Aplicação e avaliação de técnicas neuro-fuzzy para a elaboração de mapas de susceptibilidade a erosão

• Ano de defesa: 2020
• Autor(a) principal: Serikawa, Vagner de Souza
• Orientador(a): Melanda, Edson Augusto
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Urbana - PPGEU
• Palavras-chave em Português: Erosão; Geoinformação; FIS; ANFIS; SIG
• Palavras-chave em Inglês: Erosion; Geoinformation; GIS
• Área do conhecimento CNPq: ENGENHARIAS::ENGENHARIA CIVIL::GEOTECNICA
• Link de acesso: https://repositorio.ufscar.br/handle/ufscar/13076

The objective of this thesis was to analyze and contrast the existing methodological models within the geographic information system, its techniques and indicators for environmental urban planning, aiming at building bases for smart cities To achieve the objectives, the following objectives were proposed: (a) elaborate an inventory of erosive processes in the study area; (b) generate a collection of cartographic documents; (c) build training bases for different models of artificial neural networks (one-layer perceptron neural network, multilayer perceptron neural network and ANFIS); and discuss the methods and methodologies for environmental urban planning The study area is the Monjolinho River Basin inserted in the municipalities of São Carlos and Ibaté, with an area of approximately 273,77km2. The method used included: the elaboration of an input data structure for Matlab®️, extracted from a data matrix of the maps elaborated in the ArcGIS®️ 10.5 desktop version software. The training was carried out in two phases, the first with a structure of 203,496 points, extracted from a regular matrix of 100x100m, in the second phase a new training was done, with some modifications in the structure, two models were trained, design 1 with crisps data and design 2 with data normalized between 0 and 1, using the perceptron method with a layer, with a 30x30m matrix, with 13,152, 355 representing erosions. In the process, the removal of tuples with invalid data (for example, values -9999), counting the number of occurrences of repeated tuples and ending the removal of repeated tuples, all these steps were done to equalize the imbalance of the classes, totaling 710 “non-erosion” and 355 “erosion” maintaining a 2 to 1 ratio With the new training, a model was sought to more accurately represent the occurrence of erosive processes, so it is necessary that the values that represent false negatives are low, as it is critical for the system to indicate that there is no erosion when it actually exists. In this sense, the perceptron network of design 2 was the one that presented the best results, presenting less occurrence of false negative values than design 1.