@PhDThesis{GirolamoNeto:2018:IdFiCe,
author = "Girolamo Neto, Cesare Di",
title = "Identifica{\c{c}}{\~a}o de fitofisionomias de Cerrado no Parque
Nacional de Bras{\'{\i}}lia utilizando random forest aplicado a
imagens de alta e m{\'e}dia resolu{\c{c}}{\~o}es espaciais",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2018",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2018-08-28",
keywords = "sensoriamento remoto, classifica{\c{c}}{\~a}o, cerrado, textura,
minera{\c{c}}{\~a}o de dados, remote sensing, classification,
cerrado, texture, data mining.",
abstract = "Depois da Mata Atl{\^a}ntica, o Cerrado {\'e} o bioma Brasileiro
que mais passou por altera{\c{c}}{\~o}es com a
ocupa{\c{c}}{\~a}o humana, com uma perda de cobertura vegetal de
978.745 kmē. Portanto, {\'e} estrat{\'e}gico que o bioma seja
monitorado para combater o desmatamento e manter as {\'a}reas de
preserva{\c{c}}{\~a}o ambiental. Neste sentido, v{\'a}rias
pesquisas t{\^e}m sido desenvolvidas para estudar as
altera{\c{c}}{\~o}es da cobertura e uso da terra, estimar as
emiss{\~o}es de carbono, estudar o impacto do desmatamento e
degrada{\c{c}}{\~a}o da biodiversidade. A maioria dos trabalhos
para classificar a cobertura vegetal do Cerrado tem utilizado
imagens da classe Landsat, com 30 metros de resolu{\c{c}}{\~a}o
espacial, discriminando fitofisionomias de Campo, Savana e
Floresta com taxas de acerto superiores a 80%. Todavia, ainda
s{\~a}o encontradas dificuldades em classificar as
fitofisionomias com uma legenda de classifica{\c{c}}{\~a}o mais
detalhada. Esse fato mostra a necessidade do uso de imagens de
resolu{\c{c}}{\~a}o espacial melhor, as quais se mostraram
capazes de identificar a estrutura da vegeta{\c{c}}{\~a}o em
fitofisionomias semelhantes. Embora, o uso destas imagens permita
classificar a cobertura vegetal do Cerrado com mais detalhes,
n{\~a}o h{\'a} na literatura pesquisas conclusivas sobre quais
fitofisionomias podem ser mais bem discriminadas com imagens de
resolu{\c{c}}{\~a}o espacial da classe Landsat (30m) e alta
resolu{\c{c}}{\~a}o (1 a 4 m). Dentro deste contexto, o
principal objetivo deste trabalho foi avaliar imagens de alta e
m{\'e}dia resolu{\c{c}}{\~o}es, combinadas com t{\'e}cnicas de
extra{\c{c}}{\~a}o de atributos, para mapear as fitofisionomias
do Cerrado com um maior n{\'{\i}}vel de detalhamento do que a
literatura existente. Foram obtidas imagens (Landsat-8 e
WorldView-2) para o Parque Nacional de Bras{\'{\i}}lia,
regi{\~a}o que cont{\'e}m mais de 30 mil hectares de
vegeta{\c{c}}{\~a}o nativa de Cerrado. A partir destas imagens
foram gerados dados de Reflect{\^a}ncia, do Modelo Linear de
Mistura Espectral, da Transformada Tasseled Cap, de
{\'{\I}}ndices de Vegeta{\c{c}}{\~a}o e de textura. Foram
coletados pontos em campo e com interpreta{\c{c}}{\~a}o visual
para gerar um conjunto de dados com mais de mil amostras
classificadas para quatro diferentes legendas. Estas legendas
consideram as fitofisionomias de Ribeiro e Walter (2008) e foram
propostas, contendo 3, 6, 8 e 10 diferentes classes, de acordo com
seu n{\'{\i}}vel de detalhamento. A escala mais simples
diferenciou apenas as classes Campestre, Sav{\^a}nica e
Florestal, e conforme o aumento da complexidade, foram
identificadas classes como Campo Limpo, Campo Limpo {\'U}mido,
Campo Limpo {\'U}mido com Murundu, Campo Sujo, Campo Rupestre,
Cerrado Ralo, Cerrado T{\'{\i}}pico, Cerrado Denso, Veredas e
Mata de Galeria. O comportamento espectral destas fitofisionomias
revelou que elas s{\~a}o diferenci{\'a}veis apenas para a escala
mais simples. Para n{\'{\i}}veis mais complexos, existe uma
maior dificuldade de discrimina{\c{c}}{\~a}o com dados de
Reflect{\^a}ncia. A classifica{\c{c}}{\~a}o das imagens foi
feita pelo algoritmo Random Forest. Dentre os principais
resultados, a legenda mais simples de mapeamento mostra-se
adequada para ambas {\`a}s resolu{\c{c}}{\~o}es espaciais,
obtendo taxas de acerto superiores a 87%. Com o aumento da
complexidade de legenda, a imagem Landsat-8 passou a apresentar
limita{\c{c}}{\~o}es na discrimina{\c{c}}{\~a}o de classes
como Campo Limpo, Campo Sujo e Campo Rupestre. As classes de
Cerrado Ralo e Cerrado Denso apresentaram confus{\~a}o com a
classe de Cerrado T{\'{\i}}pico. Ainda foi constatado que estas
imagens s{\~a}o deficientes em representar a
transi{\c{c}}{\~a}o entre as classes de Campo Sujo e Cerrado
Ralo. A taxa de acerto para a legenda mais detalhada com a imagem
Landsat-8 foi de 65,21%. Entretanto, a imagem WorldView-2 se
mostrou capaz de identificar estas fitofisionomias com uma melhor
taxa de acerto (74,17%). O uso de atributos relacionados {\`a}
textura foi essencial para o aumento dessa taxa. Neste sentido,
por meio da imagem WorldView-2 foi poss{\'{\i}}vel identificar a
Classe de Campo Rupestre com melhor taxa de acerto, reduzindo
erros entre as classes de Campo Limpo e Campo Sujo. xii As classes
de Cerrado Ralo e Cerrado Denso reduziram sua confus{\~a}o com
Cerrado T{\'{\i}}pico. O erro de transi{\c{c}}{\~a}o entre
Campo Sujo e Cerrado Ralo ainda persiste na imagem de alta
resolu{\c{c}}{\~a}o, por{\'e}m com uma menor magnitude. Algumas
classes como Veredas, Campo Limpo {\'U}mido e Campo Limpo
{\'U}mido com Murundu n{\~a}o foram identificadas com boa
precis{\~a}o em ambas as imagens. Dentre as principais
conclus{\~o}es deste trabalho destacam-se o uso de atributos de
textura para melhorar a discrimina{\c{c}}{\~a}o de
fitofisionomias do Cerrado. Estes atributos foram capazes de
representar as varia{\c{c}}{\~o}es entre regi{\~o}es com
vegeta{\c{c}}{\~a}o arb{\'o}rea intercaladas por regi{\~o}es
com vegeta{\c{c}}{\~a}o herb{\'a}ceoarbustiva, melhorando a
discrimina{\c{c}}{\~a}o de fitofisionomias como Campo Sujo,
Cerrado Ralo, Cerrado T{\'{\i}}pico e Cerrado Denso. ABSTRACT:
After the Atlantic Forest, the Cerrado is the Brazilian biome that
has presented most changes with human occupation, with a loss of
vegetation cover of 978,745 kmē. Therefore, it is strategic that
this biome is monitored in order to decrease deforestation and
maintain the areas of environmental protection. In this sense,
several researches have been developed to study changes in land
cover and use, to estimate carbon emissions, to study the impact
of deforestation and biodiversity degradation. Most of these
studies classify Cerrado vegetation using Landsat like images,
with 30 meters of spatial resolution, discriminating classes such
as Grassland, Savanna and Woodland with accuracy higher than 80%.
However, it is still difficult to classify Cerrado
phytophysiognomies with a more detailed classification legend.
This fact shows the need of better spatial resolution images,
which were able to identify vegetation structure in similar
phytophysiognomies. Although the use of these images allows
classifying the Cerrado vegetation cover with more details, there
is no conclusive research in the literature about which
phytophysiognomie can be discriminated with better accuracy with
Landsat images (30m) and high resolution (1 to 4 m). In this
context, the aim of this work was to evaluate high and medium
resolution images, combined with feature extraction techniques, to
map Cerrado phytophysiognomies with a higher level of detail than
the existing literature. A Landsat-8 image and a WorldView-2 image
were obtained for the Brasilia National Park, a region that
contains more than 30 thousand hectares of Cerrado native
vegetation. Data of Reflectance, Spectral Linear Mixture Model,
Tasseled Cap Transformation, Vegetation Indices and texture were
obtained for these images. Samples for the classification were
collected on field and by visual interpretation, generating a
dataset with more than one thousand samples classified for four
different legends. These legends adopts the phytophysiognomies
described by Ribeiro and Walter (2008) and were proposed
containing 3, 6, 8 and 10 different classes, according to their
level of detail. The simpler scale adopted the classes of
Grassland, Savanna and Woodland. When the level of detail was
increased, the following classes were used: Open Grassland,
Flooded Grassland, Flooded Grassland with Murundu, Shrub
Grassland, Rocky Grassland, Shrub Savanna, Typical Savanna, Dense
Savanna, Flooded Plains with Palmtrees and Gallery Forest. The
spectral behavior of these phytophysiognomies revealed that they
are distinguishable only for the simplest scale, for more complex
levels, there is a greater difficulty of discrimination with
Reflectance data. The classification of the images was done by the
algorithm Random Forest. Among the main results, the simplest
mapping legend is adequate for both spatial resolutions, obtaining
hit rates higher than 87%. With the increase of the legend
complexity, the Landsat-8 images started to present difficulties
in discriminating classes like Open Grassland, Shrub Grassland and
Rocky Grassland. The classes of Shrub Savanna and Dense Savanna
were misclassified as Typical Savanna. It was still observed that
these images are deficient in representing the transition between
the classes of Shrub Grassland and Shrub Savanna. The hit rate for
the most detailed legend with the Landsat-8 image was 65.21%.
However, the WorldView-2 image was able to identify these
phytophysiognomies with a better accuracy (74.17%). The use of
texture features was essential for this increase. In this sense,
the WorldView-2 image xiv identified the Rock Grassland class with
better accuracy and also reduced the misclassification between the
Open Grasslands and Shrub Grasslands. The classes of Shrub Savanna
and Dense Savanna reduced their confusion with Typical Savanna.
The transition error between Shrub Grasslands and Shrub Savanna
still persists in WorldView-2 image, but with a smaller magnitude.
Some classes such as Flooded Plains with Palmtrees, Flooded
Grassland and Flooded Grassland with Murundu were not identified
with good accuracy on both images. The main conclusion of this
study is that the use of texture features helped to improve the
discrimination of Cerrado phytophysiognomies. These features were
able to represent the variations between regions with arboreal
vegetation interspersed by regions with herbaceous-shrub
vegetation, improving the discrimination of phytophysiognomies
such as Shrub Grasslands, Shrub Savanna, Typical Savanna and Dense
Savanna.",
committee = "Ponzoni, Fl{\'a}vio Jorge (presidente) and Fonseca, Leila Maria
Garcia (orientadora) and K{\"o}rting, Thales Sehn (orientador)
and Valeriano, Dalton de Morisson and Negri, Rog{\'e}rio Galante
and Lacerda, Camila Souza dos Anjos",
englishtitle = "Identification of Brazilian savannah physiognomies on
Bras{\'{\i}}lia National Park using random forest on high and
medium spatial resolution images",
language = "pt",
pages = "186",
ibi = "8JMKD3MGP3W34R/3RU6Q68",
url = "http://urlib.net/ibi/8JMKD3MGP3W34R/3RU6Q68",
targetfile = "publicacao.pdf",
urlaccessdate = "27 abr. 2024"
}