Uso de redes neurais artificiais no mapeamento de solos na Bacia do Rio Turvo Sujo - Viçosa MG

• Autor(a) principal: Melo, Leonardo Vaz de
• Data de Publicação: 2009
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: LOCUS Repositório Institucional da UFV
• Texto Completo: http://locus.ufv.br/handle/123456789/5473
• topic: Geoprocessamento; Bacias hidrográficas; Classificação de solos; GIS; Watersheds; Soil classification; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA::CIENCIA DO SOLO

Human occupations have a history of lack of planning and lack of organization that, many times, have compromised people s quality of life and environmental quality. It is observed that soil loss is a major environmental problem associated with the use and occupation of new areas, which, besides affecting the native flora and fauna, can seriously compromise the flow of watercourses and, therefore, people's quality of life. In this context, in any planning process, the availability and the information about the soil resources, in particular of its limitations and capabilities, are important data. Given the above, this study has aimed to: i) evaluate the effectiveness of two different approaches to soil mapping: automated via Artificial Neural Networks (ANN) and the traditional one, made with the help of soil scientists; and ii) estimate the carbon stock of the soil in Bacia do Rio Turvo Sujo, located in the Zona da Mata of Minas Gerais. The methodology consisted of modeling the distribution of soils in the basin area, using existing and new data, Geographic Information System (GIS), RNA's and field checks. The existing data were obtained in theses and documents that contained information on soil profiles in the area. The new data were collected in the field, in modal profiles for specific points of the basin, due to the absence of previous data. In the collections in the field, the profile was described and samples were collected from each horizon. We also raised some points of observation of some classes of soil. In working with GIS, a digital elevation model (DEM) was generated and from it, there were slope mapping, radiation, surface of sun exposure, horizontal distance, vertical distance, and curvature. Combining the curvature, slope, and vertical distance to drainage resulted in a map of units in the landscape of that region: top, terrace, convex and concave hillsides. In addition to these input data, three bands of an ASTER image (from 2001) and the Normalized Difference Vegetation Index (NDVI) were used. Two soil maps were then produced: one made with soil scientists, focused solely on field and laboratory data, and the other, which used GIS and ANN tools. The evaluation of the suitability of soil maps produced was made from the Kappa index. In the conventional map five classes of soil were established, and in the one produced via ANN, seven classes. The results showed greater accuracy of the map produced via ANN (Kappa = 66.6%) compared with that generated in traditional way (Kappa = 42.6%). The carbon stock in the soil, estimated up to one meter depth based on profile data from the basin of Rio Turvo, was around 4.900.000 tons, considering the depth of 1 meter. The relationship between the occurrence of soils and relief attributes, derived from DEM, were decisive in the delineation of soil mapping units in the region studied, which certainly ensured a better performance of the digital map production. Although extrapolation to other regions should be taken with caution, the results indicate that the ANNs were able to map the soils in the Bacia do Rio Turvo Sujo in an automated and appropriate manner, with more detailed results than the traditional procedure. In the conventional classification, it is difficult to perfectly integrate a wide range of information from different sources and in this respect, the ANNs result in helping to obtain a final product of better quality. Obviously, this data processing will not be able to replace human hands and minds in the survey procedures and soil classification, but should be taken as a tool to obtain results increasingly closer to reality.