@MastersThesis{Diniz:2019:AvPoDa,
author = "Diniz, Juliana Maria Ferreira de Souza",
title = "Avalia{\c{c}}{\~a}o do potencial dos dados polarim{\'e}tricos
Sentinel-1A para mapeamento do uso e cobertura da terra na
regi{\~a}o de Ariquemes - RO",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2019",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2019-02-15",
keywords = "aprendizagem de m{\'a}quina, simula{\c{c}}{\~a}o de Monte
Carlo, radares de abertura sint{\'e}tica, banda C, Floresta
Amaz{\^o}nica, machine learning, Monte Carlo simulation,
synthetic aperture radars (SAR), C band, Amazon Forest.",
abstract = "O processo de ocupa{\c{c}}{\~a}o humana vem ocorrendo de forma
cada vez mais acentuada na Amaz{\^o}nia Legal, devido
principalmente a expans{\~a}o da fronteira agr{\'{\i}}cola.
Entender o processo de mudan{\c{c}}a no uso e cobertura da terra
{\'e} fundamental para o planejamento e gerenciamento dos
recursos naturais. Atrav{\'e}s das t{\'e}cnicas de sensoriamento
remoto {\'e} poss{\'{\i}}vel mapear o uso e cobertura da terra
de forma r{\'a}pida e eficaz. A utiliza{\c{c}}{\~a}o de dados
de radares de abertura sint{\'e}tica (SAR) vem se mostrando uma
alternativa para o monitoramento e mapeamento em regi{\~o}es
tropicais, principalmente por serem pouco influenciados pela
cobertura de nuvens. Nesse sentido, o objetivo principal desta
disserta{\c{c}}{\~a}o foi avaliar o potencial do uso das imagens
dual VV-VH do sat{\'e}lite Sentinel-1A (banda C) para o
mapeamento do uso e cobertura da terra na regi{\~a}o de
Ariquemes, RO. Para isso, foram testados sete cen{\'a}rios de
classifica{\c{c}}{\~o}es a partir dos atributos
extra{\'{\i}}dos dos dados de radar: coeficientes de
retroespalhamento, decomposi{\c{c}}{\~a}o polarim{\'e}trica e
coer{\^e}ncia interferom{\'e}trica, com os classificadores SVM
(Support Vector Machine) e RF (Random Forest). O mapeamento foi
dividido em duas fases, a Fase 1 buscando discriminar as classes
Pastagem, Agricultura e Floresta e a Fase 2, realizando-se uma
estratifica{\c{c}}{\~a}o da classe Floresta nas classes Floresta
Degradada, Floresta Prim{\'a}ria e Sucess{\~o}es
Secund{\'a}rias Avan{\c{c}}ada, Intermedi{\'a}ria e Inicial. A
valida{\c{c}}{\~a}o dos mapeamentos foi realizada atrav{\'e}s
da Simula{\c{c}}{\~a}o de Monte Carlo, utilizando-se as amostras
de campo, com 1000 itera{\c{c}}{\~o}es, onde se obteve os
valores m{\'e}dios de Kappa, acur{\'a}cia Global e matriz de
confus{\~a}o para cada cen{\'a}rio. A an{\'a}lise da
diferen{\c{c}}a estat{\'{\i}}stica entre os mapeamentos foi
realizada pelo teste de McNemar. Al{\'e}m disso, foi realizado o
mapeamento com o sensor {\'o}ptico Sentinel-2B com o objetivo de
comparar os resultados em rela{\c{c}}{\~a}o ao radar. A partir
da compara{\c{c}}{\~a}o dos cen{\'a}rios, observou-se que o
cen{\'a}rio sete, que utilizou todos os atributos em conjunto,
apresentou os melhores resultados com os classificadores SVM e RF,
com uma acur{\'a}cia global igual a 81,6% e 69,6% para as Fases 1
e 2, respectivamente, com o classificador SVM e 85,7% e 71,6% para
as Fases 1 e 2 com o classificador RF, respectivamente. Al{\'e}m
disso, o algoritmo RF apresentou superioridade para o mapeamento
em rela{\c{c}}{\~a}o ao SVM, sendo considerados diferentes
estatisticamente pelo teste de McNemar com 95% de
confian{\c{c}}a. A partir da an{\'a}lise de import{\^a}ncia das
vari{\'a}veis pelo RF, notou-se que a coer{\^e}ncia
interferom{\'e}trica foi o atributo que apresentou a maior
import{\^a}ncia para a discrimina{\c{c}}{\~a}o das classes
tem{\'a}ticas para as duas fases do mapeamento. Atrav{\'e}s da
compara{\c{c}}{\~a}o do mapeamento realizado com o cen{\'a}rio
RF-7 do Sentinel-1A com o mapeamento realizado com o sensor
{\'o}ptico, observou-se uma diferen{\c{c}}a de 4,3% na
acur{\'a}cia global para a Fase 1 e de 13,8% para a Fase 2, com o
melhor desempenho do sensor {\'o}ptico. Al{\'e}m disso,
evidenciou-se a limita{\c{c}}{\~a}o da utiliza{\c{c}}{\~a}o de
dados {\'o}pticos diante dos efeitos atmosf{\'e}ricos, como a
presen{\c{c}}a de fuma{\c{c}}a. De modo geral, notou-se que os
dados de radar foram capazes de discriminar as classes
tem{\'a}ticas analisadas, apresentando valores de acur{\'a}cia
global e Kappa considerados satisfat{\'o}rios. ABSTRACT: The
human occupation process in the Legal Amazon has occurred in an
increasingly pronounced way, especially with the expansion of the
agricultural frontier. Understanding the process of land use and
land cover change is critical to the planning and management of
natural resources. Through remote sensing techniques, it is
possible to map land use and land cover quickly and efficiently.
The use of synthetic aperture radar (SAR) data have shown to be an
alternative for monitoring and mapping in tropical regions, mainly
because they are little influenced by cloud cover. So, the main
objective of this dissertation was to evaluate the potential of
the VV-VH dual images of the Sentinel-1A satellite (band C) for
mapping land use and land cover in the region of Ariquemes, RO.
Seven classification scenarios were tested using the attributes
extracted from the radar data: backscatter coefficients,
polarimetric decomposition and interferometric coherence, and the
SVM (Support Vector Machine) and RF (Random Forest) classifiers.
The mapping process was divided into two phases, Phase 1, which
aimed to discriminate the classes Pasture, Agriculture and Forest,
and Phase 2, whose Forest class was stratified in the classes:
Degraded Forest, Primary Forest and Advanced, Intermediate and
Initial Successions. The validation of the mappings was done
through the Monte Carlo Simulation, using the field samples, with
1000 iterations, where the mean values of Kappa, Global Accuracy
and confusion matrix were obtained for each scenario. The analysis
of the statistical difference between the mappings was performed
by the McNemar test. In addition, it was carried out the mapping
using the Sentinel-2B optical image in order to compare to the
results obtained using the radar data. From the scenario
comparison, it was observed that scenario seven, which used all
the attributes together, presented the best results for both SVM
and RF classifiers, with a global accuracy of 81.6% and 69.6% for
Phases 1 and 2, respectively, with SVM classifier and 85.7% and
71.6% for Phases 1 and 2 with Rf classifier, respectively. In
addition, RF algorithm showed superiority over the SVM algorithm
for mapping the region. The mappings obtained using the RF and SVM
classifiers were considered statistically different by the McNemar
test with 95% confidence. From the analysis of variable importance
by the RF, it was noticed that the interferometric coherence was
the attribute that presented the most importance for the
discrimination of the thematic classes for the two phases of the
mapping. By comparing the mapping performed with the Sentinel-1A
RF-7 scenario with the mapping performed with the optical sensor,
a difference of 4.3% was observed in the overall accuracy for
Phase 1 and 13.8% for Phase 2, with the best optical sensor
performance. In addition, it was evidenced the limitation of
optical data to the atmospheric effects, such as the presence of
smoke. In general, it was observed that the radar data were able
to discriminate the classes of land use and land cover analyzed in
this dissertation, presenting values of global accuracy and Kappa
considered satisfactory.",
committee = "Shimabukuro, Yosio Edemir (presidente) and Gama, F{\'a}bio Furlan
(orientador) and Mura, Jos{\'e} Claudio and Araujo, Luciana
Spinelli",
englishtitle = "Evaluation of the potential of polarimetric data sentinel-1a for
mapping land use and land cover at the ariquemes-RO region",
language = "pt",
pages = "132",
ibi = "8JMKD3MGP3W34R/3SL65N8",
url = "http://urlib.net/ibi/8JMKD3MGP3W34R/3SL65N8",
targetfile = "publicacao.pdf",
urlaccessdate = "27 abr. 2024"
}