Fechar
Metadados

@MastersThesis{Ibaņez:2016:UsReNe,
               author = "Ibaņez, Marilyn Menecucci",
                title = "Uso de redes neurais nebulosas e florestas aleat{\'o}rias na 
                         classifica{\c{c}}{\~a}o de imagens em um projeto de ci{\^e}ncia 
                         cidad{\~a}",
               school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
                 year = "2016",
              address = "S{\~a}o Jos{\'e} dos Campos",
                month = "2016-02-19",
             keywords = "redes neurais, florestas aleat{\'o}rias, processamento de 
                         imagens, computa{\c{c}}{\~a}o cidad{\~a}, desmatamento, 
                         sat{\'e}lites, neural network, image processing, computing 
                         citizen, desforestation, satellites.",
             abstract = "Recentemente, um projeto de ci{\^e}ncias cidad{\~a} chamado 
                         \emph{ForestWatchers} (LUZ et al., 2014) foi lan{\c{c}}ado com o 
                         objetivo de envolver os cidad{\~a}os leigos no monitoramento do 
                         desmatamento. Por meio de uma interface Web, volunt{\'a}rios de 
                         todo o mundo s{\~a}o convidados a analisar imagens MODIS de 
                         regi{\~o}es florestais e confirmar se atribui{\c{c}}{\~o}es 
                         autom{\'a}ticas de regi{\~o}es de florestas desmatadas 
                         est{\~a}o corretamente classificadas. Considerando a grande 
                         {\'a}rea em todo mundo coberta pelas florestas tropicais, 
                         torna-se fundamental o uso de um classificador r{\'a}pido que 
                         atenda a um objetivo duplo: o mapeamento de pixels em duas classes 
                         (\${'}\$Floresta\${'}\$ e \${'}\$n{\~a}o-Floresta\${'}\$) 
                         e a sele{\c{c}}{\~a}o dos pixels a serem enviados aos 
                         volunt{\'a}rios para a inspe{\c{c}}{\~a}o, com base em uma 
                         m{\'e}trica de confian{\c{c}}a. Nesta disserta{\c{c}}{\~a}o 
                         investiga-se o uso de dois m{\'e}todos distintos - rede neural de 
                         perceptrons multicamada (\emph{Multi-Layered Perceptron}, MLP) 
                         difusa e Floresta Aleat{\'o}ria (\emph{Random Forest, RF}) - na 
                         classifica{\c{c}}{\~a}o de padr{\~o}es de desmatamento na 
                         Amaz{\^o}nia brasileira, utilizando imagens MODIS. Neste sentido, 
                         foram gerados mapas de desmatamento de diversos tamanhos, de 
                         diversas {\'a}reas do estado de Rond{\^o}nia. Os resultados 
                         foram validados com os resultados de projeto PRODES, que avalia 
                         anualmente o desmatamento na Amaz{\^o}nia brasileira. Nestes 
                         testes, o classificador RF apresentou um desempenho amplamente 
                         superior ao das redes neurais \emph{Multi-Layered Perceptro e 
                         Multi-Layered Perceptron Fuzzy}. ABSTRACT: Recently, a citizen 
                         science project called ForestWatchers (LUZ et al., 2014) was 
                         launched in order to involve the laity citizens in the monitoring 
                         of deforestation. Through a Web interface, volunteers from around 
                         the world are invited to review MODIS images of forest regions and 
                         confirm that automatic assignment of cleared forest areas are 
                         properly classified. Considering the large area worldwide covered 
                         by tropical forest, it is essential to use a fast classifier that 
                         meets a double objective: the pixel mapping into two classes 
                         (\${'}\$Forest\${'}\$ and \${'}\$non-forest\${'}\$) and 
                         the selection of pixels to be sent to volunteers for inspection, 
                         based on a reliable metric. This dissertation investigates the use 
                         of two different methods - neural network multilayer perceptrons 
                         (Multi-Layered Perceptron, MLP) diffuse and Random Forest (Random 
                         Forest, RF) - the deforestation pattern classification in the 
                         Brazilian Amazon using MODIS images. In this sense, deforestation 
                         maps were generated from various sizes, from different areas of 
                         the state of Rondonia. The results were validated with the results 
                         of PRODES project, which annually evaluates deforestation in the 
                         Brazilian Amazon. In these tests, the classifier RF showed a 
                         vastly superior performance to the \emph{Multi-Layered Perceptro 
                         and Multi-Layered Perceptron Fuzzy neural networks.}.",
            committee = "Rosa, Reinaldo Roberto (presidente) and Ramos, Fernando Manuel 
                         (orientador) and Carvalho, Adenilson Roberto (orientador) and 
                         Becceneri, Jos{\'e} Carlos and Shiguemori, Elcio Hideiti",
           copyholder = "SID/SCD",
         englishtitle = "Use of fuzzy neural networks and random forest in image's 
                         classification of a citizen science project",
             language = "pt",
                pages = "120",
                  ibi = "8JMKD3MGP3W34P/3L2D5ME",
                  url = "http://urlib.net/rep/8JMKD3MGP3W34P/3L2D5ME",
           targetfile = "publicacao.pdf",
        urlaccessdate = "25 nov. 2020"
}


Fechar