@MastersThesis{Souto:2000:SeImMu,
author = "Souto, Roberto Pinto",
title = "Segmenta{\c{c}}{\~a}o de imagem multiespectral utilizando-se o
atributo matiz",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2000",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2000-08-29",
keywords = "sensoriamento remoto, cores, classifica{\c{c}}{\~a}o de imagens,
an{\'a}lise estat{\'{\i}}stica, programa{\c{c}}{\~a}o, remote
sensing, color, image classification, statistical analysis,
computer programming.",
abstract = "Este trabalho teve como objetivo, a implementa{\c{c}}{\~a}o e
avalia{\c{c}}{\~a}o de algoritmos de segmenta{\c{c}}{\~a}o e
classifica{\c{c}}{\~a}o de imagens de matiz de
composi{\c{c}}{\~a}o colorida. Desejava-se averiguar o
comportamento do matiz em regi{\~o}es onde h{\'a}
diferen{\c{c}}as na luminosidade em fun{\c{c}}{\~a}o da
topografia do terreno em dois alvos: floresta e urbano. Neste
aspecto, os resultados alcan{\c{c}}ados foram satisfat{\'o}rios,
pois o algoritmo implementado conseguiu resolver estes dois casos
na maior parte das vezes. Normalmente as imagens coloridas
s{\~a}o resultado de composi{\c{c}}{\~o}es com tr{\^e}s bandas
espectrais. Implementou-se um m{\'e}todo que {\'e} capaz de
obter o valor de matiz diretamente de quaisquer N bandas
espectrais. No entanto, nem sempre o acr{\'e}scimo de bandas
resultou em uma classifica{\c{c}}{\~a}o melhor. Isto depende de
quais bandas se escolhe a fim de gerar a imagem de matiz. Foram
gerados resultados de classifica{\c{c}}{\~a}o de matiz obtidos
atrav{\'e}s de composi{\c{c}}{\~o}es coloridas com tr{\^e}s,
quatro e cinco bandas espectrais. Avaliou-se tamb{\'e}m o
comportamento do matiz nestas tr{\^e}s composi{\c{c}}{\~o}es,
na distin{\c{c}}{\~a}o do alvo floresta em regi{\~o}es de
relevo muito acidentado, com presen{\c{c}}a forte de sombra. A
avalia{\c{c}}{\~a}o, tanto para o alvo floresta quanto para o
urbano, se deu atrav{\'e}s da compara{\c{c}}{\~a}o dos
resultados de classifica{\c{c}}{\~a}o com pontos de
refer{\^e}ncia classificados previamente por
fotoint{\'e}rpretes. Foi calculado de cada
classifica{\c{c}}{\~a}o, a partir dos pontos de refer{\^e}ncia,
um coeficiente kappa. Os diversos kappas estimados encontrados
foram comparados atrav{\'e}s de teste de hip{\'o}tese para se
verificar se havia diferen{\c{c}}as significativas entre os
resultados de classifica{\c{c}}{\~a}o de matiz alcan{\c{c}}ados
com tr{\^e}s, quatro e cinco bandas. Notou-se que h{\'a} um
desempenho superior usando quatro bandas para distinguir alvo
urbano, mas n{\~a}o foram percebidas diferen{\c{c}}as
significativas no alvo floresta. Mesmo procedimento foi feito para
comparar estas classifica{\c{c}}{\~o}es de matiz com o
m{\'e}todo de classifica{\c{c}}{\~a}o n{\~a}o-supervisionada
Isoseg, implementado no {"}software{"} SPRING. Na maior parte das
vezes o resultado de matiz superou o Isoseg, havendo no entanto,
para cinco bandas, resultados de classifica{\c{c}}{\~a}o de
Isoseg melhores que no matiz. Paralelamente a esta
avalia{\c{c}}{\~a}o por coeficiente kappa, foram desenvolvidos
algoritmos de avalia{\c{c}}{\~a}o automatizada dos erros de
classifica{\c{c}}{\~a}o, levando-se em considera{\c{c}}{\~a}o
a distribui{\c{c}}{\~a}o estat{\'{\i}}stica (binomial) destes
erros. Basicamente, esta tarefa tem por finalidade saber se o
tamanho da amostra de pontos escolhido {\'e} adequado. Isto
{\'e}, se o produtor n{\~a}o corre risco demasiado ao coletar um
tamanho pequeno de pontos para avaliar a
classifica{\c{c}}{\~a}o. Ou ainda se compensa o custo de coletar
amostragem muito grande, para correr um risco menor de ver seu
mapa rejeitado. Todos recursos de classifica{\c{c}}{\~a}o e
avalia{\c{c}}{\~a}o s{\~a}o acessados atrav{\'e}s de uma
interface gr{\'a}fica desenvolvida neste trabalho. ABSTRACT: This
work had as objective, the implementation and evaluation of
segmentation algorithms and classification of hue images from
color composit. It was desired to inquire the behavior of the hue
in regions where it has differences in the luminosity due to the
topography of the land in two targets: forest and urban. In this
aspect, the reached results had been satisfactory, since the
implemented algorithm obtained solved these two cases in the most
part of the times. Normally the color images are resulted of
composition with three spectral bands. A method was implemented
that is able to directly get the value of hue of any N spectral
bands. However, nor always the upgrade of bands mean a better
classification. It will depend on which bands are chosen, in order
to generate the hue image. Results had been generated by
classification of hue through color composit of three, four and
five spectral bands. The behavior of the hue in these three cases,
the distinction of the white forest in relief regions, with a
strong presence of shade was also evaluated. The evaluation, as
much for the white forest as for the urban one, was made through
the matching of the results of classification with control points
classified previously by photointerpreters. It was calculated of
each classification, from the control points, a coefficient kappa.
Diverse kappas estimated found had been compared through
hypothesis test to verify itself if it had significant differences
between the results of reached classificaton of hue from three,
four and five bands. It was noticed that it has an upper
performance using four bands to distinguish urban target, but had
not been perceived significant differences in the forest. Same
procedure was made to compare these classificatons of hue with the
method of unsupervised classifier Isoseg, implemented at SPRING
software. At the most part of the times, hue results overcame
Isoseg however. However, for five bands, there are better results
of classification by Isoseg method. It had been developed
algorithms of automatized evaluation of the errors of
classification, taking in account the statistical distribution
(binomial) of these errors. Basically, this task has as purpose to
know if the size of the sample of points chosen is correct. That
is, if the producer does not has too much risk when collecting a
small size of points to evaluate the classification. Or, besides,
if it compensates the cost to collect sampling very great, in
order to have a lesser risk to see its map rejected wrongly. All
features of classification and evaluation can be accessed through
a graphical user interface developed in this work.",
committee = "Dutra, Luciano Vieira (presidente/orientador) and Banon, Gerald
Jean Francis and Valeriano, Dalton de Morisson and Mattos,
Ju{\'e}rcio Tavares and Fernandes, David",
copyholder = "SID/SCD",
englishtitle = "Multispectral image segmentation using the hue attribute",
language = "pt",
pages = "173",
ibi = "6qtX3pFwXQZ3P8SECKy/Ae2JT",
url = "http://urlib.net/ibi/6qtX3pFwXQZ3P8SECKy/Ae2JT",
targetfile = "publicacao-24.pdf",
urlaccessdate = "06 maio 2024"
}