@InProceedings{LacerdaAlmeGalv:2017:AvAcSe,
author = "Lacerda, Camila Souza dos Anjos and Almeida, Cl{\'a}udia Maria de
and Galv{\~a}o, L{\^e}nio Soares",
affiliation = "{} and {Instituto Nacional de Pesquisas Espaciais (INPE)} and
{Instituto Nacional de Pesquisas Espaciais (INPE)}",
title = "Classifica{\c{c}}{\~a}o por {\'a}rvores de decis{\~a}o:
Avalia{\c{c}}{\~a}o de acur{\'a}cia com e sem a
pr{\'e}-sele{\c{c}}{\~a}o de atributos",
booktitle = "Anais...",
year = "2017",
editor = "Gherardi, Douglas Francisco Marcolino and Arag{\~a}o, Luiz
Eduardo Oliveira e Cruz de",
pages = "6407--6414",
organization = "Simp{\'o}sio Brasileiro de Sensoriamento Remoto, 18. (SBSR)",
publisher = "Instituto Nacional de Pesquisas Espaciais (INPE)",
address = "S{\~a}o Jos{\'e} dos Campos",
abstract = "The use of decision trees for image classification has grown
rapidly in recent years, since the reported developments are
promising. The current work uses the C4.5 and Random Forest
methods for selecting statistical and customized attributes
derived from a WorldView-2 (WV-2) image, meant to effectively
separate the classes of interest. The attributes comprising the
bands extracted from transforms such as Principal Component
Analysis and Minimum Noise Fraction in addition to vegetation
indices and attributes based on band ratios were used in the stage
of feature selection. Two datasets have been employed in the
herein described experiments: one consisting of a WV-2 scene with
42 image-derived attributes, and the second one containing the
same WV-2 scene and 28 pre-selected meaningful attributes.
Comparisons between the two datasets showed that for both methods
(C4.5 and RF) the use of statistical attributes plus those derived
from image transforms and arithmetical operations increases the
agreement indices accuracy. Both classifiers work as data miners,
identifying among a large set of attributes those able to
discriminate the concerned classes. The results of this article
comply with the peer-reviewed literature, for they demonstrate
that the classification integrating a greater number of input
attributes (without feature selection) attain a significantly
superior accuracy. In sum, decision tree classifiers have the
capacity to deal with weak explanatory variables, and hence, it is
not possible to assess their importance based on individual
factors alone.",
conference-location = "Santos",
conference-year = "28-31 maio 2017",
isbn = "978-85-17-00088-1",
label = "59385",
language = "pt",
organisation = "Instituto Nacional de Pesquisas Espaciais (INPE)",
ibi = "8JMKD3MGP6W34M/3PSMCRE",
url = "http://urlib.net/ibi/8JMKD3MGP6W34M/3PSMCRE",
targetfile = "59385.pdf",
type = "Classifica{\c{c}}{\~a}o e minera{\c{c}}{\~a}o de dados",
urlaccessdate = "27 abr. 2024"
}