Les images hyperspectrales (HSI) fournissent des informations spectrales détaillées sur les objets analysés. Étant donné que différents matériaux ont des signatures spectrales distinctes, les objets ayant des couleurs et des formes similaires peuvent être distingués dans le domaine spectral. Toutefois, l’énorme quantité de données peut poser des problèmes en termes de stockage et de transmission des données. De plus, la haute dimensionnalité des images hyperspectrales peut entraîner un surajustement du classificateur en cas de données d'apprentissage insuffisantes. Une façon de résoudre de tels problèmes consiste à effectuer une sélection de bande (BS), car elle réduit la taille du jeu de données tout en conservant des informations utiles et originales. Dans cette thèse, nous proposons trois méthodes de sélection de bande différentes. La première est supervisée, conçu pour utiliser seulement 20% des données disponibles. Pour chaque classe du jeu de données, une classification binaire un contre tous utilisant un réseau de neurones est effectuée et les bandes liées aux poids le plus grand et le plus petit sont sélectionnées. Au cours de ce processus, les bandes les plus corrélées avec les bandes déjà sélectionnées sont rejetées. Par conséquent, la méthode proposée peut être considérée comme une approche de sélection de bande orientée par des classes. La deuxième méthode que nous proposons est une version non supervisée du premier framework. Au lieu d'utiliser les informations de classe, l'algorithme K-Means est utilisé pour effectuer une classification binaire successive de l'ensemble de données. Pour chaque paire de grappes, un réseau de neurones à une seule couche est utilisé pour rechercher l'hyperplan de séparation, puis la sélection des bandes est effectuée comme décrit précédemment. Pour la troisième méthode de BS proposée, nous tirons parti de la nature non supervisée des auto-encodeurs. Pendant la phase d'apprentissage, le vecteur d'entrée est soumis au bruit de masquage. Certaines positions de ce vecteur sont basculées de manière aléatoire sur zéro et l'erreur de reconstruction est calculée sur la base du vecteur d'entrée non corrompu. Plus l'erreur est importante, plus les fonctionnalités masquées sont importantes. Ainsi, à la fin, il est possible d'avoir un classement des bandes spectrales de l'ensemble de données. / Hyperspectral images (HSIs) are capable of providing a detailed spectral information about scenes or objects under analysis. It is possible thanks to both numerous and contiguous bands contained in such images. Given that di_erent materials have distinct spectral signatures, objects that have similar colors and shape can be distinguished in the spectral domain that goes beyond the visual range. However, in a pattern recognition system, the huge amount of data contained in HSIs may pose problems in terms of data storage and transmission. Also, the high dimensionality of hyperspectral images can cause the overfitting of the classifer in case of insufficient training data. One way to solve such problems is to perform band selection(BS) in HSIs, because it decreases the size of the dataset while keeping both useful and original information. In this thesis, we propose three different band selection frameworks. The first one is a supervised one, and it is designed to use only 20% of the available training data. For each class in the dataset, a binary one-versus-all classification using a single-layer neural network is performed, and the bands linked to the largest and smallest coefficients of the resulting hyperplane are selected. During this process, the most correlated bands with the bands already selected are automatically discarded, following a procedure also proposed in this thesis. Consequently, the proposed method may be seen as a classoriented band selection approach, allowing a BS criterion that meets the needs of each class. The second method we propose is an unsupervised version of the first framework. Instead of using the class information, the K-Means algorithm is used to perform successive binary clustering of the dataset. For each pair of clusters, a single-layer neural network is used to find the separating hyperplane, then the selection of bands is done as previously described. For the third proposed BS framework, we take advantage of the unsupervised nature of autoencoders. During the training phase, the input vector is subjected to masking Noise - some positions of this vector are randomly flipped to zero and the reconstruction error is calculated based on the uncorrupted input vector. The bigger the error, the more important the masked features are. Thus, at the end, it is possible to have a ranking of the spectral bands of the dataset.

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.

A utilização de dados provenientes do sensoriamento remoto é alternativa para otimizar a utilização de insumos, dentre eles o nitrogênio. O presente trabalho teve como objetivo verificar a possibilidade de uso de um sensor hiperespectral em dossel na cultura da cana-de-açúcar, verificando sua capacidade em discriminar a resposta da cultura as diferentes doses de nitrogênio e estimar o teor foliar de nitrogênio, em três áreas experimentais. O trabalho foi dividido em três capítulos: O primeiro capitulo utiliza os dados hiperespectrais somente da variedade SP 81-3250, única comum em todas as áreas, de todas as datas de coleta das três áreas experimentais para verificar o potencial dos dados em diferenciar as doses de nitrogênio aplicado (0, 50, 100 e 150 kg.ha-1) e qual melhor época. Os dados espectrais foram avaliados pela estatística multivariada da análise discriminante, em que os centroides das diferentes doses foram submetidos a análise de variância. Os resultados obtidos foram que os meses de dezembro, janeiro e fevereiro discriminou todas as doses nas três áreas, o mesmo não ocorreu no mês de agosto. As bandas que apresentaram maiores significância foram na região do verde, red-edge e infravermelho próximo. No segundo capitulo foi avaliado a sensibilidade dos dados hiperespectrais em estimar a biomassa do ponteiro da cana-de-açúcar. Para isso foi utilizado somente os dados de Piracicaba. A análise espectral foi realizada aos 137, 169 e 193 Dias Após o Corte (DAC) e a avaliação biométrica foi realizada aos 345 DAC. Durante o corte de dois metros de linha, realizado manualmente. A biomassa do ponteiro foi submetida ao teste de Shapiro-Wilk, análise de variância pelo Teste F e as médias quando significativas, comparadas pelo Teste de Tukey. Posteriormente foi realizada a análise de correlação de Pearson da biomassa do ponteiro e cada comprimento de onda. Análise mostrou que existe correlação positiva entre a biomassa do ponteiro e a reflectância do dossel aos 137 DAC e 169 DAC, porém aos 193 DAC não houve nenhum comprimento de onda com correlação significativa. O comprimento de onda de 685 nm aos 137 DAC obteve a maior correlação, de 0,33. No terceiro capitulo teve por objetivo selecionar variáveis a partir de dados hiperespectrais de dossel da cana-de-açúcar para geração de modelos para predição do Teor Foliar de Nitrogênio. Para isso foi utilizado os dados das três áreas experimentais, que receberam doses de 0, 50, 100 e 150 kg.ha-1 de nitrogênio. Para redução da dimensionalidade dos dados foi utilizada a metodologia sparse Partial Least Square (sPLS), posteriormente foi feito a combinação linear das variáveis selecionadas, por meio de Regressão Linear Múltipla por Stepwise (SMLR). O modelo geral teve valores de R² ajustado e RMSE respectivamente de 0,50 e 1,67 g kg-1. Os modelos gerados para Piracicaba, Jaú e Santa Maria obtiveram R² ajustado, respectivamente, de 0,31, 0,53 e 0,54. Sensores hiperespectrais de dossel podem ser utilizados para predição do TFN e monitoramento de aplicação de nitrogênio em cana-de-açúcar. / The use of data from remote sensing is an alternative to optimize the use of agricultural inputs, including nitrogen. The present study aimed to verify the possibility of using a hyperspectral sensor in sugarcane canopy, verifying its ability to discriminate crop response to different rates of nitrogen and estimating leaf nitrogen content in three experimental areas. The work is divided in three chapters: The first chapter uses hyperspectral data of the variety SP 81-3250, which is the only one present in all the areas for all dates of collection in three of experimental areas, to check the potential of the data and the best time to differentiate between rates of nitrogen (0, 50, 100 and 150 kg.ha-1). Spectral data were evaluated by multivariate discriminant analysis, wherein the centroids of the rates were submitted to an Analysis of Variance. The results showed that the all doses in three areas of study were discriminated for the months of December, January and February, but the same thing hasn\'t happened in the month of August. The bands that showed statistically significant power difference were found in the green, red, and near-infrared edge spectral regions. In the second chapter, the sensitivity of hyperspectral data was evaluated to estimate the sugarcane biomass (pointes) for the data from Piracicaba. Spectral analysis was performed at 137, 169 and 193 Days After Harvest (DAH) and evaluation of sugarcane yield was performed 345 DAH. Biomass was analyzed using The Shapiro-Wilk test of normality, F test (analysis of variance), respectively, and when significant, compared by the Tukey test. Biomass (pointer) and each wavelength were analyzed by Pearson\'s correlation analysis. The results showed that there is a positive correlation between biomass (pointer) and the canopy reflectance to 137 DAH and 169 DAH, however there was no wavelength with a significant correlation to 193 DAH. The best power relationship was obtained at 685 nm, at 137 days. The third chapter aimed to select variables from hyperspectral data of sugarcane canopy to generate models for prediction of Foliar Nitrogen Content, for three experimental areas that received nitrogen rates (0, 50, 100 and 150 kg.ha-1). Sparse Partial Least Square (sPLS) was used to reduce the dimensionality of the data. Subsequently, the linear combination of selected variables was done through Stepwise Multiple Linear Regression (SMLR). The RMSE and adjusted R-squared statistics were 0.50 and 1.67 g.kg-1, respectively. The models to Piracicaba, Jaú and Santa Maria presented adjusted R-squared 0.31, 0.53, and 0.54, respectively. Hyperspectral sensors for canopy can be used for prediction of the TFN and monitoring of nitrogen application in sugarcane.

O objetivo do presente trabalho foi avaliar o uso potencial dos dados hiperespectrais do sensor orbital Hyperion/Earth Observing One (EO-1) e dos modelos de mistura espectral MESMA (Multiple Endmember Spectral Mixture Analysis) e SMA (Spectral Mixture Analysis) para discriminação de classes de cobertura da Planície Costeira do Rio Grande do Sul. O modelo MESMA difere do SMA por permitir que o número e o tipo de Membros de Referência (MRs), assim como sua abundância, variem pixel a pixel. A abordagem metodológica utilizada envolveu as seguintes etapas: (a) préprocessamento dos dados Hyperion e conversão dos valores de radiância para imagens atmosfericamente corrigidas de reflectância de superfície; (b) uso seqüencial das técnicas Minimum Noise Fraction (MNF), Pixel Purity Index (PPI) e Visualizador n- Dimensional, no intervalo de 454 a 2334 nm, para seleção inicial de um grupo de pixels candidatos a MRs (primeira biblioteca espectral) e de um outro grupo para fins de validação dos modelos; (c) uso do aplicativo VIPER (Visualization and Image Processing for Environmental Research) Tools para refinamento da primeira biblioteca espectral e seleção final dos MRs, utilizando as métricas EAR (Endmember Average RMSE), MASA (Minimum Average Spectral Angle) e CoB (Count Based Endmember Selection); (d) geração dos modelos MESMA e SMA com o VIPER Tools; e (e) comparação dos resultados dos modelos com base nas imagens-fração e nos valores de erro médio quadrático (RMSE). Os resultados obtidos mostraram que: (1) o uso seqüencial das técnicas MNF, PPI e Visualizador n-Dimensional pode constituir uma etapa inicial para identificar pixels candidatos a MRs, cuja seleção final pode ser feita com as métricas EAR, MASA e CoB. Usadas de forma combinada, essas métricas minimizam possíveis efeitos da baixa relação sinal-ruído do Hyperion; (2) os MRs selecionados representaram os principais componentes de cena como “água” (com clorofila, límpida e com sedimentos em suspensão), “vegetação verde” (pinus, eucalipto e gramíneas) e “solo” (dunas e campo seco); (3) Por utilizar número e tipo variáveis de MRs, o modelo MESMA produziu melhores resultados que o SMA. Quando aplicado sobre a imagem, sobre a amostra de validação e quando comparado com o SMA, o modelo MESMA de 4 componentes (Solo = dunas e campo Seco; vegetação verde = pinus, eucalipto e gramíneas; água = com Sedimentos em suspensão, sem Sedimentos e com clorofila; sombra) descreveu adequadamente a diversidade dos componentes de cena, incluindo materiais dentro de uma mesma classe (p.ex. pinus e eucalipto). O MESMA produziu menores valores de RMSE e uma maior quantidade de pixels modelados na cena (85% contra 55%) do que o SMA; (4) o VIPER mostrou-se uma ferramenta bastante eficaz para seleção dos MRs e geração dos modelos. Os resultados, como um todo, demonstraram o potencial da aplicação dos modelos MESMA com dados hiperespectrais do sensor Hyperion/EO-1, mesmo considerando a baixa relação sinal/ruído do instrumento, especialmente no infravermelho de ondas curtas (SWIR). / The objective of this work was to evaluate the potential use of the Hyperion/Earth Observing One (EO-1) hyperspectral data and of the MESMA (Multiple Endmember Spectral Mixture Analysis) and SMA (Spectral Mixture Analysis) mixture models to discriminate land covers in the Rio Grande do Sul state, South Brazil. MESMA differs from SMA because it may use a variable number and type of endmembers in each pixel. The methodology involved: (a) pre-processing of Hyperion data and conversion of radiance values into atmospherically corrected surface reflectance images; (b) sequential use of the Minimum Noise Fraction (MNF), Pixel Purity Index (PPI) and n- Dimensional Visualizer techniques, in the 454-2334 nm range, for initial selection of a general group of candidate endmembers (first spectral library) and of another group of pixels used for model validation; (c) use of VIPER (Visualization and Image Processing for Environmental Research) Tools algorithm for final selection of endmembers from the first spectral library and from the use of the metrics EAR (Endmember Average RMSE), MASA (Minimum Average Spectral Angle) and CoB (Count Based Endmember Selection); (d) use of VIPER tools to obtain MESMA and SMA models; and (e) comparison of modeling results based on the inspection of fraction images and root mean square error (RMSE) values. Results showed that: (1) the sequential use of the MNF, PPI and n-D Visualizer techniques may comprise an initial step to identify candidate endmembers. Final selection was performed using a combination of EAR, MASA and CoB to minimize possible effects of low signalnoise ratio (SNR) of Hyperion; (2) the selected endmembers represented major scene components such as water (with chlorophyll, clear or bearing in suspended sediments), green vegetation (pinus, eucalyptus and grasslands) and soil (dunes and dry grasslands); (3) By using a variable number and type of endmembers, MESMA produced better results than SMA. When applied over the image, the validation dataset and compared with SMA, the four-endmember MESMA model (soil = dunes and dry grasslands; green vegetation = pinus, eucalyptus and grasslands; water = with chlorophyll, clear and with suspended sediments; shadow) described adequately the diversity of the scene components, including materials within the same class (e.g., pinus and eucalyptus). MESMA produced lower RMSE values and greater number of modeled pixels (85% versus 55%) than SMA; (5) the VIPER tools seems to be an interesting approach for endmember selection and spectral mixture model generation. Results, as a whole, demonstrated the potential use of the MESMA with Hyperion/EO-1 hyperspectral data, even considering the low SNR of the instrument, especially in the shortwave infrared (SWIR).

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.

Hyperspectral images are made up of energy measurements at different wavelengths of light. The case is considered where these measurements are dependent on temperature, the self-emitted energy (emissivity), and reflected energy (downwelling radiance) from the surroundings. The process where the downwelling radiance is fixed and the temperature and emissivity are estimated is referred to as temperature/emissivity separation. Due to the way these terms mix, for a given set of measurements, there exist many pairs of temperatures and emissivities that satisfy the model. This creates ambiguity in the solution that must be resolved for the result to have any significance. A new model is developed which reduces this ambiguity. This model is used to form an objective function. The temperature and emissivity which maximize the value of the objective function are solved for given a set of measurements. As part of the solution, a new algorithm is developed which exploits the shape of the objective function to estimate the temperature and emissivity quickly and accurately. Extensive testing of this algorithm is performed to gain an understanding of its average speed and accuracy.

Detection of chemical plumes in hyperspectral data is a problem having solutions that focus on spectral information. These solutions neglect the presence of the spatial information in the scene. The spatial information is exploited in this work by assignment of prior probabilities to neighborhood configurations of signal presence or absence. These probabilities are leveraged in a total probability approach to testing for signal presence in a pixel of interest. The two new algorithms developed are named spatial information detection enhancement (SIDE) and bolt-on SIDE (B-SIDE). The results are explored in comparison to the clutter matched filter (CMF), a standard spectral technique, and to several supervised machine learning techniques. The results show a great improvement of SIDE over these other techniques, in some cases showing the poorest performance of the SIDE filter being much better than the CMF at its best.

Hyperspectral remote sensing image data, such as that recorded by AVIRIS with 224 spectral bands, provides rich information on ground cover types. However, it presents new problems in machine assisted interpretation, mainly in long processing times and the difficulties of class training due to the low ratio of number of training samples to the number of bands. This thesis investigates feasible and efficient feature reduction and image classification techniques which are appropriate for hyperspectral image data. The study is reported in three parts. The first concerns a deterministic approach for hyperspectral data interpretation. Multigroup and multiple threshold spectral coding procedures, and associated techniques for spectral matching and classification, are proposed and tested. By coding on subgroups of bands using one or three thresholds, spectral searching and matching becomes simple, fast and free of the need for radiometric correction. Modifications of existing statistical techniques are proposed in the second part of the investigation A block-based maximum likelihood classification technique is developed. Several subgroups are formed from the complete set of spectral bands in the data, based on the properties of global correlation among the bands. Subgroups which are poorly correlated with each other are treated independently using conventional maximum likelihood classification. Experimental results demonstrate that, when using appropriate subgroup sizes, the new method provides a compromise among classification accuracy, processing time and available training pixels. Furthermore, a segmented, and possibly multi-layer, principal components transformation is proposed as a possible feature reduction technique prior to classification, and for effective colour display. The transformation is performed efficiently on each of the highly correlated subgroups of bands independently. Selected features from each transformed subgroup can be then transformed again to achieve a satisfactory data reduction ratio and to generate the three most significant components for colour display. Classification accuracy is improved and high quality colour image display is achieved in experiments using two AVIRIS data sets.

L'objectif de cette thèse est d'étudier les potentialités de la télédétection aérospatiale pour l'aide à la gestion de sole cannière Réunionnaise. Nous avons utilisé une base de données d'images multitemporelles SPOT 4&5 (années 2002 et 2003) et organisé une campagne d'acquisition d'images hyperspectrales CASI en septembre 2002. Simultanément, nous avons assuré le déroulement et la mise en place d'un protocole de mesures au champ pour suivre l'évolution des paramètres biophysiques descriptifs de l'état du couvert de la canne (surface foliaire, taux d'azote, biomasse de la culture) et des paramètres agronomiques (suivi des coupes et des replantations). Les résultats ont montré qu'il est possible d'estimer la surface foliaire (LAI) à partir de l'indice de végétation normalisé (NDVI) ainsi que le rendement canne à partir de l'indice de végétation NDVI calculé au moment du développement maximal du couvert. Avec les données SPOT, la meilleure estimation du rendement canne à l'échelle parcellaire résulte du couplage entre le modèle de croissance Mosicas et les profils d'évolution de surface foliaire obtenus à partir des images SPOT 4&5. Les données hyperspectrales CASI permettent une meilleure estimation de la surface foliaire et de la biomasse fraîche que les données SPOT 4&5 ainsi qu'une estimation du taux d'azote foliaire qui est, en phase de maturation, un indicateur de richesse en sucre. La possibilité de discriminer des parcelles de canne en fonction de leurs états de surface (pleine végétation, coupée ou labourée) nous a permis de développer des applications opérationnelles de cartographie dynamique de la sole cannière en temps quasi réel : le suivi des coupes et des replantations.

<p>The subject of this thesis is the application of Support Vector Machines on two totally different applications, facial expressions recognition and remote sensing.</p><p>The basic idea of kernel algorithms is to transpose input data in a higher dimensional space, the feature space, in which linear operations on the data can be processed more easily. These operations in the feature space can be expressed in terms of input data thanks to the kernel functions. Support Vector Machines is a classifier using this kernel method by computing, in the feature space and on basis of examples of the different classes, hyperplanes that separate the classes. The hyperplanes in the feature space correspond to non linear surfaces in the input space.</p><p>Concerning facial expressions, the aim is to train and test a classifier able to recognise, on basis of some pictures of faces, which emotion (among these six ones: anger, disgust, fear, joy, sad, and surprise) that is expressed by the person in the picture. In this application, each picture has to be seen has a point in an N-dimensional space where N is the number of pixels in the image.</p><p>The second application is the detection of camouflage nets hidden in vegetation using a hyperspectral image taken by an aircraft. In this case the classification is computed for each pixel, represented by a vector whose elements are the different frequency bands of this pixel.</p>