Application of chemometrics to hyperspectral imaging analysis of environmental and agricultural samples

Abstract

This Thesis deals with the resolution of hyperspectral imaging data by using chemometric methods, in particular by using appropriate data pretreatment methods and by using Multivariate Curve Resolution (MCR) methods. The main contribution of the present Thesis is the study and implementation of the MCR-ALS (Multivariate Curve Resolution Alternating Least Squares) method for the resolution of hyperspectral images, collected by remote sensing (airborne or space borne Earth observation instrument) and by micro-spectroscopy imaging. Specifically, in this Thesis work, we explore the combination of chemometric and hyperspectral imaging methods for the resolution of spectra (signatures) and spatial distribution maps of the chemical constituents of a sample. The ultimate goal of this study is to improve the analysis and interpretation of hyperspectral imaging data by taking advantage of different chemometric powerful tools. Local rank/selectivity properties describing the spatial information of spectroscopic images can be used as a constraint to increase the performance of MCR methods significantly, decreasing rotation ambiguity uncertainties. Different multivariate resolution methods were compared, such as MCR-ALS, Principal Component Analysis (PCA), and Minimum Volume Simplex Analysis (MVSA), Multivariate Curve Resolution-Function Minimization (MCR-FMIN), MCR-BANDS and FAC-PACK. All these approaches have been used for the evaluation of the extension of rotation ambiguities remaining in the results after their application. Several hyperspectral images provided by standard and widely used instruments such as NASA’s Airborne Visible Infra-Red Imaging Spectrometer (AVIRIS), Raman hyperspectral imaging Spectrometer, and Infrared hyperspectral imaging Spectrometer have been used as example of data sets to test the different methods, in particular to test the MCR-ALS method. The results obtained in this Thesis show that MCR-ALS method can be successfully used for hyperspectral image resolution purposes. The spectra signatures of the pure constituents present in hyperspectral images and their concentration distribution at a pixel level can be estimated. Constituents identification can be performed using the resolved pure spectra signatures and comparing them to reference spectra from spectral libraries or from experimental spectra of reference samples. Application of image data pretreatment methods reduce significantly the presence of strong fluorescence background in Raman hyperspectral images. In contrast, infrared hyperspectral imaging is not affected by fluorescence. Kramers-Kronig transform enables to calculate absorption spectra in case only reflectance spectra can be measured for infrared spectra. The extent of rotation ambiguity associated to MCR-ALS and other resolution methods can be rather high when they are applied for hyperspectral image resolution with high noise. The correct resolution of hyperspectral images can only be guaranteed if additional constraints are applied, such as those providing information about the local rank properties of the image, i.e. about the presence or absence of the different constituents (components) in the image pixels. Only in this way it is possible to increase the reliability of the solutions provided by MCR methods and decrease the uncertainties associated to them. Appropriate use of local rank and selectivity constraints can improve significantly the quality of the pure spectra (signatures) and of the constituent distribution maps resolved by MCR-ALS analysis of hyperspectral images in remote sensing studies. Use of correlation coefficients between selected spectra and image pixel spectra is shown to provide an alternative way for the application of the selectivity constraint in hyperspectral images for the first time. This alternative method resulted to be satisfactory when pure pixels exist. MCR-BANDS method can be used to get estimations of the extension of rotation ambiguities in MCR resolved results. The Area of Feasible Solutions represents feasible solutions geometrically. The range of rotation ambiguity calculated by MCR-BANDS and AFS are in agreement. MCR-ALS with the trilinearity constraint is an effective way to characterize and resolve Excitation-Emission Matrix fluorescence spectra (EEM).

Esta Tesis trata de la resolución de datos de imágenes hiperespectrales utilizando métodos quimiométricos, en particular mediante el uso de métodos de pretratamiento de datos y utilizando métodos de resolución multivariante de curvas (MCR). La principal contribución de la presente Tesis es el estudio y la aplicación del método MCR-ALS (resolución multivariante de curvas mediante mínimos cuadrados alternados) para la resolución de imágenes hiperespectrales, adquiridas mediante técnicas de teledetección y mediante técnicas de micro-espectroscopia. Específicamente, en el trabajo de esta Tesis, se explora la combinación de los métodos quimiométricos y de los métodos de análisis de imágenes hiperespectrales, para la resolución de los espectros (firmas) y de los mapas de distribución de los componentes químicos de la muestra. El objetivo final de este estudio es mejorar el análisis y la interpretación de los datos de imágenes hiperespectrales mediante el aprovechamiento de diferentes herramientas quimiométricas poderosas. La detección del rango local y las propiedades de selectividad que describen la información espacial de los componentes presentes en las imágenes espectroscópicas. Se han comparado diferentes métodos de resolución, tales como MCR-ALS, MVSA (Mínimo Volumen Simplex Análisis), PCA (Análisis de Componentes Principales), y MCR-FMIN. Los métodos MCR-BANDS y FAC-PACK se han utilizado para la evaluación de la extensión de las ambigüedades rotacionales existentes en los resultados después de la aplicación de estos métodos de resolución multivariante. En esta Tesis se han analizado diversos conjuntos de datos compuestos por varias imágenes hiperespectrales proporcionadas por instrumentos estándar tales como el espectrómetro de imágenes hiperespectrales en el visible y en el infrarrojo AVIRIS de la NASA, y diversos espectrómetros de imágenes hiperespectrales Raman y infrarrojo de laboratorio. La eficacia del procedimiento MCR-ALS se ilustra proporcionando comparaciones exhaustivas con otros métodos de resolución de mezclas espectrales a partir de conjuntos de datos hiperespectrales simulados y reales.