@PhDThesis{Ferreira:2017:DeEsAr,
author = "Ferreira, Matheus Pinheiro",
title = "Detec{\c{c}}{\~a}o de esp{\'e}cies arb{\'o}reas em floresta
estacional semidecidual por sensoriamento remoto hiperespectral e
modelagem de transfer{\^e}ncia radiativa",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2017",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2017-05-19",
keywords = "sensoriamento remoto hiperespectral, Floresta Atl{\^a}ntica,
WorldView-3, DART. hyperspectral remote sensing, Atlantic Forest,
WorldView-3, DART.",
abstract = "O mapeamento da distribui{\c{c}}{\~a}o espacial de esp{\'e}cies
arb{\'o}reas em ambientes tropicais fornece
informa{\c{c}}{\~o}es valiosas para ec{\'o}logos e gestores
florestais. Esse procedimento pode reduzir custos com trabalhos de
campo, auxiliar o monitoramento da diversidade flor{\'{\i}}stica
do dossel e auxiliar a localiza{\c{c}}{\~a}o de {\'a}rvores
matrizes para a coleta de sementes em iniciativas de
restaura{\c{c}}{\~a}o florestal. Entretanto, a
detec{\c{c}}{\~a}o de esp{\'e}cies arb{\'o}reas em florestas
tropicais com dados de sensoriamento remoto {\'e} um desafio
devido {\`a} elevada diversidade espectral e flor{\'{\i}}stica.
O objetivo principal desta pesquisa de doutorado foi explorar a
utiliza{\c{c}}{\~a}o de dados reais e simulados de sensoriamento
remoto multi e hiperespectral para mapear e classificar
esp{\'e}cies arb{\'o}reas da Floresta Estacional Semidecidual.
Os objetivos espec{\'{\i}}ficos inclu{\'{\i}}ram: (i) avaliar
o desempenho de m{\'e}todos estat{\'{\i}}sticos de
classifica{\c{c}}{\~a}o na discrimina{\c{c}}{\~a}o de
esp{\'e}cies arb{\'o}reas da {\'a}rea de estudo; (ii)
identificar regi{\~o}es e bandas espectrais, no intervalo de 450
a 2400 nm, prop{\'{\i}}cias {\`a} classifica{\c{c}}{\~a}o das
esp{\'e}cies; (iii) avaliar a utiliza{\c{c}}{\~a}o de
{\'{\i}}ndices de vegeta{\c{c}}{\~a}o de bandas estreitas na
discrimina{\c{c}}{\~a}o das esp{\'e}cies; (iv) quantificar a
variabilidade espectral intra e interespec{\'{\i}}fica, bem como
sua influ{\^e}ncia, na classifica{\c{c}}{\~a}o das
esp{\'e}cies; (v) desenvolver um m{\'e}todo para o mapeamento
autom{\'a}tico das esp{\'e}cies ao n{\'{\i}}vel de
{\'a}rvores individuais; (vi) desenvolver uma abordagem de
simula{\c{c}}{\~a}o da resposta espectral das esp{\'e}cies ao
n{\'{\i}}vel de dossel a partir da utiliza{\c{c}}{\~a}o de
modelagem de transfer{\^e}ncia radiativa em tr{\^e}s
dimens{\~o}es (3D) e (vii) comparar a resposta espectral simulada
e medida das esp{\'e}cies. Tr{\^e}s m{\'e}todos de
classifica{\c{c}}{\~a}o supervisionada foram testados para
discriminar oito esp{\'e}cies arb{\'o}reas: An{\'a}lise
Discriminante Linear (LDA), Support Vector Machines com kernel
linear (L-SVM) e fun{\c{c}}{\~a}o de base radial (RBF-SVM) e
Random Forest (RF). Uma exatid{\~a}o de classifica{\c{c}}{\~a}o
m{\'e}dia de 70\% foi obtida ao se utilizar as bandas do
vis{\'{\i}}vel/infravermelho pr{\'o}ximo (VNIR, 450-919 nm). A
inclus{\~a}o de bandas do infravermelho de ondas curtas (SWIR,
1045-2400 nm) elevaram a exatid{\~a}o para 84\%.
{\'{\I}}ndices de vegeta{\c{c}}{\~a}o de bandas estreitas
tamb{\'e}m foram testados e elevaram a exatid{\~a}o em 5\%
quando combinados a bandas do VNIR. Dados reais e simulados do
sensor WorldView-3 (WV-3) foram utilizados para fins de
classifica{\c{c}}{\~a}o. Enquanto as bandas VNIR simuladas de
sensor forneceram uma exatid{\~a}o de 57,4\%, o conjunto
VNIR+SWIR aumentou a exatid{\~a}o para 74,8\%. Este padr{\~a}o
se manteve na classifica{\c{c}}{\~a}o de dados WV-3 reais
(aumento de 3,2 \% ap{\'o}s a inclus{\~a}o de bandas SWIR). O
grau de sobreposi{\c{c}}{\~a}o das variabilidades intra e
interespec{\'{\i}}fica influenciou diretamente as
classifica{\c{c}}{\~o}es. O m{\'e}todo para mapeamento de
{\'a}rvores desenvolvido produziu mapas fidedignos da
distribui{\c{c}}{\~a}o espacial das esp{\'e}cies e elevou a
exatid{\~a}o em rela{\c{c}}{\~a}o {\`a}s
classifica{\c{c}}{\~o}es ao n{\'{\i}}vel de pixel em at{\'e}
6\%. Um procedimento de sele{\c{c}}{\~a}o de atributos baseado
em regress{\~a}o stepwise identificou bandas localizadas ao redor
do pico do verde (550 nm), na fei{\c{c}}{\~a}o de
absor{\c{c}}{\~a}o do vermelho (650 nm) e no SWIR em 1200, 1700,
2100 e 2300 nm, como {\'u}teis para discriminar as esp{\'e}cies.
A resposta espectral das esp{\'e}cies no VNIR foi acuradamente
simulada pelo modelo Discrete Anisotropic Radiative Transfer
(DART) que opera em tr{\^e}s dimens{\~o}es (3D). Por meio de uma
estrutura simplificada da copa, a resposta espectral das
esp{\'e}cies no topo do dossel foi simulada com at{\'e} 1,5\%
de erro quadr{\'a}tico m{\'e}dio. A invers{\~a}o do modelo na
imagem hiperespectral gerou rela{\c{c}}{\~o}es
estat{\'{\i}}sticas significantes (R\$^{²}\$=0,65) entre o
conte{\'u}do de clorofila (C\$_{ab}\$) e o {\'{\i}}ndice
MCARI (Modified Chlorophyll Absorption in Reflectance Index),
sendo C\$_{ab}\$ estatisticamente diferente entre as
esp{\'e}cies. A abordagem de simula{\c{c}}{\~a}o desenvolvida
pode ser utilizada para reproduzir aquisi{\c{c}}{\~o}es
hiperespectrais de florestas tropicais. ABSTRACT: Accurately
mapping the spatial distribution of tree species in tropical
environments provides valuable insights for ecologists and forest
managers. This process may play an important role in reducing
fieldwork costs, monitoring changes in canopy biodiversity, and
locating parent trees to collect seeds for forest restoration
efforts. However, mapping tree species in tropical forests with
remote sensing data is a challenge because of high floristic and
spectral diversity. The main objective of this study was to
explore the use of experimental and simulated multi and
hyperspectral remotely sensed data for tree species discrimination
and mapping in a tropical seasonal semi-deciduous forest.
Specifically we aimed: (i) to evaluate the performance of machine
learning methods in the discrimination of tree species of the
study area; (ii) to identify spectral regions and bands, in
450-2400 nm range, suitable for species classification; (iii) to
evaluate the use of narrow band vegetation indices in species
discrimination; (iv) to quantify within- and among-species
spectral variability as well as its influence on species
classification; (v) develop a method for tree species mapping at
the individual tree level; (vi) to develop a modeling approach to
simulate the spectral response of the species at the canopy level
using three-dimensional (3D) radiative transfer modeling and (vii)
to compare simulated and measured spectral responses of the
species. Three classifiers were tested to discriminate eight tree
species: Linear Discriminant Analysis (LDA), Support Vector
Machines with linear kernel (L-SVM) and radial base function
(RBF-SVM) and Random Forest (RF). An average classification
accuracy of 70\% was obtained when using the
visible/near-infrared bands (VNIR, 450-919 nm). The inclusion of
short-wave infrared (SWIR, 1045-2400 nm) bands changed the
accuracy to 84\%. Narrow-band vegetation indices were also tested
and increased the classification accuracy by up to 5\% when
combined with VNIR features. Experimental and simulated data of
the WorldView-3 (WV-3) sensor were used for classification
purposes. While the simulated VNIR sensor bands provided an
accuracy of 57.4\%, the VNIR + SWIR set increased accuracy to
74.8\%. This pattern was also observed in the classification of
experimental WV-3 data (increase of 3.2\% after inclusion of SWIR
bands). The degree of overlap between the within- and
among-species spectral variability influenced the classifications.
The developed tree mapping method produced reliable maps of the
spatial distribution of the species and increased accuracy in
relation to pixel-level classifications by up to 6\%. The use of
a reduced set of hyperspectral bands did not significantly affect
the classification accuracies but allowed us to depict the most
important wavelengths to discriminate the species. These
wavelengths were located around the green reflectance peak (550
nm), at the red absorption feature (650 nm) and in the SWIR range
at 1200, 1700, 2100 and 2300 nm. The spectral response of the
species in the VNIR was accurately simulated by the Discrete
Anisotropic Radiative Transfer (DART) model that operates in 3D.
By means of a simplified crown structure, the spectral response of
the species at the top of the canopy was simulated with an error
of 1.5\%. The inversion of the model in the hyperspectral image
provided statistical significant relationships (R\$^{²}\$=0.65)
between chlorophyll content (C\$_{ab}\$) and MCARI (Modified
Chlorophyll Absorption in Reflectance Index), being C\$_{ab}\$
statistically different among the species. The developed modeling
approach can be used to simulate hyperspectral acquisitions of
tropical forests.",
committee = "Galv{\~a}o, L{\^e}nio Soares (presidente) and Shimabukuro, Yosio
Edemir (orientador) and Souza Filho, Carlos Roberto de
(orientador) and Arag{\~a}o, Luiz Eduardo Oliveira de Cruz de and
Almeida, Teodoro Isnard Ribeiro de and Silva, Thiago Sanna
Freire",
englishtitle = "Tree species discrimination in tropical semi-deciduous forest with
remotely sensed data and radiative transfer modeling",
language = "pt",
pages = "148",
ibi = "8JMKD3MGP3W34P/3NSEFF5",
url = "http://urlib.net/ibi/8JMKD3MGP3W34P/3NSEFF5",
targetfile = "publicacao.pdf",
urlaccessdate = "27 abr. 2024"
}