ITC/USA 2013 Conference Proceedings / The Forty-Ninth Annual International Telemetering Conference and Technical Exhibition / October 21-24, 2013 / Bally's Hotel & Convention Center, Las Vegas, NV / Compressive Sensing (CS) is a set of techniques that can faithfully acquire a signal from sub- Nyquist measurements, provided the class of signals have certain broadly-applicable properties. Reconstruction (or exploitation) of the signal from these sub-Nyquist measurements requires a forward model - knowledge of how the system maps signals to measurements. In high-dimensional CS systems, determination of this forward model via direct measurement of the system response to the complete set of impulse functions is impractical. In this paper, we will discuss the development of a parameterized forward model for the Adaptive, Feature-Specific Spectral Imaging Classifier (AFSSI-C), an experimental compressive spectral image classifier. This parameterized forward model drastically reduces the number of calibration measurements.

Remote Sensing Group (RSG) at the University of Arizona has a long history of using ground-based test sites for the calibration of airborne- and satellite-based sensors. Often, ground-truth measurements at these tests sites are not always successful due to weather and funding availability. Therefore, RSG has also employed automated ground instrument approaches and cross-calibration methods to verify the radiometric calibration of a sensor. The goal in the cross-calibration method is to transfer the calibration of a well-known sensor to that of a different sensor.This dissertation presents a method for determining the radiometric calibration of a hyperspectral imager using multispectral imagery. The work relies on a multispectral sensor, Moderate-resolution Imaging Spectroradiometer (MODIS), as a reference for the hyperspectral sensor Hyperion. Test sites used for comparisons are Railroad Valley in Nevada and a portion of the Libyan Desert in North Africa. A method to predict hyperspectral surface reflectance using a combination of MODIS data and spectral shape information is developed and applied for the characterization of Hyperion. Spectral shape information is based on RSG's historical in situ data for the Railroad Valley test site and spectral library data for the Libyan test site. Average atmospheric parameters, also based on historical measurements, are used in reflectance prediction and transfer to space. Results of several cross-calibration scenarios that differ in image acquisition coincidence, test site, and reference sensor are found for the characterization of Hyperion. These are compared with results from the reflectance-based approach of vicarious calibration, a well-documented method developed by the RSG that serves as a baseline for calibration performance for the cross-calibration method developed here. Cross-calibration provides results that are within 2% of those of reflectance-based results in most spectral regions. Larger disagreements exist for shorter wavelengths studied in this work as well as in spectral areas that experience absorption by the atmosphere.

A platform technology is identified for grain handling facilities to improve grading and determine non-destructively different quality parameters of wheat. In this study, a near-infrared (NIR) hyperspectral imaging system was used to scan four wheat classes namely, Canada Western Red Spring (CWRS), Canada Prairie Spring Red (CPSR), Canada Western Hard White Spring (CWHWS), and Canada Western Soft White Spring (CWSWS) that were collected from across various growing regions in Manitoba, Saskatchewan, and Alberta in 2007, 2008, and 2009 crop years. A database of the near-infrared (NIR) hyperspectral image cubes of bulk samples of four wheat classes at three moisture levels for each class was created. These image cubes were acquired in the wavelength region of 960-1700 nm with 10 nm intervals. Wheat classification was done using the non-parametric statistical and a four-layer back propagation neural network (BPNN) classifiers. Average classification accuracies of 93.1 and 83.9% for identifying wheat classes using the linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA), respectively, were obtained for two-class identification models that included variations of moisture levels, growing locations, and crop years of samples. In the pair-wise moisture discrimination study, near-perfect classifications were achieved for wheat samples which had difference in moisture levels of about 6%. The NIR wavelengths of 1260-1380 nm had the highest factor loadings for the first principal component using the principal components analysis (PCA). A four-layer BPNN classifier was used for two-class identification of wheat classes and moisture levels. Overall average pair-wise classification accuracies of 83.7% were obtained for discriminating wheat samples based on their moisture contents. Classification accuracies of 83.2, 75.4, 73.1%, on average, were obtained for identifying wheat classes for samples with 13, 16, and 19% moisture content (m.c.), respectively. Ten-factor partial least squares regression (PLSR) and principal components regression (PCR) models were developed using a ten-fold cross validation for prediction. Prediction performances of PLSR and PCR models were assessed by calculating the estimated mean square errors of prediction (MSEP), standard error of cross-validation (SECV), and correlation coefficient (r). Overall, PLSR models demonstrated better prediction performances than the PCR models for predicting protein contents and hardness of wheat.

A platform technology is identified for grain handling facilities to improve grading and determine non-destructively different quality parameters of wheat. In this study, a near-infrared (NIR) hyperspectral imaging system was used to scan four wheat classes namely, Canada Western Red Spring (CWRS), Canada Prairie Spring Red (CPSR), Canada Western Hard White Spring (CWHWS), and Canada Western Soft White Spring (CWSWS) that were collected from across various growing regions in Manitoba, Saskatchewan, and Alberta in 2007, 2008, and 2009 crop years. A database of the near-infrared (NIR) hyperspectral image cubes of bulk samples of four wheat classes at three moisture levels for each class was created. These image cubes were acquired in the wavelength region of 960-1700 nm with 10 nm intervals. Wheat classification was done using the non-parametric statistical and a four-layer back propagation neural network (BPNN) classifiers. Average classification accuracies of 93.1 and 83.9% for identifying wheat classes using the linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA), respectively, were obtained for two-class identification models that included variations of moisture levels, growing locations, and crop years of samples. In the pair-wise moisture discrimination study, near-perfect classifications were achieved for wheat samples which had difference in moisture levels of about 6%. The NIR wavelengths of 1260-1380 nm had the highest factor loadings for the first principal component using the principal components analysis (PCA). A four-layer BPNN classifier was used for two-class identification of wheat classes and moisture levels. Overall average pair-wise classification accuracies of 83.7% were obtained for discriminating wheat samples based on their moisture contents. Classification accuracies of 83.2, 75.4, 73.1%, on average, were obtained for identifying wheat classes for samples with 13, 16, and 19% moisture content (m.c.), respectively. Ten-factor partial least squares regression (PLSR) and principal components regression (PCR) models were developed using a ten-fold cross validation for prediction. Prediction performances of PLSR and PCR models were assessed by calculating the estimated mean square errors of prediction (MSEP), standard error of cross-validation (SECV), and correlation coefficient (r). Overall, PLSR models demonstrated better prediction performances than the PCR models for predicting protein contents and hardness of wheat.

This dissertation develops new approaches for improving the performance of hyperspectral target detection. Different aspects of hyperspectral target detection are reviewed and studied to effectively distinguish target features from background interference. The contributions of this dissertation are detailed as follows. 1) Propose an adaptive background characterization method that integrates region segmentation with target detection. In the experiments, not only unstructured matched filter based detectors are considered, but also two hybrid detectors combining fully constrained least squared abundance estimation with statistic test (i.e., adaptive matched subspace detector and adaptive cosine/coherent detector) are investigated. The experimental results demonstrate that using local adaptive background characterization, background clutters can be better suppressed than the original algorithms with global characterization. 2) Propose a new approach to estimate abundance fractions based on the linear spectral mixture model for hybrid structured and unstructured detectors. The new approach utilizes the sparseness constraint to estimate abundance fractions, and achieves better performance than the popular non-negative and fully constrained methods in the situations when background endmember spectra are not accurately acquired or estimated, which is very common in practical applications. To improve the dictionary incoherence, the use of band selection is proposed to improve the sparseness constrained linear unmixing. 3) Propose random projection based dimensionality reduction and decision fusion approach for detection improvement. Such a data independent dimensionality reduction process has very low computational cost, and it is capable of preserving the original data structure. Target detection can be robustly improved by decision fusion of multiple runs of random projection. A graphics processing unit (GPU) parallel implementation scheme is developed to expedite the overall process. 4) Propose nonlinear dimensionality reduction approaches for target detection. Auto-associative neural network-based Nonlinear Principal Component Analysis (NLPCA) and Kernel Principal Component Analysis (KPCA) are applied to the original data to extract principal components as features for target detection. The results show that NLPCA and KPCA can efficiently suppress trivial spectral variations, and perform better than the traditional linear version of PCA in target detection. Their performance may be even better than the directly kernelized detectors.

<p dir="ltr">Advancements in machine learning are narrowing the gap in visual capabilities between machines and healthcare professionals, resulting in a transformation of the way we understand and address health challenges. Despite these advances, underlying limitations persist in addressing real-world problems, particularly in the precise capture of biological and physiological information. This is primarily because traditional trichromatic cameras fall short of representing reflectance spectra due to their limited spectral information. To overcome these limitations, hyperspectral imaging has emerged as a powerful tool for biomedical applications. By collecting a wealth of information at different wavelengths, hyperspectral imaging provides a comprehensive view of electromagnetic spectra, allowing non-invasive clinical analysis for accurate diagnostics. Snapshot hyperspectral imaging, in particular, is a competitive alternative to traditional cameras as it can capture a hyperspectral image in a single shot without the need for scanning individual wavelengths. Here, we introduce a computational snapshot hyperspectral imaging method, achieved through the integration of a machine learning approach with a streamlined optical system. We design an explainable machine learning algorithm by incorporating optical and biological knowledge into the algorithm. Therefore, the algorithm can reconstruct hyperspectral images with high spectralspatial resolution comparable to those of scientific spectrometers, despite the use of sparse information captured from the optical system. To demonstrate its versatility in biomedical applications, we extract hemodynamic parameters of peripheral microcirculation from embryonic model systems, tissue phantom samples, and human conjunctivas. Furthermore, we validate high accuracy of the results using conventional hyperspectral imaging and functional near-infrared spectroscopy. This learning-powered imaging method, characterized by high resolution and simplified hardware requirements, has the potential to offer solutions for various biomedical challenges by surpassing the constraints of conventional cameras and hyperspectral imaging.</p>

The objective of this dissertation is to investigate all the necessary components in spectral mixture analysis (SMA) for hyperspectral imagery under an unsupervised circumstance. When SMA is linear, referred to as linear spectral mixture analysis (LSMA), these components include estimation of the number of endmembers, extraction of endmember signatures, and calculation of endmember abundances that can automatically satisfy the sum-to-one and non-negativity constraints. A simple approach for nonlinear spectral mixture analysis (NLSMA) is also investigated. After SMA is completed, a color display is generated to present endmember distribution in the image scene. It is expected that this research will result in an analytic system that can yield optimal or suboptimal SMA output without prior information. Specifically, the uniqueness in each component is described as follow. 1)A new signal subspace-based approach is developed to determine the number of endmembers with relatively robust performance and the least parameter requirement. 2)The best implementation strategy is determined for endmember extraction algorithms using simplex volume maximization and pixel spectral similarity; and algorithm with the special consideration for anomalous pixels is developed to improve the quality of extracted endmembers. 3)A new linear mixture model (LMM) is deployed where the number of endmembers and their types can be changed from pixel to pixel such that the resulting endmember abundances are sum-to-one and nonnegative as required; and fast algorithms are developed to search for a sub-optimal endmember set for each pixel. 4)A simple approach for NLSMA based on LMM is investigated and an automated approach is developed to determine either linear or nonlinear mixing is actually experienced. 5)A color display strategy is developed to present SMA results with high class/endmember separability.

A Hyperspectral imager aboard a cubesat would be able to provide images which could be used to identify ships and determine the ship's length and breadth and heading. Depending on the size of the ship, the speed the ship is traveling can be determined as well; however the speed and size determination is limited by the spatial resolution of 100 meters. The spectral signature of the boat is dramatically different from the spectral signature of the open Ocean especially within the range of 400 to 1000 nanometers, and this threshold is the basis for extracting ship data. Hyperspectral Imagers are ideal for minimization with few optical errors introduced, and designs range in durability making them useful on board small satellites especially in the visible and near infrared region. Placing an imager on a satellite allows for consistent observation over a region to identify patterns in ship movement over time. / Master of Science

Agricultural productivity, ecosystem health, and wetland restoration rely on soil organic carbon (SOC) as vital for microbial activity and plant health. This study assessed: (1) accuracy of topographic-based non-linear models for predicting SOC; and (2) the effect of analytic strategies and soil condition on performance of spectral-based models for predicting SOC. SOC data came from 28 agriculturally converted Delmarva Bays sampled down to 1 meter. R2 was used as an indicator of model performance. For topographic-based modeling, correlation coefficients and condition indices reduced 50 terrain-related values to three datasets of 16, 11, and 7 variables. Five types of non-linear models were examined: Generalized Linear Mondel (GLM) ridge, GLM LASSO, Generalized Additive Model (GAM) non-penalized, GAM cubic splice, and partial least-squares regression. Carbon stocks varied widely, 50 to 219 Mg/ha, with the average around 93 Mg/ha. Topography shared a weak relationship to SOC with most attributes showing a correlation coefficient less than 0.3. GLM ridge and both GAMs achieved moderate accuracy at least once, usually using the 16 or 11 variable datasets. GAMs consistently performed the best. Prior to carbon analysis, hyperspectral signatures were recorded for the topmost soil horizons under different conditions: moist unground, dry unground, and dry ground. Twenty-four math treatment and smoothing technique combinations were run on each hyperspectral dataset. R2 varied greatly within datasets depending on analytic strategy, but all datasets returned an R2 greater than 0.9 at least twice. Moist unground soil models outperformed the others when comparing the best models among datasets. / Master of Science

Le démelangeage spectral est un domaine de recherche actif qui trouve des applications dans des domaines variés comme la télédétection, le traitement des signaux audio ou la chimie. Dans le contexte des capteurs hyper spectraux, les images acquises sont souvent de faible résolution spatiale, principalement à cause des limites technologiques liées aux capteurs. Ainsi, les pixels sont constitués des mélanges des différentes signatures spectrales des matériaux présents dans la scène observée. Le démélangeage hyperspectral correspond à la procédure inverse permettant d'identifier la présence de ces matériaux ainsi que leur abondance par pixel. Déterminer le nombre total de matériaux dans l'image et par pixel est un problème difficile. Des approches à base de modèle de mélange linéaire ont été développées mais l’hypothèse sous-jacente de linéarité est parfois mise à mal dans des scénarios réels. Le problème est amplifié lorsqu'un même matériel présente une forte variabilité de signatures spectrales. De plus, la présence de nombreuses signatures parasites (ou anomalies) rend l'estimation plus difficile. Ces différents problèmes sont abordés dans cette thèse au travers de solutions théoriques et algorithmiques. La première contribution porte sur un démélangeage non-linéaire parcimonieux basé sur des approches à noyaux (SAGA+), qui estime et enlevé de l'analyse simultanément les anomalies. La deuxième contribution majeure porte sur une méthode de démélangeage supervisée basée sur la théorie du transport optimal (OT-unmixing) et permet d'intégrer la variabilité potentielle des matériaux observés. Un cas d'étude réel, dans le contexte du projet CATUT, et visant l'estimation des températures de surface par imagerie aéroportée, est finalement décrit dans la dernière partie de ce travail. / Spectral Unmixing has been an active area of research during the last years and found its application in domains including but not limited to remote sensing, audio signal processing and chemistry. Despite their very high spectral resolution, hyperspectral images (HSI) are known to be of low spatial resolution. This low resolution is a relative notion and is due to technological limitations of the HSI captors. As a consequence the values of HSI pixels are likely to be mixtures Of diferent materials in the scene. hyperspectral Unmixing then can be dened as an inverse procedure that consists in identifying in each pixel the amount of pure elements contributing to the pixels mixture. The total number of pure elements (also called endmembers) and the number of them included in one pixel are two informations tricky to retrieve. The simplest situation is when both the total number and type of endmembers within the scene are known and associated with a linear mixing process assumption. Though efficient in some situations, this linearity assumption does not generally hold in real world scenarios. Also in most cases the knowledge regarding the endmember signature of a specic material is not exact, raising the need to account for variations among different representations of the same material. Last but not least existence of anomalies and noise is a ubiquitous issue affecting the accuracy of the estimations. In this thesis, the three aforementioned issues were mainly brought into light and by introducing two original algorithms, defined within different mathematical frameworks, solutions to these open problems has provided. The first contribution using the applications of kernel theory proposes a new unsupervised algorithm (SAGA+) for representation of the non-linear manifold embedding the data while through a simultaneous anomaly detection procedure makes sure that the representation of the manifold hall is not being distorted at the presence of anomalies. The second major contribution of this PhD focuses mainly on the issue of endmember variability and by exploiting the notion of overcomplete dictionary tries to address this problem. This supervised algorithm (OT-unmixing) which is based on the optimal transport theory is comparable to the second step of SAGA+, as it solves an inversion problem and calculates the sparse representation of the original pixels through generation of the abundance maps. A case study in the context of CATUT project for land surface temperature estimation is described in the last part of this work where the two algorithms used for unmixing of airborne hyperspectral remote sensing.