Advanced Kernel Methods For Remote Sensing Image Classification

Résumé Suite aux recentes avancées technologiques, les archives d'images digitales ont connu une croissance qualitative et quantitative sans précédent. Malgré les énormes possibilités qu'elles offrent, ces avancées posent de nouvelles questions quant au traitement des masses de données saisies. Cette question est à la base de cette Thèse: les problèmes de traitement d'information digitale à très haute résolution spatiale et/ou spectrale y sont considérés en recourant à des approches d'apprentissage statistique, les méthodes à noyau. Cette Thèse étudie des problèmes de classification d'images, c'est à dire de catégorisation de pixels en un nombre réduit de classes refletant les propriétés spectrales et contextuelles des objets qu'elles représentent. L'accent est mis sur l'efficience des algorithmes, ainsi que sur leur simplicité, de manière à augmenter leur potentiel d'implementation pour les utilisateurs. De plus, le défi de cette Thèse est de rester proche des problèmes concrets des utilisateurs d'images satellite sans pour autant perdre de vue l'intéret des méthodes proposées pour le milieu du machine learning dont elles sont issues. En ce sens, ce travail joue la carte de la transdisciplinarité en maintenant un lien fort entre les deux sciences dans tous les développements proposés. Quatre modèles sont proposés: le premier répond au problème de la haute dimensionalité et de la redondance des données par un modèle optimisant les performances en classification en s'adaptant aux particularités de l'image. Ceci est rendu possible par un système de ranking des variables (les bandes) qui est optimisé en même temps que le modèle de base: ce faisant, seules les variables importantes pour résoudre le problème sont utilisées par le classifieur. Le manque d'information étiquétée et l'incertitude quant à sa pertinence pour le problème sont à la source des deux modèles suivants, basés respectivement sur l'apprentissage actif et les méthodes semi-supervisées: le premier permet d'améliorer la qualité d'un ensemble d'entraînement par interaction directe entre l'utilisateur et la machine, alors que le deuxième utilise les pixels non étiquetés pour améliorer la description des données disponibles et la robustesse du modèle. Enfin, le dernier modèle proposé considère la question plus théorique de la structure entre les outputs: l'intègration de cette source d'information, jusqu'à présent jamais considérée en télédétection, ouvre des nouveaux défis de recherche. Advanced kernel methods for remote sensing image classification Devis Tuia Institut de Géomatique et d'Analyse du Risque September 2009 Abstract The technical developments in recent years have brought the quantity and quality of digital information to an unprecedented level, as enormous archives of satellite images are available to the users. However, even if these advances open more and more possibilities in the use of digital imagery, they also rise several problems of storage and treatment. The latter is considered in this Thesis: the processing of very high spatial and spectral resolution images is treated with approaches based on data-driven algorithms relying on kernel methods. In particular, the problem of image classification, i.e. the categorization of the image's pixels into a reduced number of classes reflecting spectral and contextual properties, is studied through the different models presented. The accent is put on algorithmic efficiency and the simplicity of the approaches proposed, to avoid too complex models that would not be used by users. The major challenge of the Thesis is to remain close to concrete remote sensing problems, without losing the methodological interest from the machine learning viewpoint: in this sense, this work aims at building a bridge between the machine learning and remote sensing communities and all the models proposed have been developed keeping in mind the need for such a synergy. Four models are proposed: first, an adaptive model learning the relevant image features has been proposed to solve the problem of high dimensionality and collinearity of the image features. This model provides automatically an accurate classifier and a ranking of the relevance of the single features. The scarcity and unreliability of labeled. information were the common root of the second and third models proposed: when confronted to such problems, the user can either construct the labeled set iteratively by direct interaction with the machine or use the unlabeled data to increase robustness and quality of the description of data. Both solutions have been explored resulting into two methodological contributions, based respectively on active learning and semisupervised learning. Finally, the more theoretical issue of structured outputs has been considered in the last model, which, by integrating outputs similarity into a model, opens new challenges and opportunities for remote sensing image processing.

Prognostic relevance of Masaoka and Müller-Hermelink classification in patients with thymic tumors.

BACKGROUND: To compare the prognostic relevance of Masaoka and Müller-Hermelink classifications. METHODS: We treated 71 patients with thymic tumors at our institution between 1980 and 1997. Complete follow-up was achieved in 69 patients (97%) with a mean follow up-time of 8.3 years (range, 9 months to 17 years). RESULTS: Masaoka stage I was found in 31 patients (44.9%), stage II in 17 (24.6%), stage III in 19 (27.6%), and stage IV in 2 (2.9%). The 10-year overall survival rate was 83.5% for stage I, 100% for stage IIa, 58% for stage IIb, 44% for stage III, and 0% for stage IV. The disease-free survival rates were 100%, 70%, 40%, 38%, and 0%, respectively. Histologic classification according to Müller-Hermelink found medullary tumors in 7 patients (10.1%), mixed in 18 (26.1%), organoid in 14 (20.3%), cortical in 11 (15.9%), well-differentiated thymic carcinoma in 14 (20.3%), and endocrine carcinoma in 5 (7.3%), with 10-year overall survival rates of 100%, 75%, 92%, 87.5%, 30%, and 0%, respectively, and 10-year disease-free survival rates of 100%, 100%, 77%, 75%, 37%, and 0%, respectively. Medullary, mixed, and well-differentiated organoid tumors were correlated with stage I and II, and well-differentiated thymic carcinoma and endocrine carcinoma with stage III and IV (p < 0.001). Multivariate analysis showed age, gender, myasthenia gravis, and postoperative adjuvant therapy not to be significant predictors of overall and disease-free survival after complete resection, whereas the Müller-Hermelink and Masaoka classifications were independent significant predictors for overall (p < 0.05) and disease-free survival (p < 0.004; p < 0.0001). CONCLUSIONS: The consideration of staging and histology in thymic tumors has the potential to improve recurrence prediction and patient selection for combined treatment modalities.

Classification of urban multi-angular image sequences by aligning their manifolds

When dealing with multi-angular image sequences, problems of reflectance changes due either to illumination and acquisition geometry, or to interactions with the atmosphere, naturally arise. These phenomena interplay with the scene and lead to a modification of the measured radiance: for example, according to the angle of acquisition, tall objects may be seen from top or from the side and different light scatterings may affect the surfaces. This results in shifts in the acquired radiance, that make the problem of multi-angular classification harder and might lead to catastrophic results, since surfaces with the same reflectance return significantly different signals. In this paper, rather than performing atmospheric or bi-directional reflection distribution function (BRDF) correction, a non-linear manifold learning approach is used to align data structures. This method maximizes the similarity between the different acquisitions by deforming their manifold, thus enhancing the transferability of classification models among the images of the sequence.

Serrated polyps of the colon: how reproducible is their classification?

For several years, the lack of consensus on definition, nomenclature, natural history, and biology of serrated polyps (SPs) of the colon has created considerable confusion among pathologists. According to the latest WHO classification, the family of SPs comprises hyperplastic polyps (HPs), sessile serrated adenomas/polyps (SSA/Ps), and traditional serrated adenomas (TSAs). The term SSA/P with dysplasia has replaced the category of mixed hyperplastic/adenomatous polyps (MPs). The present study aimed to evaluate the reproducibility of the diagnosis of SPs based on currently available diagnostic criteria and interactive consensus development. In an initial round, H&E slides of 70 cases of SPs were circulated among participating pathologists across Europe. This round was followed by a consensus discussion on diagnostic criteria. A second round was performed on the same 70 cases using the revised criteria and definitions according to the recent WHO classification. Data were evaluated for inter-observer agreement using Kappa statistics. In the initial round, for the total of 70 cases, a fair overall kappa value of 0.318 was reached, while in the second round overall kappa value improved to moderate (kappa = 0.557; p < 0.001). Overall kappa values for each diagnostic category also significantly improved in the final round, reaching 0.977 for HP, 0.912 for SSA/P, and 0.845 for TSA (p < 0.001). The diagnostic reproducibility of SPs improves when strictly defined, standardized diagnostic criteria adopted by consensus are applied.

Monitoring network optimisation for spatial data classification using support vector machines

The paper presents a novel method for monitoring network optimisation, based on a recent machine learning technique known as support vector machine. It is problem-oriented in the sense that it directly answers the question of whether the advised spatial location is important for the classification model. The method can be used to increase the accuracy of classification models by taking a small number of additional measurements. Traditionally, network optimisation is performed by means of the analysis of the kriging variances. The comparison of the method with the traditional approach is presented on a real case study with climate data.

Characterization and classification of matrix effects in biological samples analyses.

An exhaustive classification of matrix effects occurring when a sample preparation is performed prior to liquid-chromatography coupled to mass spectrometry (LC-MS) analyses was proposed. A total of eight different situations were identified allowing the recognition of the matrix effect typology via the calculation of four recovery values. A set of 198 compounds was used to evaluate matrix effects after solid phase extraction (SPE) from plasma or urine samples prior to LC-ESI-MS analysis. Matrix effect identification was achieved for all compounds and classified through an organization chart. Only 17% of the tested compounds did not present significant matrix effects.

Speciation of naturally-accumulated uranium in an organic-rich soil of an alpine region (Switzerland)

Very high concentrations of uranium (up to 4000 ppm) were found in a natural soil in the Dischma valley, an alpine region in the Grisons canton in Switzerland. The goal of this study was to examine the redox state and the nature of uranium binding in the soil matrix in order to understand the accumulation mechanism. Pore water profiles collected from Dischma soil revealed the establishment of anoxic conditions with increasing soil depth. A combination of chemical extraction methods and spectroscopy was applied to characterize the redox state and binding environment of uranium in the soil. Bicarbonate extraction under anoxic conditions released most of the uranium indicating that uranium occurs predominantly in the hexavalent form. Surprisingly, the uranium redox state did not vary greatly as a function of depth. X-ray absorption near edge spectroscopy (XANES), confirmed that uranium was present as a mixture of U(VI) and U(IV) with U(VI) dominating. Sequential extractions of soil samples showed that the dissolution of solid organic matter resulted in the simultaneous release of the majority of the soil uranium content (>95%). Extended X-ray absorption fine structure (EXAFS) spectroscopy also revealed that soil-associated uranium in the soil matrix was mainly octahedrally coordinated, with an average of 1.7 axial (at about 1.76 Å) and 4.6 to 5.3 equatorial oxygen atoms (at about 2.36 Å) indicating the dominance of a uranyl-like (UO22+) structure presumably mixed with some U(IV). An additional EXAFS signal (at about 3.2 Å) identified in some spectra suggested that uranium was also bound (via an oxygen atom) to a light element such as carbon, phosphorus or silicon. Gamma spectrometric measurements of soil profiles failed to identify uranium long-life daughter products in the soil which is an indication that uranium originates elsewhere and was transported to its current location by water. Finally, it was found that the release of uranium from the soil was significantly promoted at very low pH values (pH 2) and increased with increasing pH values (between pH 5 and 9).

Typologies textuelles et partitions musicales : dissimilarités, classification et autocorrélation

Axée dans un premier temps sur le formalisme et les méthodes, cette thèse est construite sur trois concepts formalisés: une table de contingence, une matrice de dissimilarités euclidiennes et une matrice d'échange. À partir de ces derniers, plusieurs méthodes d'Analyse des données ou d'apprentissage automatique sont exprimées et développées: l'analyse factorielle des correspondances (AFC), vue comme un cas particulier du multidimensional scaling; la classification supervisée, ou non, combinée aux transformations de Schoenberg; et les indices d'autocorrélation et d'autocorrélation croisée, adaptés à des analyses multivariées et permettant de considérer diverses familles de voisinages. Ces méthodes débouchent dans un second temps sur une pratique de l'analyse exploratoire de différentes données textuelles et musicales. Pour les données textuelles, on s'intéresse à la classification automatique en types de discours de propositions énoncées, en se basant sur les catégories morphosyntaxiques (CMS) qu'elles contiennent. Bien que le lien statistique entre les CMS et les types de discours soit confirmé, les résultats de la classification obtenus avec la méthode K- means, combinée à une transformation de Schoenberg, ainsi qu'avec une variante floue de l'algorithme K-means, sont plus difficiles à interpréter. On traite aussi de la classification supervisée multi-étiquette en actes de dialogue de tours de parole, en se basant à nouveau sur les CMS qu'ils contiennent, mais aussi sur les lemmes et le sens des verbes. Les résultats obtenus par l'intermédiaire de l'analyse discriminante combinée à une transformation de Schoenberg sont prometteurs. Finalement, on examine l'autocorrélation textuelle, sous l'angle des similarités entre diverses positions d'un texte, pensé comme une séquence d'unités. En particulier, le phénomène d'alternance de la longueur des mots dans un texte est observé pour des voisinages d'empan variable. On étudie aussi les similarités en fonction de l'apparition, ou non, de certaines parties du discours, ainsi que les similarités sémantiques des diverses positions d'un texte. Concernant les données musicales, on propose une représentation d'une partition musicale sous forme d'une table de contingence. On commence par utiliser l'AFC et l'indice d'autocorrélation pour découvrir les structures existant dans chaque partition. Ensuite, on opère le même type d'approche sur les différentes voix d'une partition, grâce à l'analyse des correspondances multiples, dans une variante floue, et à l'indice d'autocorrélation croisée. Qu'il s'agisse de la partition complète ou des différentes voix qu'elle contient, des structures répétées sont effectivement détectées, à condition qu'elles ne soient pas transposées. Finalement, on propose de classer automatiquement vingt partitions de quatre compositeurs différents, chacune représentée par une table de contingence, par l'intermédiaire d'un indice mesurant la similarité de deux configurations. Les résultats ainsi obtenus permettent de regrouper avec succès la plupart des oeuvres selon leur compositeur.