Cette thèse porte sur la statistique spatiale multivariée et l’apprentissage appliqués aux images hyperspectrales et multimodales. Les thèmes suivants sont abordés :Fusion d'images :Le microscope électronique à balayage (MEB) permet d'acquérir des images à partir d'un échantillon donné en utilisant différentes modalités. Le but de ces études est d'analyser l’intérêt de la fusion de l'information pour améliorer les images acquises par MEB. Nous avons mis en œuvre différentes techniques de fusion de l'information des images, basées en particulier sur la théorie de la régression spatiale. Ces solutions ont été testées sur quelques jeux de données réelles et simulées.Classification spatiale des pixels d’images multivariées :Nous avons proposé une nouvelle approche pour la classification de pixels d’images multi/hyper-spectrales. Le but de cette technique est de représenter et de décrire de façon efficace les caractéristiques spatiales / spectrales de ces images. Ces descripteurs multi-échelle profond visent à représenter le contenu de l'image tout en tenant compte des invariances liées à la texture et à ses transformations géométriques.Réduction spatiale de dimensionnalité :Nous proposons une technique pour extraire l'espace des fonctions en utilisant l'analyse en composante morphologiques. Ainsi, pour ajouter de l'information spatiale et structurelle, nous avons utilisé les opérateurs de morphologie mathématique. / This thesis focuses on multivariate spatial statistics and machine learning applied to hyperspectral and multimodal and images in remote sensing and scanning electron microscopy (SEM). In this thesis the following topics are considered:Fusion of images:SEM allows us to acquire images from a given sample using different modalities. The purpose of these studies is to analyze the interest of fusion of information to improve the multimodal SEM images acquisition. We have modeled and implemented various techniques of image fusion of information, based in particular on spatial regression theory. They have been assessed on various datasets.Spatial classification of multivariate image pixels:We have proposed a novel approach for pixel classification in multi/hyper-spectral images. The aim of this technique is to represent and efficiently describe the spatial/spectral features of multivariate images. These multi-scale deep descriptors aim at representing the content of the image while considering invariances related to the texture and to its geometric transformations.Spatial dimensionality reduction:We have developed a technique to extract a feature space using morphological principal component analysis. Indeed, in order to take into account the spatial and structural information we used mathematical morphology operators

Monitoring the working environment of laboratories and industries for pollutants is of primary concern to ensure the healthiness of working personnel. Semiconducting metal oxides (SMOs) are sensitive to the gas ambience and can be tuned for sensing purpose. As SMOs are not selective, an array of sensors with differential selectivity may resolve to great extent. The objective of the thesis is to understand the physicochemical properties and gas sensing characteristics of Cr1-xFexNbO4. Applying principal component analysis to the sensor response data either for selection of features or for differentiation of analysts is also of concern. Preparation of Cr1-xFexNbO4, phase characterization, lattice parameters estimation, morphological and micro chemical analysis (SEM & EDX), electrical characterization by direct current (DC & AC) in the temperature range of 423 K to 573 K, weighted magnetic moment of iron and chromium deduced from susceptibility measurements, spin nature of iron and surface compositions of different valences of chromium and iron deduced from X-ray photoelectron spectroscopy of are presented. The wide dynamic range hydrogen sensing characteristics of CrNbO4 bulk pellets at different temperatures along with the cross-sensitivity towards NH3, NOx(NO+NO2) and PG (petroleum gas) are investigated. The preparation of Cr1-xFexNbO4 thick and thin films by screen-printing and PLD are also presented. The thick films are tested at different temperatures towards hydrogen. The n-type or p-type nature of thick films towards hydrogen with varying iron concentration in Cr1-xFexNbO4 is reported. The thin films are characterized for phase formation, morphology by XRD, SEM and AFM. XPS performed surface characterization. Electrical resistance measurements at different temperatures and preliminary experiments on hydrogen sensing are presented. The probable hydrogen sensing mechanism of CrNbO4 was revealed by X-ray photoelectron spectroscopy. The experimentally observed reduction in metal ion oxidation states upon interacting with hydrogen is best illustrated by Kröger Vink notation. Principal component analysis was applied for three different types of studies: i) The fit parameters of the transient response of CrNbO4 thick films towards hydrogen are analyzed for finding out the better feature for calibration, ii) Different thick films of CrNbO4, Cr0.5Fe0.5NbO4 and FeNbO4 operated at various temperatures for testing H2 and VOCs are analyzed for redundancy in sensor behaviour and iii) Cr0.8Fe0.2NbO4 thick films are studied for sensing H2, NH3 and their mixtures and usefulness of PCA in resolving them in PC-space. In addition, H2 and VOCs are tested at different temperatures and redundancy in temperature is deduced to construct a sensor array with a minimum number of sensors. Finally, a sensor array consisting of Cr0.8Fe0.2NbO4 thick films, operating at different temperatures is built, and qualitative discrimination of analysts in PC-space is demonstrated. Finally, the major findings of the present investigations and suggestions for future aspects of experimentation are provided

The traditional way of estimating the formant frequencies from articulatory data presupposes knowledge of how the vocal tract cross-sectional area varies for a given articulatory shape (Fant 1960/1970). Accordingly, in order to derive the formant pattern of a given articulation, the three-dimensional shape of the vocal tract (VT) needs to be known. In the past cross-sectional areas have typically been derived by means of ‘d-to-A rules’ that use the mid-sagittal cross-distance d at each point along the VT to produce a corresponding cross-sectional area A. X-ray and MRI data have been used to calibrate such rules (Heinz & Stevens 1964, Sundberg et al. 1987, Ericsdotter 2005). Although this procedure has produced many useful results it is time consuming and laborious. It is speaker-specific. It presupposes access to information on the three-dimensional shape of the VT, which is not experimentally readily accessible. Such observations raise the question whether sufficiently accurate alternative approaches can be developed. Is it possible to go straight from articulatory data to formant frequencies without having to construct a cross-sectional area function? If such methods could be developed it would have many uses both in phonetics and practical applications. This paper reports an attempt to map the time variations of selected articulatory parameters from X-ray profiles directly on the formant tracks using multiple regression analysis. Preliminary results for F1 indicate that multiple regression analysis can indeed be useful for making such predictions. The prospects of extending the present analyses to other formants are discussed.

Magister Scientiae (Biodiversity and Conservation Biology) - MSc (Biodiv and Cons Biol) / This research assessed landuse changes and trends in vegetation cover in the Sandveld, using remote sensing images. Landsat TM satellite images of 1990, 2004 and 2007 were classified using the maximum likelihood classifier into seven landuse classes, namely water, agriculture, fire patches, natural vegetation, wetlands, disturbed veld, and open sands. Change detection using remote sensing algorithms and landscape metrics was performed on these multi-temporal landuse maps using the Land Change Modeller and Patch Analyst respectively. Markov stochastic modelling techniques were used to predict future scenarios in landuse change based on the classified images and their transitional probabilities. MODIS NDVI multi-temporal datasets with a 16day temporal resolution were used to assess seasonal and annual trends in vegetation cover using time series analysis (PCA and time profiling).Results indicated that natural vegetation decreased from 46% to 31% of the total landscape between 1990 and 2007 and these biodiversity losses were attributed to an increasing agriculture footprint. Predicted future scenario based on transitional probabilities revealed a continual loss in natural habitat and increase in the agricultural footprint. Time series analysis results (principal components and temporal profiles) suggested that the landscape has a high degree of overall dynamic change with pronounced inter and intra-annual changes and there was an overall increase in greenness associated with increase in agricultural activity. The study concluded that without future conservation interventions natural habitats would continue to disappear, a condition that will impact heavily on biodiversity and significant waterdependent ecosystems such as wetlands. This has significant implications for the long-term provision of water from ground water reserves and for the overall sustainability of current agricultural practices. / South Africa

Panna 6 på Högdalenverket i södra Stockholm (P6) med tillhörande ångturbin producerar kraftvärme genom förbränning av utsorterat returbränsle från industri och samhälle. För att minimera underhållskostnader och öka anläggningens tillgänglighet är det viktigt att fel och oönskat processbeteende kan upptäckas i ett tidigt skede. I detta syfte testas här en metod för detektering av avvikande händelser med hjälp av principalkomponentanalys (PCA) på produktionsprocessen för kraftvärme. En PCA-modell med reducerad dimension skapas utifrån processdata från en problemfri driftperiod och används som mall för inkommande data att jämföras med i ett kontrolldigram. Avvikelser ifrån modellen bör vara en indikation på att ett onormalt drifttillstånd har uppkommit och orsaker till avvikelsen analyseras. Som avvikande händelse testas två fall av tubläckage som uppstod i ett av tubpaketen för kylning av rökgaserna under 2014 och 2015. Resultatet visar att processavvikelser ifrån normallägesmodellerna tydligt syns i kontrolldiagrammen vid båda tubläckagen och avvikelserna kan härledas till variabler som är kopplade till tubläckage. Det finns potential för att tillämpa metoden för övervakning av processen, en svårighet ligger i att skapa en modell som representerar processen när den är stabil på grund av att det finns många varierande driftfall som anses stabila, detta kräver vidare arbete. Metoden kan redan användas som analysverktyg exempelvis vid misstanke om tubläckage. / Boiler 6 at the Högdalen facility in southern Stockholm (P6) combined with a a steam turbine produces Combined Heat and Power (CHP) through combustion of treated industry waste. In order to minimise maintenance costs and increase plant availability it is of importance to detect process faults and deviations at an early state. In this study a method for outlier detection using Principal Component Analysis (PCA) is applied on the CHP production process. A PCA model with reduced dimension is created using process data from a problem free period and is used as a template for new operating data to be compared with in a control chart. Deviations from the model should be an indication of the presence of abnormal conditions and the reasons for the deviations are analysed. Two cases of tube failure in 2014 and 2015 are used to study the deviations. The result shows that process deviations from the models can be detected in the control chart in both cases of tube failure and the variables known to be associated with tube failure contributes highly to the deviating behaviour. There is potential for applying this method for process control, a difficulty lies in creating a model that represents the stable process when there are big variances within what is considererd a stable process state. The method can be used for data analysis when suspecting a tube failure.

Les exigences en matière de bruit rayonné par les navires de la Marine ou de recherche engendrent le développement de nouvelles méthodes pour améliorer leurs caractérisations. Le propulseur, qui est la source la plus importante en champ lointain, est généralement étudié en tunnel hydrodynamique. Cependant, compte tenu de la réverbération dans le tunnel et du niveau élevé du bruit de couche limite turbulente (CLT), la caractérisation peut s’ avérer délicate. L'objectif de la thèse est d'améliorer les capacités de mesures acoustiques du Grand Tunnel Hydrodynamique (GTH) de la DGA en matière de bruits émis par les maquettes testées dans des configurations d'écoulement.Un modèle de propagation basé sur la théorie des sources images est utilisé afin de prendre en compte le confinement du tunnel. Les coefficients de réflexion associés aux parois du tunnel sont identifiés par méthode inverse et à partir de la connaissance de quelques fonctions de transfert. Un algorithme de débruitage qui repose sur l’ Analyse en Composantes Principales Robuste est également proposé. Il s'agit de séparer, de manière aveugle ou semi-aveugle, l’ information acoustique du bruit de CLT en exploitant, respectivement, la propriété de rang faible et la structure parcimonieuse des matrices interspectrales du signal acoustique et du bruit. Ensuite, une technique d'imagerie basée sur la méthode des sources équivalentes est appliquée afin de localiser et quantifier des sources acoustiques corrélées ou décorrélées. Enfin, la potentialité des techniques proposées est évaluée expérimentalement dans le GTH en présence d'une source acoustique et d'un écoulement contrôlé. / The noise requirements for naval and research vessels lead to the development of new characterization methods. The propeller, which is the most important source in the far field, is usually studied in a water tunnel. However, due to the reverberation in the tunnel and the high level of flow noise, the characterization may be difficult. The aim of the thesis is to improve the measurement capabilities of the DGA Hydrodynamic tunnel (GTH) in terms of noise radiated by models in flow configurations.The propagation model is described through the image source method. Unfortunately, the reflection coefficients of the tunnel walls are generally unknown and it is proposed to estimate these parameters using an inverse method and the knowledge of some reference transfer functions. The boundary layer noise (BLN) may be stronger than the acoustic signal, therefore a Robust Principal Component Analysis is introduced in order to separate, blindly or semi-blindly, the acoustic signal from the noise. This algorithm is taking advantage of the low rank and sparse structure of the acoustic and the BLN cross-spectrum matrices. Then an acoustic imaging technique based on the equivalent source method is applied in order to localize and quantify correlated or decorrelated sources. Finally, the potentiality of the proposed techniques is evaluated experimentally in the GTH in the presence of an acoustic source and a controlled flow.

Les lésions résultantes d’un Accident Ischémique Cérébral (AIC) restent invisibles pendant plusieurs heures voire dizaines d’heures sur les techniques conventionnelles d’imagerie médicale (scanner X ou IRM). De nouvelles séquences d’IRM (pondérées en diffusion et en perfusion) permettent de visualiser les atteintes immédiatement. Les images de diffusion et de perfusion sont le reflet de mécanismes physiopathologiques complexes et contiennent aussi des informations sur le devenir tissulaire. Un modèle probabiliste d’évolution des lésions visibles au stade précoce a été développé de manière à prédire les extensions et rétrécissements des dommages. Le résultat se présente sous la forme d’une cartographie probabiliste de nécrose où la valeur de chaque voxel de l’image représente le risque de mort cellulaire au stade chronique. Cette cartographie pourra être, ensuite, prise en compte dans la décision thérapeutique. Deux types d’approche ont été mises en place : une méthode de segmentation, permettant d’extraire les régions à risque, et une méthode de prédiction des voxels à risque de nécrose, de type régression logistique. Ces modèles fournissent des performances similaires par rapport aux quelques modèles décrits dans la littérature, mais sont plus performants sur les petites lésions et présentent une grande amélioration en raison de leur caractère entièrement automatique / The lesions resulting from an ischemic stroke remain invisible for several hours or even tens of hours on conventional techniques of medical imaging (CT scan or MRI). New MRI sequences (diffusion-weighted and perfusion-weighted images) reveal the damages immediately. The diffusion and perfusion images reflect complex pathophysiological mechanisms and also contain some information on the fate of the tissue. A probabilistic model of evolution of the lesions visible at an early stage was set up so as to predict expansions and contractions of the damages. The result is presented as a probabilistic mapping of necrosis where the value of each voxel of the image represents the risk of cellular death in the chronic phase. This mapping can then be taken into account in the therapeutic decision. We implemented two types of approach: a segmentation method, allowing extracting regions atrisk, and a prediction method, type of logistic regression, to predict voxels at risk of necrosis. These models provide similar performances with regard to the few models described in the literature, but are more efficient for small lesions and show a big improvement because they are fully automatic

The present thesis proposes new methods for obtaining objective and accurate diagnoses in modern healthcare. Non-invasive techniques have been used to examine or diagnose three different medical conditions, namely neuropathy among diabetics, radiotherapy induced erythema (skin redness) among breast cancer patients and diagnoses of cutaneous malignant melanoma. The techniques used were Near-InfraRed spectroscopy (NIR), Multi Frequency Bio Impedance Analysis of whole body (MFBIA-body), Laser Doppler Imaging (LDI) and Digital Colour Photography (DCP). The neuropathy for diabetics was studied in papers I and II. The first study was performed on diabetics and control subjects of both genders. A separation was seen between males and females and therefore the data had to be divided in order to obtain good models. NIR spectroscopy was shown to be a viable technique for measuring neuropathy once the division according to gender was made. The second study on diabetics, where MFBIA-body was added to the analysis, was performed on males exclusively. Principal component analysis showed that healthy reference subjects tend to separate from diabetics. Also, diabetics with severe neuropathy separate from persons less affected. The preliminary study presented in paper III was performed on breast cancer patients in order to investigate if NIR, LDI and DCP were able to detect radiotherapy induced erythema. The promising results in the preliminary study motivated a new and larger study. This study, presented in papers IV and V, intended to investigate the measurement techniques further but also to examine the effect that two different skin lotions, Essex and Aloe vera have on the development of erythema. The Wilcoxon signed rank sum test showed that DCP and NIR could detect erythema, which is developed during one week of radiation treatment. LDI was able to detect erythema developed during two weeks of treatment. None of the techniques could detect any differences between the two lotions regarding the development of erythema. The use of NIR to diagnose cutaneous malignant melanoma is presented as unpublished results in this thesis. This study gave promising but inconclusive results. NIR could be of interest for future development of instrumentation for diagnosis of skin cancer.

In this research, hyperspectral and multispectral images were utilized for vegetation studies in the greenbelt corridor near Denton. EO-1 Hyperion was the hyperspectral image and Landsat Thematic Mapper (TM) was the multispectral image used for this research. In the first part of the research, both the images were classified for land cover mapping (after necessary atmospheric correction and geometric registration) using supervised classification method with maximum likelihood algorithm and accuracy of the classification was also assessed for comparison. Hyperspectral image was preprocessed for classification through principal component analysis (PCA), segmented principal component analysis and minimum noise fraction (MNF) transform. Three different images were achieved after these pre-processing of the hyperspectral image. Therefore, a total of four images were classified and assessed the accuracy. In the second part, a more precise and improved land cover study was done on hyperspectral image using linear spectral unmixing method. Finally, several vegetation constituents like chlorophyll a, chlorophyll b, caroteoids were distinguished from the hyperspectral image using feature-oriented principal component analysis (FOPCA) method and which component dominates which type of land cover particularly vegetation were correlated.

Otimizar simultaneamente os interesses dos usuários, anunciantes e publicadores é um grande desafio na área de publicidade computacional. Mais precisamente, a ordenação de anúncios, ou ad ranking, desempenha um papel central nesse desafio. Por outro lado, nem mesmo as melhores fórmulas ou algoritmos de ordenação são capazes de manter seu status por um longo tempo em um ambiente que está em constante mudança. Neste trabalho, apresentamos uma análise orientada a dados que mostra a importância de combinar diferentes dimensões de publicidade computacional por meio de uma abordagem evolutiva para ordenação de anúncios afim de responder a mudanças de forma mais eficaz. Nós avaliamos as dimensões de valor comercial, desempenho histórico de cliques, interesses dos usuários e a similaridade textual entre o anúncio e a página. Nessa avaliação, nós averiguamos o desempenho e a correlação das diferentes dimensões. Como consequência, nós desenvolvemos uma abordagem evolucionária para combinar essas dimensões. Essa abordagem é composta por três partes: um repositório de configurações para facilitar a implantação e avaliação de experimentos de ordenação; um componente evolucionário de avaliação orientado a dados; e um motor de programação genética para evoluir fórmulas de ordenação de anúncios. Nossa abordagem foi implementada com sucesso em um sistema real de publicidade computacional responsável por processar mais de quatorze bilhões de requisições de anúncio por mês. De acordo com nossos resultados, essas dimensões se complementam e nenhuma delas deve ser neglicenciada. Além disso, nós mostramos que a combinação evolucionária dessas dimensões não só é capaz de superar cada uma individualmente, como também conseguiu alcançar melhores resultados do que métodos estáticos de ordenação de anúncios. / Simultaneous optimization of users, advertisers and publishers\' interests has been a formidable challenge in online advertising. More concretely, ranking of advertising, or more simply ad ranking, has a central role in this challenge. However, even the best ranking formula or algorithm cannot withstand the ever-changing environment of online advertising for a long time. In this work, we present a data-driven analysis that shows the importance of combining different aspects of online advertising through an evolutionary approach for ad ranking in order to effectively respond to changes. We evaluated aspects ranging from bid values and previous click performance to user behavior and interests, including the textual similarity between ad and page. In this evaluation, we assessed commercial performance along with the correlation between different aspects. Therefore, we proposed an evolutionary approach for combining these aspects. This approach was composed of three parts: a configuration repository to facilitate deployment and evaluation of ranking experiments; an evolutionary data-based evaluation component; and a genetic programming engine to evolve ad ranking formulae. Our approach was successfully implemented in a real online advertising system that processes more than fourteen billion ad requests per month. According to our results, these aspects complement each other and none of them should be neglected. Moreover, we showed that the evolutionary combination of these aspects not only outperformed each of them individually, but was also able to achieve better overall results than static ad ranking methods.