Uso de informações contextuais no processo de classificação de imagens do sensoriamento remoto

• Ano de defesa: 2008
• Autor(a) principal: Assis, Leonardo Campos de
• Orientador(a): Vieira, Carlos Antonio Oliveira
• Banca de defesa: Gleriani, José Marinaldo , Silva, Fabyano Fonseca e , Oliveira, Leonardo Castro de , Vasconcellos, José Carlos Penna
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Viçosa
• Programa de Pós-Graduação: Mestrado em Engenharia Civil
• Departamento: Geotecnia; Saneamento ambiental
• País: BR
• Palavras-chave em Português: Inferência Bayesiana; Regressão logística binomial; Imagens fundamentadas em valores de probabilidade
• Palavras-chave em Inglês: Bayesian model; Binomial Logistic Regression; Imagery founded on probability values
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS::ENGENHARIA CIVIL
• Link de acesso: http://locus.ufv.br/handle/123456789/3706

The exclusive use of spectral information applied to image classification has shown less efficiency through the complex conditions that technological incoming of earth observation offers. For this reason, the research presented on this dissertation purposes to consider the contextual information to help in the remote sensing digital image classification process. To reveal the purpose viability, the methodology was applied in the study case for determination of vegetation areas at Belo Horizonte Municipal District. Through the partnership with some organisms of the Belo Horizonte Administration, the spatial data providers, informational classes to be discovered by the process was determined. The data were composed by vector information layers and QuickBird 2 artificial satellite multispectral images. After the edition of part of the vector layers data and subsequent conversion to raster format, combined to the atmospheric effects attenuation process and imagery orthorrectification, the application of the method was initiated. At first, the helpful contextual information to map categories identification was established, following by the necessary procedures to get them. Two different approaches to deal with the context modeling were chosen: one direct and other indirect one. The direct modeling was featured by images intersection operations with Boolean data types, and, the indirect modeling offered the analyst option to operate continuous data types. The direct modeling was applied to identify the incidence of vegetation areas on streets and blocks, obtained from the spectral environmental variable NDVI. And then, an intersection operation was performed between vegetation image, the streets and blocks layers. To apply the indirect modeling, the contextual information were defined to better featuring the vegetation types (native forest, crop forest, savannah, savannah field and crop field), in terms of the topographic environmental variables HCDEM (Hydrologically Consistent Digital Elevation Model), SDM (Slope Digital Model), NDM (North face Digital Model also called Northness), and EDM (East face Digital Model also called Eastness); and the spectral environmental variable NDVI. For each informational class, one specific arrangement of contextual information was determined and used on a Binomial Logistic Regression procedure, to find out initial values of higher likelihood for class happening. In this way, images was generated, one for each class, with defined likelihood values defined by regression operation. The Bayesian Paradigm was applied, through the Bayesian probabilistic model, to obtain images founded on updated probability values based on specialist knowledge. Beta-Binomial was the Bayesian probabilistic model adopted. Model index was done by α and β hyperparameters values, expressing the analyst opinion about the distribution position and dispersion, respectively. Values assigned for the hyperparameters has been acquired by simulations and iterations, following by intermediate results analysis, successively made until a proper reality representation was obtained. Imagery founded on probability values, determined by Bayesian model, was used like initial probability values in the maximum likelihood classification method. To verify the efficiency of the contextual information inclusion method, named Contextual Classification process, it was compared against the Traditional Classification method by the Maximum Likelihood Classifier algorithm. Results of both procedures were evaluated by contingency matrixes build by comparison between the thematic imagery produced by the classification methods and the reference image. Reference image was generated from the training sites, refined (or purified) by the Mahalanobis statistical distance algorithm, employing a 50% arbitrary value threshold of similarity criterion. From the contingency matrixes, Kappa coefficient value was estimated, presenting a difference about 6.5% between the Classification methods, indicating superiority of the Contextual (0.9199) one, in relation to Traditional (0.8528) one. To certify that the difference was, although, significant, a two-sided Z test was applied among the methods at 5% significance level. By Z values obtained, was observed that, at 95% confidence level, the nullity hypothesis (equivalency between methods) must been rejected and, thus, the methods were verified different. Consequently, the Contextual Classification method was employed to generate the vegetation thematic image of Belo Horizonte (MG) Municipal District. Results show that, from the available data and through the simplification hypothesis assumed, the Contextual Classification method purposed was really superior to the Traditional Classification method and, thus, its application is recommended.