To obtain intrinsic reflectance of the scene by hyperspectral imaging systems has been a scientific and engineering challenge. Factors such as illumination variations, atmospheric effects and viewing geometries are common artefacts which modulate the way of light reflections from the object into the sensor and that they are needed to be corrected. Some of these factors induce highly scattered and diffuse irradiance which can artificially modify the intrinsic spectral reflectance of the surface, such as that in the shadows. This research is attempted to compensate the shadows in the hyperspectral imagery. In this study, three methods known as the Diffuse Irradiance Compensation (DIC), Linear Regression (LR) and spectro-polarimetry technique (SP) have been proposed to compensate the shadow effect. These methods have various degrees of shadow compensation capabilities, and their pros and cons are elucidated within the context of their classification performances over several data sets recorded within and outside of the laboratory. The spectro-polarimetry (SP) technique has been found to be the most versatile and powerful method for shadow compensation, and it has achieved over 90% of classification accuracy for the scenes with ~30% of shadow areas.

To obtain intrinsic reflectance of the scene by hyperspectral imaging systems has been a scientific and engineering challenge. Factors such as illumination variations, atmospheric effects and viewing geometries are common artefacts which modulate the way of light reflections from the object into the sensor and that they are needed to be corrected. Some of these factors induce highly scattered and diffuse irradiance which can artificially modify the intrinsic spectral reflectance of the surface, such as that in the shadows. This research is attempted to compensate the shadows in the hyperspectral imagery. In this study, three methods known as the Diffuse Irradiance Compensation (DIC), Linear Regression (LR) and spectro-polarimetry technique (SP) have been proposed to compensate the shadow effect. These methods have various degrees of shadow compensation capabilities, and their pros and cons are elucidated within the context of their classification performances over several data sets recorded within and outside of the laboratory. The spectro-polarimetry (SP) technique has been found to be the most versatile and powerful method for shadow compensation, and it has achieved over 90% of classification accuracy for the scenes with ~30% of shadow areas.

Dans ce travail de thèse, nous développons des protocoles d'imagerie optique pour l'observation des nanoparticules sur des coupes de tissus afin de relier leur localisation et leur «comportement» à l'environnement biologique, en particulier son éventuel état pathologique. Nous avons synthétisé des agents de contraste bimodaux, sous forme de nanoparticules -NP- visibles en résonance magnétique et en optique, à base d'oxydes de fer et de zinc (Zn(Fe)O) avec une nouvelle méthode de polyol azéotropique dans des solvants glycoliques (DEG et PG). L'élimination de l'eau à l'aide de l'appareil Dean-Stark est une nouvelle stratégie pour la synthèse de NP dans une solution de polyol, avec un rendement élevé et produisant des particules de petite taille. Les NP les plus visibles, selon leur contraste IRM, ont été revêtus de carboxyméthyl pullulane, de polyéthylène glycol, de carboxyméthyl dextrane et de fucoïdane, ce dernier étant un polysaccharide capable de se lier spécifiquement à la paroi vasculaire. Les NPs montrent de bonnes propriétés magnétiques et optiques à température ambiante. Les NP recouvertes ont été injectées dans un modèle de rat d’athérothrombose pour localiser le thrombus par IRM avant sacrifice et collecte des tissus pour étude des coupes histologiques par microscopie optique. La différence entre les images IRM avant et après l'injection de fucoïdane-NPs et de CMD-NPs est claire. Les résultats montrent que les NP recouvertes de fucoïdane sont liées au thrombus. Certains types de microscopies, tels que la microscopie de fluorescence, la microscopie en champ sombre, la microscopie hyperspectrale à champ sombre et la microscopie interférentielle à champ sombre ont été développés pour la détection des NPs en milieu liquide et dans les tissus. En analysant le spectre de chaque pixel et en le comparant au spectre des matériaux de référence, la microscopie hyperspectrale peut détecter la présence de NPs sur des coupes de tissus, les localiser, les identifier et les caractériser. Zn(Fe)O NPs constituerait donc un agent de contraste bimodal potentiel pour l’IRM et l’imagerie optique. Cependant, bien que de nombreux outils optiques avancés aient été développés, nous avons constaté qu'il est toujours difficile d'identifier de manière fiable les NP dans le tissu. / In this thesis work, we develop optical imaging protocols for the observation of then anoparticles on tissue slices in order to further link their localization and their “behaviour” to the biological pathological environment. Bimodal zinc and iron oxide-based MRI/optical nanoparticle contrast agents (Zn(Fe)O) have been synthesised with a novel azeotropicpolyol method in glycol solvents (DEG and PG). The most potent NPs, as regard to their MR contrast power, have been coated with carboxymethyl pullulan, polyethylene glycol,carboxymethyl dextran (CMD) and fucoidan, the latter being a polysaccharide able to specifically bind to the vascular wall. The coated NPs were injected into rat to locate atherothrombosis by MRI. Then the histological slices of harvested diseased tissue were imaged with our homemade optical microscope. Water removal using Dean-Stark apparatus is a novel strategy for the synthesis of NPs in polyol solution with high yield and small size.The NPs show the good magnetic and optical properties at room temperature. The coated nanoparticles were injected into an atherothrombotic rat model to locate the thrombus by MRI prior to sacrifice of the animals and tissue collection for histological study by optical microscopy. The difference of MRI images between before and after injection with Fucoidan-NPs and CMD-NPs is clear. The results indicated that fucoidan-NPs are linked to the thrombus. Some type of microscopies, such as fluorescent microscopy, dark field microscopy, hyperspectral dark field microscopy and interference dark field microscopy have been developed for the detection of NPs in liquid medium and in the histological tissue. By analyzing the spectrum of every pixel and comparing to the spectrum of reference materials, hyperspectral microscopy can detect the presence of nanomaterial on exposed tissue slices, locate, identify, and characterize them. Zn(Fe)O NPs would therefore constitute a potential bimodal contrast agent for MRI and optical imaging. Although many advance optical tools have been developed, but we found it is still a challenge to identify reliably the NPs in the tissue.

Hyperspectral image data has long been an important tool for many areas of sci- ence. The addition of spectral data yields significant improvements in areas such as object and image classification, chemical and mineral composition detection, and astronomy. Traditional capture methods for hyperspectral data often require each wavelength to be captured individually, or by sacrificing spatial resolution. Recently there have been significant improvements in snapshot hyperspectral captures using, in particular, compressed sensing methods. As we move to a compressed sensing image formation model the need for strong image priors to shape our reconstruction, as well as sparse basis become more important. Here we compare several several methods for representing hyperspectral images including learned three dimensional dictionaries, sparse convolutional coding, and decomposable nonlocal tensor dictionaries. Addi- tionally, we further explore their parameter space to identify which parameters provide the most faithful and sparse representations.

I DEVELOPED A NEW HIGH THROUGHPUT DUAL CHANNEL HYPERSPECTRAL IMAGING CONFIGURATION BASED ON ACOUSTO-OPTIC TUNABLE FILTER. THE DEVELOPED IMAGING SYSTEM WAS CHARACTERIZED AND EVALUATED IN COMPARISON WITH OTHER CONVENTIONAL CONFIGURATIONS. THE NEW IMAGING SYSTEM PROVED HIGHER THROUGHPUT WITH RESPECT TO THE CURRENTLY USED CONFIGURATIONS.THE IMAGING SYSTEM WAS THEN USED TO QUANTITATIVELY ASSESS AND PRECISELY CLASSIFY SKIN ERYTHEMA INDUCED ARTIFICIALLY ON VOLUNTEERS AND NATURALLY ON SKIN CANCER PATIENTS DUE TO RADIOTHERAPY TREATMENT. / Recent cancer statistics show that 40% of Canadians might contract cancer during their life and 25% of Canadians might die due to cancer. In skin, head and neck cancers, surgery and radiation therapies are the most prevalent treatment options, while radiation therapy is the most commonly used approach. A common problem in radiation therapy is tumors behave differently against ionizing radiation. For instance, with the same dose, some tumors are fully damaged or shrunk, while others are less affected. The difference in individual tumor response to therapy is transformed into a research question: how to quantitatively assess tumor response to radiation and how to tune radiation therapy to achieve full destruction for tumor cells? Few past studies addressed the question, although no definite answer was realized. This work is a part of a project that investigates the hypothesis that radiation response of skin is correlated to individual tumor response. In the case of high correlation, the skin’s faster response to ionizing radiation can be used to modify the irradiation dose to achieve the maximum destruction of individual’s tumor. To examine the project hypothesis, radiation-induced skin redness or erythema was selected as an acute skin reaction to being objectively quantified. Hence, the overall goal of the research thesis work is to objectively assess and precisely quantify radiation-induced erythema or radiation dermatitis. Skin erythema was assessed formerly by multiple optical and non-optical modalities. The current gold standard is the visual assessment (VA). Unfortunately, VA lacks objectiveness, precise communication, and quantification. To push the limitations of VA and past techniques, hyperspectral imaging (HSI) was proposed to be used for erythema assessment. The work detailed in this thesis aims to create more confidence in HSI to be utilized toward objectively quantify skin erythema. To reach this goal, initially, a new high-throughput dual channel acousto-optic tunable filter (AOTF)-based-HSI instrument was developed for monitoring radiation dermatitis. AOTF-HSI instrument design, implementation, and full characterization are presented. Second, the developed AOTF-HSI instrument is evaluated against a liquid crystal tunable filter (LCTF) instrument. Third, to be prepared for clinical operation, the AOTF-HSI equipment was used to classify an artificially-induced erythema on healthy volunteers in an exploratory study. A robust linear discriminant analysis (LDA)-based classification method was developed for the purpose of image classification. Finally, HSI instrument and LDA classification method were utilized in a preliminary clinical study to properly monitor and precisely quantify radiation dermatitis for skin cancer patients. In the clinical study, erythema indices were computed using Dawson’s method. Least square fitting was used to fit the acquired absorbance data, and thus quantify the hemoglobin concentration change along the study duration. Moreover, LDA was used to contrast spectral and digital imaging for erythema classification. In sum, the work documented in this thesis was willfully directed to achieve an efficient, portable, user-friendly hyperspectral imaging system which has the opportunity to be a benchtop in the clinical daily procedure in the near future. / Thesis / Doctor of Philosophy (PhD)

This dissertation develops new techniques to reduce computational complexity for hyperspectral remote sensing image analysis. Specific techniques are applied with regards to different applications of hyperspectral imagery, i.e., classification, target detection. The contribution of this dissertation can be summarized as follows. 1. A time efficient version combining multiple collaborative representations model is proposed for hyperspectral image classification. Collaborative representation (CR) can be implemented either with a dictionary containing training samples of all-classes or class-specific. A collaborative representation optimized classifier with Tikhonov regularization (CROCT) is proposed to avoid the redundant operations in all-class and class-specific versions. 2. An efficient probabilistic collaborative representation is presented for hyperspectral image classification. Its performance is evaluated on different types of spatial features of hyperspectral imagery including shape feature (i.e., extended multi-attribute feature), global feature (i.e., Gabor feature), and local feature (i.e., Local Binary Pattern). Experimental results show the probabilistic collaborative representation based classifier (PROCRC) has excellent performance in terms of both accuracy and computational cost compared with the original CRC and regularized versions of CRC. 3. Fast nonlinear classification and an explicit kernel approach are built for multispectral and hyperspectral imagery respectively to improve the kernel version of collaborative representation based algorithms. Experimental results show that using artificial bands generated from a simple band ratio function can yield better classification accuracy than the nonlinear kernel method and also reduce computational cost. In addition, the explicit kernel mapping approach can yield high accuracy as the original kernel versions of CR-based algorithms but with similarly low computational cost as in the original linear CRC classifiers. 4. Efficient band selection approaches are proposed for hyperspectral target detection. A maximum-sub-maximum ratio (MSR) metric has been applied for band selection, which can well gauge the target background separation. Efficient evolutionary searching methods such as particle swarm optimization and firefly algorithm are used in conjunction with maximum-sub-maximum ratio metric for band selection. Experimental results show that the proposed band selection approach can select a small band set while yielding similar detection performance compared with using all the original bands.

Developments in sensor technology have made high resolution hyperspectral remote sensing data available to the remote sensing analyst for ground cover classification and target recognition tasks. Further, with limited ground-truth data in many real-life operating scenarios, such hyperspectral classification systems often employ dimensionality reduction algorithms. In this thesis, the efficacy of spectral derivative features for hyperspectral analysis is studied. These studies are conducted within the context of both single and multiple classifier systems. Finally, a modification of existing classification techniques is proposed and tested on spectral reflectance and derivative features that adapts the classification systems to the characteristics of the dataset under consideration. Experimental results are reported with handheld, airborne and spaceborne hyperspectral data. Efficacy of the proposed approaches (using spectral derivatives and single or multiple classifiers) as quantified by the overall classification accuracy (expressed in percentage), is significantly greater than that of these systems when exploiting only reflectance information.

No description available.

Processing hyperspectral image data can be computationally expensive and difficult to employ for real-time applications due to its extensive spatial and spectral information. Further, applications in which computational resources may be limited can be hindered by the volume of data that is common with airborne hyperspectral image data. This paper proposes utilizing band selection to down-select the number of spectral bands to consider for a given classification task such that classification can be done at the edge. Specifically, we consider the following state of the art band selection techniques: Fast Volume-Gradient-based Band Selection (VGBS), Improved Sparse Subspace Clustering (ISSC), Maximum-Variance Principal Component Analysis (MVPCA), and Normalized Cut Optimal Clustering MVPCA (NC-OC-MVPCA), to investigate their feasibility at identifying discriminative bands such that classification performance is not drastically hindered. This would greatly benefit applications where time-sensitive solutions are needed to ensure optimal outcomes. In this research, an NVIDIA AGX Xavier module is used as the edge device to run trained models on as a simulated deployed unmanned aerial system. Performance of the proposed approach is measured in terms of classification accuracy and run time.

Pulses are a major source of human protein intake nowadays and will continue to be so because of their high protein content. Pulse crops are members of the family Leguminosae. The five major pulse crops grown in Canada are chick peas, green peas, lentils, pinto bean and kidney beans. Over the past 20 years, Canada has emerged as the world’s largest exporter of lentils and one of world’s top five exporters of beans. These contribute more than $2 billion income to the Canadian economy. The major causes of fungal infection in these pulses are Aspergillus flavus and Penicillium commune. Early stages of fungal infections in pulses are not detectable with human eyes. Near infrared (NIR) hyperspectral imaging system is an advanced technique widely used for detection of insect infestation and fungal infection in cereal grains and oil seeds. A typical NIR instrument captures images across the electromagnetic spectrum at evenly spaced wavelengths from 700 to 2500 nm (a system at the University of Manitoba captures images in the 960 nm to 1700 nm range). From the captured images, the spatial relationships for different spectra in the neighborhood can be found allowing more elaborate spectral-spatial methods for a more accurate classification of the images. The primary objective of this study was to assess the feasibility of the NIR hyperspectral system to identify fungal infections in pulses. Hyperspectral images of healthy and fungal infected chick peas, green peas, lentils, pinto bean and kidney beans were acquired and features (statistical and histogram) were used to develop classification models to identify fungal infection caused by Aspergillus flavus and Penicillium commune. Images of healthy and fungal infected kernels were acquired at 2 week intervals (0, 2, 4, 6, 8 and 10 weeks from artificial inoculation). Six-way (healthy vs the five different stages of infection) and two-way (healthy vs every stage of infection) models were developed and classifications were done using linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) classifiers. The LDA classifier identified with 90-94% accuracy while using the six-way model, and with 98-100% accuracy when using the two-way models for all five types of pulses and for both types of fungal infections. The QDA classifier also showed promising results as it identified 85-90% while using the six-way model and 96-100% when using the two-way models. Hence, hyperspectral imaging is a promising and non-destructive method for the rapid detection of fungal infections in pulses, which cannot be detected using human eyes. / October 2015