Aplicação de componentes principais e regressões logísticas múltiplas em sistema de informações geográficas para a predição e o mapeamento digital de solos

• Ano de defesa: 2008
• Autor(a) principal: Caten, Alexandre Ten
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; BR; Agronomia; UFSM; Programa de Pós-Graduação em Ciência do Solo
• Palavras-chave em Português: Pedometria; Mapeamento preditivo; Análise estatística multivariada; Modelos logísticos politómicos; Índice de umidade topográfica; Pedometry; Predictive mapping; Statistical multivariate analysis; Polytomous logistic regression; Topographic wetness index; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA
• Link de acesso: http://repositorio.ufsm.br/handle/1/5483

Social demands on soil information have grown dramatically, meanwhile the soil surveys are seldom carried out in the country. Digital soil mapping techniques can be applied to infer the spatial distribution of soil from existing soil maps or from reference areas, extrapolating this information to areas not mapped. The purpose of this study was to apply in a Geographic Information System the Multiple Logistic Regressions (MLR) using Principal Components (PC) as explanatory variables to predict soil classes spatial distribution. The study area was the region of municipality São Pedro do Sul / RS. For the development of predictive models a set of nine terrain attributes were used. Model training was executed on an existing soil map and with a survey carried out in a reference area, both in a 1:50.000 scale. The first three retained PC explained 65.57% of the data variability. The predictive models which used PC had lower values of kappa index. The most accurate predicted map reached a kappa value of 63.20% and was generated by using the nine attributes of land as predictive covariates. The mapping accuracy is sensitive to similarities between the mapped classes, and mapping in a more homogeneous categorical level reduces the accuracy of the predicted maps. Soil classes relatively not representative in the training maps are not properly spatialized. The use of MLR allows spatializing of soil classes to areas not mapped, although the use of PC needs to be tested with a larger number of covariates.