the intrinsic dimensionality of the information contained in data. Our solution is based on the use of principal component analysis to build signal sub-spaces adapted to measured spectra, but with the distinctive feature that they are generated from theoretical datasets that simulate the target scene. Therefore, not only a notable dimensionality and multicolinearity reduction is achieved, but also noise levels are decreased, making easier to retrieve column density and temperature. This methodology is also validated for the experimental setup previously considered, and it is jointly applied with different generalization strategies, that aim to optimize data retrieval exploring different approaches like function approximation (with a multilayer perceptron) and non linear regression techniques (through radial basis neural networks and support vector machines). Another important factor responsible for the great volume of interferograms is oversampling. In that case, an algorithm is proposed to squeeze the redundant samples, so that they are rearranged to reduce noise and promote a more efficient use. This thesis concludes testing all the exposed work by dealing with a metrology topic of current interest: the accurate measurement of high temperatures (>2400 K) in combustion processes, in the context of a European research project. Our approach is based on the use of emission spectroscopy techniques along with a temperature retrieval methodology, that also provides column density estimations of major species. However, this is not a straightforward procedure, mainly due to the difficulties that entails to observe dynamic phenomenons with a Michelson interferometer, in which the collected radiation temporarily fluctuates during mirror’s sweep. This causes a superposition between both interferometer modulation and external modulation, that has been corrected to avoid artifacts in the associated spectra. Finally, the Monte Carlo method applied to measurements reveals an uncertainty in the estimation of temperature of about ±5K.
Description
Keywords
Teledetección de gases, Imagen hiperespectral, Transformadas de Fourier