Hyperspectral imagery is a highly dimensional type of data resulting in high computational costs during analysis. Band selection aims to reduce the original hyperspectral image to a smaller subset that reduces these costs while preserving the maximum amount of spectral information within the data. This thesis explores various types of band selection techniques used in hyperspectral image processing. Modifying Neural network-based techniques and observing the effects on the band selection process due to the change in network architecture or objective are of particular focus in this thesis. Herein, a generalized neural network-based band selection technique is developed and compared to state-of-the-art algorithms that are applied to a unique dataset and the Pavia City Center dataset where the subsequent selected bands are fed into a classifier to gather comparison results.

Over the past decade, graphene/nanomaterial hybrids have gained a great interest in various applications due to their unique optical properties. This work explores lanthanide doped upconverting nanoparticles (UCNPs)/graphene hybrid nanomaterials. Here, core/shell structures comprising β-NaGdF4:Y b3+(20%),Er3+(2%)@NaGdF4 and α-NaGdF4:Y b3+(20%), Er3+(2%)@NaGdF4 with oleate as capping agent were synthesized and characterized. The choice of lanthanide ions (Yb3+ and Er3+) and their concentrations plays an important role to make these nanoparticles undergo two optical processes (upcoversion and downshifting) capable to convert near-infrared excitation to visible and near-infrared emission. In order to make hybrid systems, these nanoparticles were combined with graphene films. The morphology and the optical behavior of the hybrid samples were studied by microscope and hyperspectral imaging. The multi-energy sublevels from the 4f electronic configuration of lanthanides, their long excited state lifetime and the high carrier mobility of the graphene expected to open an exciting possibility of interaction, however, UCNPs/Graphene hybrid nanomaterial exhibits a minimal response when subjected to 980 nm laser illumination.

<p>Nowadays, soybean is one the most consumed crops in the world. As the human population continuously increases, new phenotyping technology is needed to help plant scientists breed soybean that has high-yield, stress-tolerant, and disease-tolerant traits. Hyperspectral imaging (HSI) is one of the most commonly used technologies for phenotyping. The current HSI techniques include HSI tower and remote sensing on an unmanned aerial vehicle (UAV) or satellite. There are several noise sources the current HSI technologies suffer from such as changes in lighting conditions, leaf angle, and other environmental factors. To reduce the noise on HS images, a new portable, leaf-level, high-resolution HSI device was developed for corn leaves in 2018 called LeafSpec. Due to the previous design requiring a sliding action along the leaf which could damage the leaf if used on a soybean leaf, a new design of the LeafSpec was built to meet the requirements of scanning soybean leaves. The new LeafSpec device protects the leaf between two sheets of glass, and the scanning action is automated by using motors and servos. After the HS images have been collected, the current modeling method for HS images starts by averaging all the plant pixels to one spectrum which causes a loss of information because of the non-uniformity of the leaf. When comparing the two modeling methods, one uses the mean normalized difference vegetation index (NDVI) and the other uses the NDVI heatmap of the entire leaf to predict the nitrogen content of soybean plants. The model that uses NDVI heatmap shows a significant increase in prediction accuracy with an R2 increase from 0.805 to 0.871. Therefore, it can be concluded that the changes occurring within the leaf can be used to train a better prediction model. </p> <p>Although the LeafSpec device can provide high-resolution leaf-level HS images to the researcher for the first time, it suffers from two major drawbacks: intensive labor needed to gather the image data and slow throughput. A new idea is proposed to use a UAV that carries a 6 degree of freedom (DOF) robotic arm with a LeafSpec device as an end-effect to automatically gather soybean leaf HS images. A new UAV is designed and built to carry the large payload weight of the robotic arm and LeafSpec.</p>

Doctor of Philosophy / Department of Food Science / Randall K. Phebus / Intervention techniques to control foodborne pathogens, and rapid identification of pathogens in food are of vital importance to ensure food safety. Therefore, the first objective of this research was to study the efficacy of radio frequency dielectric heating (RFDH) against C. sakazakii and Salmonella spp. in nonfat dry milk (NDM) at 75, 80, 85, or 90°C. Using thermal-death-time (TDT) disks, D-values of C. sakazakii in high heat (HH)- and low heat (LH)-NDM were 24.86 and 23.0 min at 75°C, 13.75 and 7.52 min at 80°C, 8.0 and 6.03 min at 85°C, and 5.57 and 5.37 min at 90°C, respectively. D-values of Salmonella spp. in HH- and LH-NDM were 23.02 and 24.94 min at 75°C, 10.45 and 12.54 min at 80°C, 8.63 and 8.68 min at 85°C, and 5.82 and 4.55 min at 90°C, respectively. The predicted (TDT) and observed (RFDH) destruction of C. sakazakii and Salmonella spp. were in agreement, indicating that the organisms' behavior was similar regardless of the heating system (conventional vs. RFDH). However, RFDH can be used as a faster and more uniform heating method for NDM to achieve the target temperatures. The second objective of this research was to study if hyperspectral imaging can be used for the rapid identification and differentiation of various foodborne pathogens. Four strains of C. sakazakii, 5 strains of Salmonella spp., 8 strains of E. coli, and 1 strain each of L. monocytogenes and S. aureus were used in the study. Principal component analysis and kNN (k-nearest neighbor) were used to develop classification models, which were then validated using a cross-validation technique. Classification accuracy of various strains within genera including C. sakazakii, Salmonella spp. and E. coli, respectively was 100%; except within C. sakazakii, strain BAA-894, and within E. coli, strains O26, O45 and O121 had 66.67% accuracy. When all strains were studied together (irrespective of their genera) for the classification, only C. sakazakii P1, E. coli O104, O111 and O145, S. Montevideo, and L. monocytogenes had 100% classification accuracy; whereas, E. coli O45 and S. Tennessee were not classified (classification accuracy of 0%).

Le phénotypage consiste à caractériser les plantes et leur comportement en vue de la sélection génétique. Cette étude a évalué le potentiel de l'imagerie hyperspectrale de proximité pour répondre à ces besoins. Elle s'appuie sur le lien existant entre la physiologie des plantes et leurs propriétés optiques. Cette étude a montré qu'il est possible de retrouver la réflectance des feuilles en dépit d'un éclairage naturel variable. La procédure de correction mise en place permet de retrouver la réflectance vraie de feuilles à plat et introduit un effet additif (dû à la réflexion spéculaire), un effet multiplicatif (dû au niveau d'éclairement) et un effet non linéaire (dû aux réflexions multiples) sur les feuilles inclinées des plantes au champ. Cependant, nous avons montré également que, grâce à des pré-traitements des spectres adéquats et à la PLS (Partial Least Square regression), la concentration en azote est accessible à partir de la réflectance (400-1000~nm) de feuilles fraîches sur pied. L'étude de spectres simulés a montré que la non prise en compte des réflexions multiples dans l'étalonnage d'un modèle conduisait à une surestimation de la concentration en azote des feuilles subissant des réflexions multiples. Enfin, cette étude a illustré l'intérêt de l'imagerie hyperspectrale de proximité par rapport à la spectrométrie ponctuelle. Le fait d'avoir une image, combiné à la haute résolution spatiale permet d'obtenir des données plus représentatives de la parcelle et de calculer une vitesse de fermeture de couvert. La réalisation de cartographies d'azote permet de suivre la concentration en azote dans différents étages foliaires ou parties d'une même feuille. / Henotyping consists in characterising plants and their behavior with the aim of the genetic selection. This study estimated the potential of the close-range hyperspectral imaging to meet these needs. It leans on the link existing between plant physiology and their optical properties. This study showed that it is possible to find leaf reflectance in spite of a variable natural lighting. The developed correction procedure allows finding the true reflectance of flat leaves and introduces an additive effect (due to specular reflection), a multiplicative effect (due to illumination level) and a not linear effect (due to the multiple reflections) on inclinated leaves of plants in the field. However, we also showed that, thanks to adequate preprocessing of the spectra and to PLS (Partial Least Square regression), the nitrogen concentration is accessible from the reflectance (400-1000~nm) of fresh leaves on standing plants. The study of simulated spectra showed that the not consideration of the multiple reflections in the calibration of a model lead to an overestimation of the nitrogen concentration leaves undergoing multiple reflections. Finally, this study illustrated the interest of close-range hyperspectral imaging with regard to the punctual spectrometry. The fact of having an image, combined with the high spatial resolution allows to obtain more representative data of the plot and to calculate a speed of cover closure. Nitrogen mappings allow following the nitrogen concentration in various leaf level or parts of the same leaf.

Fungal infection and mycotoxin contamination in agricultural products are a serious food safety issue. The detection of fungal infection and mycotoxin contamination in food products should be in a rapid way. A Near-infrared (NIR) hyperspectral imaging system was used to detect fungal infection in 2013 crop year canola, wheat, and barley at different periods after inoculation and different concentration levels of ochratoxin A in wheat and barley. Artificially fungal infected (Fungi: Aspergillus glaucus, Penicillium spp.) kernels of canola, wheat and barley, were subjected to single kernel imaging after 2, 4, 6, 8, and 10 weeks post inoculation in the NIR region from 1000 to 1600 nm at 61 evenly distributed wavelengths at 10 nm intervals. The acquired image data were in the three-dimensional hypercube forms, and these were transformed into two-dimensional data. The two-dimensional data were subjected to principal component analysis to identify significant wavelengths based on the highest principal component factor loadings. Wavelengths 1100, 1130, 1250, and 1300 nm were identified as significant for detection of fungal infection in canola kernels, wavelengths 1280, 1300, and 1350 nm were identified as significant for detection of fungal infection in wheat kernels, and wavelengths 1260, 1310, and 1360 nm were identified as significant for detection of fungal infection in barley kernels. The linear, quadratic and Mahalanobis statistical discriminant classifiers differentiated healthy canola kernels with > 95% and fungal infected canola kernels with > 90% classification accuracy. All the three classifiers discriminated healthy wheat and barley kernels with > 90% and fungal infected wheat and barley kernels with > 80% classification accuracy. The wavelengths 1300, 1350, and 1480 nm were identified as significant for detection of ochratoxin A contaminated wheat kernels, and wavelengths 1310, 1360, 1480 nm were identified as significant for detection of ochratoxin A contaminated barley kernels. All the three statistical classifiers differentiated healthy wheat and barley kernels and ochratoxin A contaminated wheat and barley kernels with a classification accuracy of 100%. The classifiers were able to discriminate between different durations of fungal infections in canola, wheat, and barley kernels with classification accuracy of more than 80% at initial periods (2 weeks) of fungal infection and 100% at the later periods of fungal infection. Different concentration levels of ochratoxin A contamination in wheat and barley kernels were discriminated with a classification accuracy of > 98% at ochratoxin A concentration level of ≤ 72 ppb in wheat kernels and ≤ 140 ppb in barley kernels and with 100% classification accuracy at higher concentration levels. / May 2016

L’imagerie hyperspectrale est maintenant très développée dans les applications de télédétection. Il y a principalement deux manières de construire les imageurs associés : la première méthode utilise un réseau et une fente, et l’image spectrale est acquise ligne par la ligne le long de la trajectoire du porteur. La seconde est basée sur le principe de la spectrométrie par transformée de Fourier (TF). Certains des systèmes utilisés sont construits de manière à enregistrer l’interférogramme de chaque point de la scène suivant le déplacement dans le champ. Le spectre de la lumière venant d’un point de la scène est alors calculé par la transformée de Fourier de son interférogramme. Les imageurs classiques basés sur des réseaux sont plus simples à réaliser et les données qu’ils fournissent sont souvent plus faciles à interpréter. Cependant, les spectro-imageurs par TF fournissent un meilleur rapport signal sur bruit si la source principale de bruit vient du détecteur.Dans la première partie de cette thèse, nous étudions l’influence de différents types de bruit sur les architectures classiques et TF afin d’identifier les conditions dans lesquelles ces dernières présentent un avantage. Nous étudions en particulier l’influence des bruits de détecteur, de photons, des fluctuations de gain et d’offset du détecteur et des propriétés de corrélation spatiale des fluctuations d’intensité du spectre mesuré. Dans la seconde partie, nous présentions la conception, la réalisation et les premiers résultats d’un imageur basé sur un interféromètre de Michelson à dièdres statique nommé DéSIIR (Démonstrateur de Spectro-Imagerie Infrarouge). Les premiers résultats montrent, qu’en mode spectromètre simultané, DéSIIR permet la restitution du spectre avec les spécifications requises dans le cadre des applications recherchées, c'est-à-dire détecter avec une résolution d environ 25 cm-1 un object de quelques degrés plus chaud que le fond de la scène et présentant une signature spectrale entre 3 et 5 juin. En mode spectromètre imageur, après recalage des images, il est possible de reconstruire le spectre de chaque point de la scène observée. / Hyperspectral imaging is now very important in remote sensing applications. There are two main ways to build such imagers : the first one uses a grating and a slit, and the spectral image is acquired line by line along the track of the carrier. The second way is to use the principle of Fourier transform (FT) spectrometry. Some of these systems are built in such a way that they record the interferogram of each point of the scene as it moves through the field of view. The spectrum of the light coming from a particular point is then calculated by the Fourier transform of its interferogram. Classical gratting-based spectral imagers are easier to build and the data they provide a better signal to noise ratio if the main source of noise comes from the detector.In the first part of this thesis, we study the influence of various types of noise on the classic and TF-based architectures to identify the conditions in which these last ones present an advantage. We study particularly the influence of detector noise, photons noise, detector gain and offset fluctuations and spatial correlation properties of the intensity fluctuations. In the second part, we present the conception, the realization and the first results of an imager bases on a Michelson interferometer with dihedrons named DéSIIR (“Démonstrateur de Spectro Imagerie Infrarouge”). The first results show that, in simultaneous spectrometer mode, DéSIIR allows the reconstruction of the spectrum with respect to the specific requirements, which are to be able to detect an objet of some degrees warmer than the background of the scene observed with a resolution of about 25 cm -1. In imager mode, this reconstruction is performed for each point of the scene.

Cette thèse contribue à la recherche de méthodes de détection de signaux inconnus à très faible Rapport Signal-à-Bruit. Ce travail se concentre sur la définition, l’étude et la mise en œuvre de méthodes efficaces capables de discerner entre observations caractérisées seulement par du bruit de celles qui au contraire contiennent l’information d’intérêt supposée parcimonieuse. Dans la partie applicative, la pertinence de ces méthodes est évaluée sur des données hyperspectrales. Dans la première partie de ce travail, les principes à la base des tests statistiques d’hypothèses et un aperçu général sur les représentations parcimonieuses, l’estimation et la détection sont introduits. Dans la deuxième partie du manuscrit deux tests d’hypothèses statistiques sont proposés et étudiés, adaptés à la détection de signaux parcimonieux. Les performances de détection des tests sont comparés à celles de méthodes fréquentistes et Bayésiennes classiques. Conformément aux données tridimensionnelles considérées dans la partie applicative, et pour se rapprocher de scénarios plus réalistes impliquant des systèmes d’acquisition de données, les méthodes de détection proposées sont adaptées de façon à exploiter un modèle plus précis basé sur des dictionnaires qui prennent en compte l’effet d’étalement spatio-spectral de l’information causée par les fonctions d’étalement du point de l’instrument. Les tests sont finalement appliqués à des données astrophysiques massives de type hyperspectral dans le contexte du Multi Unit Spectroscopic Explorer de l’Observatoire Européen Austral. / This thesis deals with the problem of detecting unknown signals at low Signal- to- Noise Ratio. This work focuses on the definition, study and implementation of efficient methods able to discern only-noise observations from those that presumably carry the information of interest in a sparse way. The relevance of these methods is assessed on hyperspectral data as an applicative part. In the first part of this work, the basic principles of statistical hypothesis testing together with a general overview on sparse representations, estimation and detection are introduced. In the second part of the manuscript, two statistical hypotheses tests are proposed and studied. Both are adapted to the detection of sparse signals. The behaviors and the relative differences between the tests are theoretically investigated through a detailed study of their analytical and structural characteristics. The tests’ detection performances are compared with those of classical frequentist and Bayesian methods. According to the three-dimensional data sets considered in the applicative part, and to be closer to realistic scenarios involving data acquisition systems, the proposed detection strategies are then adapted in order to: i) account for spectrally variable noise; ii) exploit the spectral similarities of neighbors pixels in the spatial domain and iii) exploit the greater accuracy brought by dictionary-based models, which take into account the spatiospectral blur of information caused by instrumental Point Spread Functions. The tests are finally applied to massive astrophysical hyperspectral data in the context of the European Southern Observatory’s Multi Unit Spectroscopic Explorer.

Seizures containing large numbers of units of narcotics, goods dangerous to health and doping are often sent to the Swedish National Forensic Centre (NFC). Only a fraction of these capsules or tablets can be analyzed, therefore the samples need to represent the whole seizure. If the samples show content variations, Near Infrared (NIR) spectroscopy in combination with hyperspectral imaging has been shown to be a promising tool to gauge the homogeneity in the seizures based on chemical content. The objective of this thesis was to further develop and then validate a method for the separation of pregabalin and gabapentin capsules using NIR hyperspectral imaging and Principal Component Analysis (PCA). Capsules containing different amounts of pregabalin and gabapentin were prepared and analyzed. Additionally, authentic seizures were analyzed to confirm that the method fulfilled its purpose. The result of this study showed that use of hyperspectral data in the wavelength range 1650-1750 nm gave the best differentiation between pregabalin and gabapentin capsules. Capsules containing the ratio 70-30 % gabapentin and pregabalin could be separated distinctively from capsules containing pure gabapentin. Multiple authentic seizures could be separated into groups correctly depending on the capsules or tablets content.

<p>This dissertation explores computational optical imaging methods to circumvent the physical limitations of classical sensing. An ideal imaging system would maximize resolution in time, spectral bandwidth, three-dimensional object space, and polarization. Practically, increasing any one parameter will correspondingly decrease the others.</p><p>Spectrometers strive to measure the power spectral density of the object scene. Traditional pushbroom spectral imagers acquire high resolution spectral and spatial resolution at the expense of acquisition time. Multiplexed spectral imagers acquire spectral and spatial information at each instant of time. Using a coded aperture and dispersive element, the coded aperture snapshot spectral imagers (CASSI) here described leverage correlations between voxels in the spatial-spectral data cube to compressively sample the power spectral density with minimal loss in spatial-spectral resolution while maintaining high temporal resolution.</p><p>Photography is limited by similar physical constraints. Low f/# systems are required for high spatial resolution to circumvent diffraction limits and allow for more photon transfer to the film plain, but require larger optical volumes and more optical elements. Wide field systems similarly suffer from increasing complexity and optical volume. Incorporating a multi-scale optical system, the f/#, resolving power, optical volume and wide field of view become much less coupled. This system uses a single objective lens that images onto a curved spherical focal plane which is relayed by small micro-optics to discrete focal planes. Using this design methodology allows for gigapixel designs at low f/# that are only a few pounds and smaller than a one-foot hemisphere.</p><p>Computational imaging systems add the necessary step of forward modeling and calibration. Since the mapping from object space to image space is no longer directly readable, post-processing is required to display the required data. The CASSI system uses an undersampled measurement matrix that requires inversion while the multi-scale camera requires image stitching and compositing methods for billions of pixels in the image. Calibration methods and a testbed are demonstrated that were developed specifically for these computational imaging systems.</p> / Dissertation