Fechar
Metadados

@MastersThesis{Neves:2017:MiDaSe,
               author = "Neves, Alana Kasahara",
                title = "Minera{\c{c}}{\~a}o de dados de sensoriamento remoto para 
                         detec{\c{c}}{\~a}o e classifica{\c{c}}{\~a}o de {\'a}reas de 
                         pastagem na Amaz{\^o}nia Legal",
               school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
                 year = "2017",
              address = "S{\~a}o Jos{\'e} dos Campos",
                month = "2017-02-15",
             keywords = "pasto limpo, pasto sujo, {\'a}rvore de decis{\~a}o, s{\'e}ries 
                         temporais, TerraClass, herbaceous pasture, shrubby pasture, 
                         decision tree, time series.",
             abstract = "Aproximadamente 60\% das {\'a}reas desflorestadas na 
                         Amaz{\^o}nia Legal s{\~a}o ocupadas por pastagens. A 
                         expans{\~a}o das {\'a}reas de pastagem sobre {\'a}reas de 
                         floresta pode ser associada a fatores como o mercado de terras e a 
                         perda de produtividade da pastagem ao longo do tempo. De toda 
                         forma, essa expans{\~a}o permanece como um obst{\'a}culo ao 
                         combate ao desflorestamento. A detec{\c{c}}{\~a}o e 
                         avalia{\c{c}}{\~a}o das condi{\c{c}}{\~o}es das pastagens 
                         permitem melhores monitoramento e controle, assim como a 
                         identifica{\c{c}}{\~a}o de {\'a}reas prop{\'{\i}}cias para a 
                         recupera{\c{c}}{\~a}o. Dentro deste contexto, o objetivo deste 
                         trabalho {\'e} desenvolver uma metodologia de reconhecimento de 
                         {\'a}reas de pastagem na Amaz{\^o}nia, com base na 
                         detec{\c{c}}{\~a}o e classifica{\c{c}}{\~a}o por meio de 
                         atributos de s{\'e}ries temporais e t{\'e}cnicas de 
                         minera{\c{c}}{\~a}o de dados, de acordo com as 
                         condi{\c{c}}{\~o}es da cobertura vegetal. A {\'a}rea de estudo 
                         consiste em tr{\^e}s {\'o}rbitas/ponto do sat{\'e}lite Landsat 
                         8 distribu{\'{\i}}das em quatro estados (AC, MT, RO e AM): 
                         001/067, 226/068 e 231/067. Foram utilizadas imagens de 
                         reflect{\^a}ncia de superf{\'{\i}}cie do sensor OLI do Landsat 
                         8, referentes ao per{\'{\i}}odo entre abril de 2013 e dezembro 
                         de 2015. As nuvens e sombras de nuvens foram detectadas pelo 
                         algoritmo FMask e exclu{\'{\i}}das. A classifica{\c{c}}{\~a}o 
                         foi feita em duas etapas: detec{\c{c}}{\~a}o de pastagem 
                         (diferenci{\'a}-las de outros alvos da cena) e, posteriormente, a 
                         classifica{\c{c}}{\~a}o das pastagens em Pasto Limpo e Pasto 
                         Sujo. Para a caracteriza{\c{c}}{\~a}o destas classes, os 
                         seguintes atributos foram usados: {\'{\i}}ndices de 
                         vegeta{\c{c}}{\~a}o (NDVI, EVI, EVI2, SAVI e NDII), 
                         fra{\c{c}}{\~o}es do Modelo Linear de Mistura Espectral 
                         (fra{\c{c}}{\~o}es vegeta{\c{c}}{\~a}o, NPV e solo), 
                         componentes da transforma{\c{c}}{\~a}o \emph{Tasselled Cap 
                         (greenness, brightness e wetness)}, outros atributos espectrais e 
                         atributos texturais. As duas etapas da classifica{\c{c}}{\~a}o 
                         foram realizadas utilizando tr{\^e}s classificadores ({\'a}rvore 
                         de decis{\~a}o, \emph{random forest} e rede neural) e duas 
                         abordagens: por pixel e baseada em objetos. A 
                         avalia{\c{c}}{\~a}o dos resultados de classifica{\c{c}}{\~a}o 
                         baseou-se em trabalho de campo e interpreta{\c{c}}{\~a}o visual 
                         de imagens do sat{\'e}lite RapidEye. Os resultados mostraram 
                         melhoras nas taxas de acerto quando houve a utiliza{\c{c}}{\~a}o 
                         de segmentos, uma vez que as pastagens possuem uma grande 
                         quantidade de mistura de elementos em sua composi{\c{c}}{\~a}o. 
                         Os modelos criados e avaliados na mesma cena obtiveram altas taxas 
                         de acerto (pr{\'o}ximas a 90\%), entretanto n{\~a}o foram 
                         capazes de classificar outras cenas com a mesma efici{\^e}ncia. 
                         Quando amostras de duas cenas diferentes foram combinadas para a 
                         gera{\c{c}}{\~a}o do modelo, as taxas de acerto ficaram 
                         parecidas entre as imagens, por volta de 80\%. A maior 
                         dificuldade esteve na separa{\c{c}}{\~a}o entre Pasto Limpo e 
                         Pasto Sujo, uma vez que as pastagens na Amaz{\^o}nia variam de 
                         acordo com muitos fatores: manejo adotado, tipo de solo, regime de 
                         chuvas, tipo de gram{\'{\i}}nea utilizada e outros. ABSTRACT: 
                         The highest percentage of deforested areas in the Legal Amazon is 
                         occupied by pastures. The expansion of pasture areas over the 
                         forest may be associated with factors such as land speculation and 
                         loss of productivity over time. In any case, this expansion 
                         remains an obstacle to fight against deforestation. The detection 
                         and evaluation of pasture conditions allow better monitoring and 
                         control it, as well as the identification of suitable areas for 
                         recovery. In this context, the objective of this work is to 
                         develop a methodology for the recognition of pasture areas in the 
                         Amazon, based on the detection and classification by time series 
                         attributes and data mining techniques, according to the vegetation 
                         conditions. The study area consists of three path/rows of Landsat 
                         8 satellite distributed in four Brazilian states (AC, MT, RO and 
                         AM): 001/067, 226/068 and 231/067. Surface reflectance images from 
                         OLI sensor (Landsat 8 satellite) were used for the period between 
                         April 2013 and December 2015. Clouds and cloud shadows were 
                         detected by the FMask algorithm and excluded from the dataset. The 
                         classification was carried out in two steps: pasture detection 
                         (differentiate them from other targets in the scene) and, later, 
                         the classification of pastures between Herbaceous Pasture and 
                         Shrubby Pasture. For the characterization of these classes, the 
                         following attributes were used: vegetation indices (NDVI, EVI, 
                         EVI2, SAVI and NDII), fractions of the Linear Spectral Mixture 
                         Model (vegetation, NPV and soil), bands of the Tasselled Cap 
                         Transformation (Greenness, Brightness and Wetness), other spectral 
                         attributes and textured attributes. The two steps of 
                         classification were performed using three classifiers (decision 
                         tree, random forest and neural network) and two different 
                         approaches: per pixel and object-based. The evaluation of the 
                         classification results was based on fieldwork and visual 
                         interpretation of RapidEye satellite images. The results showed 
                         improvements in the accuracy when segments were used instead of 
                         pixels, since pastures have a large amount of mixture of elements 
                         in their composition. The models created and evaluated in the same 
                         scene obtained high accuracy (close to 90\%), but they were not 
                         able to classify other scenes with the same efficiency. When 
                         samples from two different scenes were combined for model 
                         generation, the accuracy was similar between the images, around 
                         80\%. The greatest difficulty was in the separation between 
                         Herbaceous Pasture and Shrubby Pasture, since pastures in Amazon 
                         may vary according to some factors, such as: adopted management, 
                         soil type, rainfall regime and type of grass used.",
            committee = "Escada, Maria Isabel Sobral (presidente) and K{\"o}rting, Thales 
                         Sehn (orientador) and Fonseca, Leila Maria Garcia (orientador) and 
                         Adami, Marcos and Esquerdo, J{\'u}lio C{\'e}sar Dalla Mora",
           copyholder = "SID/SCD",
         englishtitle = "Remote sensing data mining to detect and classify pasture lands in 
                         the Legal Amazon",
             language = "pt",
                pages = "101",
                  ibi = "8JMKD3MGP3W34P/3NAADAS",
                  url = "http://urlib.net/rep/8JMKD3MGP3W34P/3NAADAS",
           targetfile = "publicacao.pdf",
        urlaccessdate = "23 nov. 2020"
}


Fechar