Classificação supervisionada de solos por redes neurais artificiais na Serra do Cipó - MG

• Ano de defesa: 2009
• Autor(a) principal: Souza, Eliana de
• Orientador(a): Fernandes Filho, Elpídio Inácio
• Banca de defesa: Vieira, Carlos Antonio Oliveira , Chagas, César da Silva , Simas, Felipe Nogueira Bello
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Viçosa
• Programa de Pós-Graduação: Mestrado em Solos e Nutrição de Plantas
• Departamento: Fertilidade do solo e nutrição de plantas; Gênese, Morfologia e Classificação, Mineralogia, Química,
• País: BR
• Palavras-chave em Português: Mapemaneto digital de solos; Pedologia; Redes neurais artificiais
• Palavras-chave em Inglês: Digital mapping of soil; Pedology; Artificial neural networks
• Área do conhecimento CNPq: CNPQ::CIENCIAS AGRARIAS::AGRONOMIA::CIENCIA DO SOLO
• Link de acesso: http://locus.ufv.br/handle/123456789/5456

The supervised classification of soils, especially in recent decades, is being carried out using mathematical and statistical models, amongst which the model of neural networks, stand out by greater accuracy of mapping comparing to classical models, such as Maximum Likelihood (MaxVer), helping the conventional method of mapping. Neural Networks model has been performed mostly for soil properties, with little application for soil classes. This work aimed to undertake the classification of soil using neural networks and MaxVer for an area located in Serra do Cipó, in the State of Minas Gerais. The map units was defined based on information from 55 soil profiles classified accord to the Brazilian System of Soils Classification to the fourth categorical level, whereas compound units were done with basis on the similarity of soil properties and the characteristics of physics environment. The discriminates variables used included: six scenes of Landsat satellite image sensor ETM+, four indexes derived from this image (Clay minerals, Ferrous minerals, Iron oxide and NDVI), Digital Elevation Model and derived attributes: altitude, slope, compound topographic index, aspect, solar radiation, curvature and elevation amplitude, in addiction to geological and soil maps. Several variables combinations were tested in both classifiers, selecting those that best contribute to classify the soil with high accuracy on the two supervised mapping approaches. The classification by neural networks was performed using the Stuttgart Neural Network Simulator and the backpropagation algorithm, the framework and classification parameters were selected by training and statistical tests. The results obtained with both classifiers, neural networks and MaxVer were compared using ground data as reference. The same set of reference points was used to validate the soil map obtained by the conventional method of mapping. Maps obtained by the two classifiers using the group of variables that provided the best performance to the classification showed a good accuracy index, with no statistical difference in overall accuracy of the maps. The map generated by MaxVer showed a kappa index of 0.58, while the map from neural network showed an index of 0.60. Although the accuracy of the two maps was statistically similar, the classifiers efficiency in individual discrimination of soil units differed significantly, with two units being best classified by MaxVer, three units by neural networks and four units with similar accuracies in both approach classification. The overall accuracy of soil maps made by the conventional method was 82%. The soils of the first component in mapping units agreed in 48% with reference soils.