Modelagem de risco de incêndios florestais utilizando redes neurais artificiais aplicada às regiões metropolitanas

• Ano de defesa: 2019
• Autor(a) principal: Luiza Cintra Fernandes
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; IGC - DEPARTAMENTO DE CARTOGRAFIA; Programa de Pós-Graduação em Análise e Modelagem de Sistemas Ambientais; UFMG
• Palavras-chave em Português: Risco de incêndio; Redes neurais artificiais; Modelagem; Modelagem de dados – Aspectos ambientais; Incêndios florestais – Previsão – Minas Gerais; Incêndios florestais – Prevenção e controle; Redes neurais (Computação)
• Link de acesso: http://hdl.handle.net/1843/33835

Forest Fires represents one of the greats concerns related with the environment, as it causes impacts in the vegetation, soil, fauna and flora. The fire risk index consists in verify the probability of fire occurrence in such place, so it is important to determine higher fire risk areas and planning activities for your prevention. From the existing methodologies for your calculation, the Artificial Neural Networks (ANN) is a methodology that have been achieving good results. The ANN it is a set of simple process units working in parallel, which stores experimental knowledge and make it available to use. This knowledge is acquired from a training process and stored in synaptic weights (connections between neurons). So, this work has the aim of mapping the fire risk areas in the Metropolitan Region of Belo Horizonte (RMBH) using an artificial neural network model with supervised training. This region presents accentuated dry seasons, due atmospheric systems activities; this season became the region propitious for fire ignition. 12 input variables were used: distance to urban areas, distance to roads, slope, aspect, land use, NDVI, minimum relative humidity, maximum air temperature, air pressure, solar radiation, wind speed and total precipitation. For network training it was used monthly data from 2014 to 2016, as output, fire active data from MODIS and VIIRS were used. A model evaluation was done with 2017 data. The RNA model results showed high accuracy and good consistency with fire data, but your validation underestimated risk areas in dry season. So complementary models were developed, training dry and rainy seasons separately. These models showed good results. In the dry model the high risk area covers almost all the region, in the other hand, the rainy model showed high risk only in locations near urban areas. The general model even dough it did not showed the best result, it was trained with characteristics of both seasons, selecting just the factors that really affects the fire, and can be used as a complementary model for the others. Among the input variables those which have more influence in the model are the distance to urban areas and meteorological variables. So, the ANN showed a good methodology for this end and can be used as a forecast model and be reapplied in other regions.