Mapeamento digital de solos por redes neurais artificiais da bacia hidrogr?fica do rio Guapi-Macacu, RJ

• Ano de defesa: 2012
• Autor(a) principal: Pinheiro, Helena Saraiva Koenow
• Orientador(a): Anjos, L?cia Helena Cunha dos
• Banca de defesa: Fernandes Filho, Elp?dio In?cio, Vasques, Gustavo de Mattos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal Rural do Rio de Janeiro
• Programa de Pós-Graduação: Programa de P?s-Gradua??o em Agronomia - Ci?ncia do Solo
• Departamento: Instituto de Agronomia
• País: Brasil
• Palavras-chave em Português: Modelagem digital; Pedologia; Pedometria; Digital modelling. Pedology. Pedometry
• Área do conhecimento CNPq: Agronomia
• Link de acesso: https://tede.ufrrj.br/jspui/handle/jspui/3644

The soil survey includes a description of the morphological, physical and chemical properties of soils in a given area, including technical description and interdisciplinary information, and supporting various purposes. The digital soil mapping makes possible to improve the products of pedological surveys through the use of geographic information system (GIS) tools, and knowledge of the soil genesis, morphology and classification, in regards the choice of attributes that represent a consistent approach to the reality, and seeking to produce a precise soil survey with efficiency (cost x time). The overall objective of this study was to characterize the landscape components to subsidize the prediction of soil classes, using artificial neural networks (ANN's), to produce the digital soil map of the Rio Guapi-Macacu watershed, located in Rio de Janeiro State. The mapping involved 100 sample points which were analysed for chemical and physical properties and morphological described, according to standard soil surveys methods in Brazil. The digital soil mapping involved the acquisition of cartographic database, creating digital models that represent landscape attributes relevant to pedogenesis and soil classification, analysis of soil-landscape relationships, and ultimately supervised classification by ANN and validation of the mapping result. Given the approach used and the processing tools of the data available, the digital elevation models (DEM's) were analysed, for the spatial resolution and how to acquire the model, to select the appropriate DEM to derive the morphometric attributes. The analyses showed as having superior quality the DEM with appropriate spatial resolution of 30m cell size, obtained by interpolation of elevation data, contour lines and elevation points, SRTM sensor data. After defining the MDE and derivation of attributes, a study was made for recognition of patterns and geomorphic characterization of pedoenvironment involving sample collection and description of profiles in predefined locations, through the program Hypercube Latin Conditioned Sampling (cLHS). The predominant soils were: Oxisols, Ultisols, Inceptisols, Aquents and other Entisols. The usage of GIS tools allowed the selection of variables to compose sets of discriminants used in the classification stage by ANN. The selected variables were: altimetry, slope, curvature, combined topographic index, euclidean distance, clay minerals, iron oxide, NDVI and geology. There were trained eleven sets of ANN's, with different combinations as to the discriminating variables (input layer). The criteria used in evaluating the performance of the ANN were the rates of overall accuracy and Kappa, considering the generalization of output classes. Validation was performed using 120 control points corresponding to the soil profiles that were not used for ANN training. After the analysis of the mean square error of the different sets, the architecture with 10 hidden nodes was chosen. The evaluation criteria of classification underlined the best networks performance of the sets 1, 7 and 10, corresponding to all variables, excluding the geology, and excluding the index NDVI, which did not differ. The generalization and validation showed that the set 10 allowed the best final classification.