Remotely-sensed hyperspectral imagery has become one the most advanced tools for analyzing the processes that shape the Earth and other planets. Effective, rapid analysis of high-volume, high-dimensional hyperspectral image data sets demands efficient, automated techniques to identify signatures of known materials in such imagery. In this thesis, we develop a framework for automatic material identification in hyperspectral imagery using adaptive similarity measures. We frame the material identification problem as a multiclass similarity-based classification problem, where our goal is to predict material labels for unlabeled target spectra based upon their similarities to source spectra with known material labels. As differences in capture conditions affect the spectral representations of materials, we divide the material identification problem into intra-domain (i.e., source and target spectra captured under identical conditions) and inter-domain (i.e., source and target spectra captured under different conditions) settings. The first component of this thesis develops adaptive similarity measures for intra-domain settings that measure the relevance of spectral features to the given classification task using small amounts of labeled data. We propose a technique based on multiclass Linear Discriminant Analysis (LDA) that combines several distinct similarity measures into a single hybrid measure capturing the strengths of each of the individual measures. We also provide a comparative survey of techniques for low-rank Mahalanobis metric learning, and demonstrate that regularized LDA yields competitive results to the state-of-the-art, at substantially lower computational cost. The second component of this thesis shifts the focus to inter-domain settings, and proposes a multiclass domain adaptation framework that reconciles systematic differences between spectra captured under similar, but not identical, conditions. Our framework computes a similarity-based mapping that captures structured, relative relationships between classes shared between source and target domains, allowing us apply a classifier trained using labeled source spectra to classify target spectra. We demonstrate improved domain adaptation accuracy in comparison to recently-proposed multitask learning and manifold alignment techniques in several case studies involving state-of-the-art synthetic and real-world hyperspectral imagery.

Orientadores: Rubens Augusto Camargo Lamparelli, Jansle Vieira Rocha / Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Engenharia Agrícola / Made available in DSpace on 2018-08-22T16:59:27Z (GMT). No. of bitstreams: 1 Bispo_RafaelCarlos_M.pdf: 6682271 bytes, checksum: 3672bc3cd1d328c4edf329eabbd31ffe (MD5) Previous issue date: 2013 / Resumo: A produção de café esteve intimamente ligada ao desenvolvimento econômico do Brasil e ainda hoje o café é um importante produto da agricultura nacional. O Estado de Minas Gerais responde atualmente por 52% de toda a área de café do Brasil. Dessa forma, dada a importância da cafeicultura para a economia brasileira, é necessário desenvolver e melhorar as metodologias para seu monitoramento. Dados de sensoriamento remoto podem fornecer informações para o monitoramento e o mapeamento de café de maneira mais rápida e menos onerosa do que os métodos convencionais. Nesse contexto, os objetivos desta pesquisa foram identificar a bienalidade da cultura de café por meio de dados do sensor MODIS, juntamente com dados de estações meteorológicas, entre os anos de 2004 a 2012, e avaliar a eficácia das imagens-fração derivadas do sensor MODIS no mapeamento automático das áreas de café do município de Monte Santo de Minas/MG. Foi utilizada uma série temporal com 163 imagens da banda NIR do MODIS, produto MOD13Q1, para se extrair os valores de refletância dos pixels com pelo menos 80% de café. Dados diários de temperatura e precipitação foram agrupados de acordo com a resolução temporal das imagens (16 dias) para o cálculo do balanço hídrico. Para o mapeamento das áreas de café, foram utilizadas imagens do MODIS, bandas MIR, NIR e RED, dos períodos seco e chuvoso. Através do Modelo Linear de Mistura Espectral foram derivadas imagens-fração de solo, café e água/sombra. Estas imagens-fração serviram como dados de entrada para a classificação automática supervisionada com o método SVM - Support Vector Machine. Os resultados mostraram que para o monitoramento do café os dados de refletância dos períodos de colheita apresentaram maior correlação com a alternância da quantidade da produção. A partir da matriz de erro montada entre as classificações e as máscaras de referência, observou-se que os melhores resultados de Exatidão Global e Índice Kappa foram obtidos na classificação do período seco, sendo 67% e 0,41, respectivamente. Análises estatísticas de correlação e coeficiente de variação aplicadas sobre as imagens-fração de café permitiram melhor compreensão da complexidade do mapeamento do café / Abstract: Coffee production was closely linked to the economic development of Brazil and even today coffee is an important product of national agriculture. The State of Minas Gerais currently accounts for 52% of the whole area of coffee in Brazil. Thus, given the importance of the coffee crops to Brazilian economy, it is necessary to develop and improve methodologies for its monitoring. Then, remote sensing data can provide information for monitoring and mapping of coffee crops faster and cheaper than conventional methods. In this context, the objectives of this study were to identify the biennial yield of the coffee crop using data from MODIS and meteorological stations, over the period between 2004 and 2012, and assess the effectiveness of the fraction-images derived from MODIS in the automatic mapping of the areas of coffee in Monte Santo de Minas/MG. Were used a time series of 163 images of NIR band from MODIS, MOD13Q1 product, to extract the values of reflectance of pixels with at least 80% of coffee. Daily data of air temperature and precipitation were compiled to 16-day intervals to match the temporal resolution of MODIS imagery and to calculate the water balance. For coffee mapping, we used MODIS imagery, MIR, NIR and RED bands, of dry and rainy seasons. Through the Spectral Linear Mixing Model were derived fraction images of soil, coffee and water/shadow. These fraction images served as input data for supervised classification with SVM - Support Vector Machine approach. The results showed that for coffee monitoring the reflectance data of harvest period presented higher correlation with the alternation of coffee production. From the error matrix between the classifications and reference masks, it was observed that the best results of Overall Accuracy and Kappa Index were obtained in the classification of the dry season, with 67% and 0.41, respectively. Statistical analyses of correlation and coefficient of variation applied over images fraction of coffee allowed a better understanding about the complexity of mapping coffee / Mestrado / Planejamento e Desenvolvimento Rural Sustentável / Mestre em Engenharia Agrícola

No description available.

Invasive Pflanzenarten, die sich durch hohe Vermehrungsratem und breite Umwelttoleranz auszeichnen, verdrängen oft einheimische Flora, was zu einem Verlust an biologischer Vielfalt und zu wirtschaftlichen Kosten führt. Aufgrund des zunehmenden globalen Handels und des Klimawandels wird erwartet, dass sich die Dynamik biologischer Invasionen in Zukunft intensivieren wird. In Deutschland nimmt die Zahl der invasiven oder potenziell invasiven Pflanzenarten zu, deren Ausbreitung erhebliche ökologische Schäden verursacht. Ziel dieser Arbeit ist die Entwicklung prädiktiver Habitatmodelle, welche sowohl klimatische und andere abiotische Faktoren als auch Habitattypen als Umweltvariablen integrieren. Die Habitattypen wurden aus einer fernerkundlichen Klassifikation des Zielgebiets unter Verwendung von MODIS-Oberflächenreflexionsdaten und einem Ensemble von drei Machine-Learning-Algorithmen abgeleitet. Dabei wurden naturnahe Gebiete anhand ihrer spektralen Ähnlichkeit mit Natura-2000-Lebensraumtypen einer von 18 Klassen zugeordnet. Die Klassifizierung ermöglichte es, quantitative Informationen über die räumliche Verteilung von Lebensräumen zu gewinnen und zwischen Störungsereignissen und graduellen Verschiebungen der Habitatzusammensetzung zu unterscheiden. Im Anschluss wurden Artverbreitungsmodelle für 46 invasive Pflanzenarten entwickelt, um potenziell geeignete Lebensräume unter gegenwärtigen und zukünftigen klimatischen Bedingungen mithilfe zweier Machine-Learning-Algorithmen zu ermitteln. Als Prädiktorvariablen wurden insgesamt 18 Umweltfaktoren verwendet, darunter Bodentyp, Höhe, Landnutzung, Habitattyp, Verkehrsinfrastruktur, Temperatur und Niederschlag. Die Modelle legen nahe, dass die Mehrheit der Arten ihren potenziell geeigneten Lebensraum noch nicht besetzt hat. Unter zukünftigen Klimabedingungen ist für die meisten Arten eine Zunahme der Habitateignung zu erwarten. Dabei üben Klimafaktoren, Bodentyp und Verkehrsinfrastruktur den größten relativen Einfluss auf die Artverbreitung aus. Der Habitattyp ist lediglich für eine geringe Zahl von Arten von Bedeutung, was darauf hindeutet, dass es nur begrenzt möglich ist, interspezifische Konkurrenz oder nicht näher spezifizierte biotische Interaktionen in diesem Maßstab zu berücksichtigen. Monitoring-Systeme stellen das kosteneffizienteste Mittel zur Verhinderung der Ausbreitung invasiver Arten dar und sind somit ein entscheidendes Element für deren Management. Auf Basis der Ergebnisse dieser Arbeit wurde eine öffentlich zugängliche Monitoring-Website entwickelt, IPS Monitor, welche digitale Webkarten sowie Pflanzensteckbriefe beinhaltet. / Invasive plant species, characterised by high reproductive capacity and broad environmental tolerance, often outcompete native flora, leading to biodiversity loss and economic costs. Due to increasing global trade and climate change, the dynamics of biological invasions are expected to increase in the future. In Germany, the number of invasive or potentially invasive plant species is increasing, causing significant damage due to their aggressive spread and ecological threats. This thesis aims to develop predictive habitat models that integrate climatic and other abiotic factors as well as habitat types as environmental variables. Habitat types were derived from a remote sensing classification of the target area using MODIS surface reflectance data and an ensemble of three machine learning algorithms, assigning natural areas to one of 18 classes based on their similarity to Natura 2000 habitat types. The classification framework was able to provide quantitative information on the spatial distribution of habitats and differentiate between disturbance events and gradual shifts in ecosystem composition. Large-scale species distribution models of 46 invasive plant species were then developed to identify the potentially suitable habitat under present and future climatic conditions using two machine learning techniques. A total of 18 environmental variables were used as predictor variables, including soil type, altitude, land use, habitat type, transport infrastructure, temperature and precipitation. The models indicate that the majority of species have not yet occupied their potentially suitable habitat. An increase in habitat suitability under predicted climatic conditions is implied for most species, with climate factors, soil type and transport infrastructure having the greatest relative influence on habitat suitability. Habitat type is only relevant for few species, indicating a limited ability to map interspecific competition or unspecified biotic interactions at this spatial scale. As the most cost-effective means of preventing the spread of invasive species, monitoring systems represent a crucial element in their management and enable early detection and rapid response mechanisms. Therefore, a publicly accessible monitoring tool, IPS Monitor, including digital web maps and plant fact sheets has been developed from the results of this thesis.