@MastersThesis{Toniol:2017:UsDiCl,
author = "Toniol, Alana Carla",
title = "Uso de diferentes classificadores e de simula{\c{c}}{\~a}o
estoc{\'a}stica para discrimina{\c{c}}{\~a}o de fitofisionomias
do Cerrado usando atributos hiperespectrais do sensor
Hyperion/EO-1",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2017",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2017-04-11",
keywords = "Cerrado, classifica{\c{c}}{\~a}o, hiperespectral,
simula{\c{c}}{\~a}o estoc{\'a}stica. Cerrado, classification,
hyperspectral, stochastic simulation.",
abstract = "O Cerrado brasileiro {\'e} considerado um dos mais importantes
ecossistemas do mundo tanto pela riqueza de fauna quanto por sua
ampla diversidade de esp{\'e}cies herb{\'a}ceas, arbustivas e
arb{\'o}reas que ocorrem em um gradiente de vegeta{\c{c}}{\~a}o
bem definido. Tendo em vista a import{\^a}ncia do monitoramento
desse hotspot de biodiversidade, o sensoriamento remoto
hiperespectral pode fornecer informa{\c{c}}{\~o}es sobre as
caracter{\'{\i}}sticas biof{\'{\i}}sicas e
bioqu{\'{\i}}micas de sua cobertura vegetal. O objetivo deste
trabalho foi identificar o melhor conjunto de atributos
hiperespectrais do sensor Hyperion/Earth Observing One (EO-1),
testando o desempenho de diferentes t{\'e}cnicas de
classifica{\c{c}}{\~a}o supervisionadas com esses atributos para
discrimina{\c{c}}{\~a}o de fitofisionomias do Cerrado. Na etapa
de classifica{\c{c}}{\~a}o foram consideradas duas imagens
referentes {\`a} esta{\c{c}}{\~a}o chuvosa (13/01/2015) e seca
(24/06/2015). A {\'a}rea de estudo foi o Parque Nacional de
Bras{\'{\i}}lia (PNB). Os atributos testados foram: (a) a
reflect{\^a}ncia de 146 bandas do sensor Hyperion; (b) a primeira
derivada da reflect{\^a}ncia; (c) 22 {\'{\i}}ndices de
vegeta{\c{c}}{\~a}o (IVs) de bandas estreitas; (d) a
profundidade, {\'a}rea, largura e assimetria das bandas de
absor{\c{c}}{\~a}o de clorofila em 680 nm; {\'a}gua foliar em
980 e 1200 nm; lignina e celulose em 1700, 2100 e 2300 nm; e (e)
todos os atributos em conjunto. Os classificadores testados foram
{\'A}rvore de Decis{\~a}o J48 (AD), Random Forest (RF), Spectral
Angle Mapper (SAM) e Support Vector Machine (SVM). Os resultados
mostraram que a maior quantidade de atributos selecionados no
per{\'{\i}}odo chuvoso compensou as confus{\~o}es espectrais
associadas {\`a} estrutura da vegeta{\c{c}}{\~a}o durante esse
per{\'{\i}}odo. Bandas mais profundas de absor{\c{c}}{\~a}o de
{\'a}gua foram observadas no per{\'{\i}}odo chuvoso para as
forma{\c{c}}{\~o}es arb{\'o}reas que apresentaram tamb{\'e}m
maiores taxas de varia{\c{c}}{\~a}o espectral associadas {\`a}
borda vermelha (primeira derivada). As classifica{\c{c}}{\~o}es
do per{\'{\i}}odo chuvoso apresentaram desempenho levemente
superior {\`a}s classifica{\c{c}}{\~o}es do per{\'{\i}}odo
seco, especialmente para tipologias que inclu{\'{\i}}am
esp{\'e}cies invasoras, embora a maioria das diferen{\c{c}}as em
exatid{\~a}o de classifica{\c{c}}{\~a}o n{\~a}o tenham sido
estatisticamente diferentes. As maiores exatid{\~o}es totais
foram atribu{\'{\i}}das {\`a}s classifica{\c{c}}{\~o}es com
todos os atributos em conjunto, enquanto que as menores
exatid{\~o}es foram relacionadas aos atributos par{\^a}metros de
bandas de absor{\c{c}}{\~a}o e derivada de 1\$^{ª}\$ ordem.
Pelos mapas de entropia de Shannon e de moda, observou-se que as
maiores incertezas entre os classificadores ocorreram para os
atributos derivada de 1\$^{ª}\$ ordem e par{\^a}metros de
bandas de absor{\c{c}}{\~a}o, estando associadas com as
fitofisionomias campestres. Pelo processo de simula{\c{c}}{\~a}o
estoc{\'a}stica foram confirmados os resultados obtidos pelos
mapas de classifica{\c{c}}{\~a}o. Considerando um intervalo de
credibilidade de 99\%, pode-se concluir que os melhores
resultados de classifica{\c{c}}{\~a}o nos per{\'{\i}}odos
chuvoso e seco foram observados para RF e SVM. Usando estes
classificadores, as maiores percentagens de acerto de
classifica{\c{c}}{\~a}o foram observadas com todos os atributos
em conjunto para as forma{\c{c}}{\~o}es campestres e com IVs,
reflect{\^a}ncia e todos os atributos para as
forma{\c{c}}{\~o}es arb{\'o}reas. A utiliza{\c{c}}{\~a}o de
simula{\c{c}}{\~a}o estoc{\'a}stica foi importante para a
complementa{\c{c}}{\~a}o e confirma{\c{c}}{\~a}o dos
resultados estat{\'{\i}}sticos associados aos processos de
classifica{\c{c}}{\~a}o de imagens Hyperion. ABSTRACT: The
Brazilian savanna, locally known as Cerrado, is one of the most
important ecosystems of the world because of the high biodiversity
of trees, shrubs and grasses associated with a well-defined
vegetation gradient. In order to monitor this important world's
hotspot, hyperspectral remote sensing can provide information on
biophysical and biochemistry vegetation properties. The objective
of this study was to identify the best set of hyperspectral
attributes to be used as input to different classification
techniques for discriminating the Cerrado physiognomies. In the
classification phase, two images were considered in rainy season
(01/13/2015) and dry (06/24/2015).The study area is the Parque
Nacional de Bras{\'{\i}}lia (PNB). The attributes tested were,
as follows: (a) the reflectance of 146 Hyperion bands; (b) the
first-order derivative of reflectance; (c) 22 narrowband
vegetation indices (VIs); (d) the depth, area, width and asymmetry
of the 680-nm chlorophyll absorption band; the 980-nm and 1200-nm
leaf water features; the 1700-nm, 2100-nm and 2300-nm
lignin/cellulose absorption bands; and (e) all sets of attributes.
The classifiers used in the data analysis were Decision Tree J48
(DT), Random Forest (RF), Spectral Angle Mapper (SAM) and Support
Vector Machine (SVM). The results showed that the greater spectral
confusion in the rainy season than in the dry season was
compensated by the selection of a greater number of hyperspectral
attributes in the classification procedure. Deeper leaf water
absorption bands were observed in the rainy season for the
tree-wooded savannas, which showed also greater rates of
reflectance changes in the red-edge interval (first-order
derivative). Classification accuracy in the rainy season was
slightly higher than in the dry season, especially for classes
with invasive species, but most of the differences were not
statistically significant. The highest classification accuracy was
obtained with the use of all hyperspectral attributes, while the
lowest values were noted for the absorption band parameters and
first-order derivative of reflectance. These results were
confirmed by the Shannon entropy and mode maps, which showed that
the greatest uncertainties in the classification were associated
with the grassland/shrub savanna physiognomies. From the
stochastic simulation at 99\% confidence level, it was concluded
that the best classification results in both seasons were observed
for RF and SVM. Using these classifiers, the largest percentages
of correct classification were obtained with all attributes for
the grassland/shrub savannas and with reflectance, VIs and all
attributes for the tree-wooded physiognomies. Overall, the
stochastic simulation was important for complementing and
confirming the statistical results associated with the
classification of the Hyperion images.",
committee = "Ponzoni, Fl{\'a}vio Jorge (presidente) and Galv{\~a}o,
L{\^e}nio Soares (orientador) and Sanches, Ieda Del'Arco and
Sano, Edson Eyji",
englishtitle = "Use of different classifiers and stochastic simulation for the
discrimination of Cerrado physiognomies using hyperspectral
attributes of the Hyperion/EO-1 sensor.",
language = "pt",
pages = "144",
ibi = "8JMKD3MGP3W34P/3NNR6DB",
url = "http://urlib.net/ibi/8JMKD3MGP3W34P/3NNR6DB",
targetfile = "publicacao.pdf",
urlaccessdate = "27 abr. 2024"
}