%0 Conference Proceedings
%4 sid.inpe.br/mtc-m16d/2019/11.27.13.54
%2 sid.inpe.br/mtc-m16d/2019/11.27.13.54.11
%@issn 2179-4847
%T Exploiting parallelism to generate meta-features for land use and land cover classification with remote sensing time series
%D 2019
%A Oliveira, Sávio S. T. de,
%A Cardoso, Marcelo de C.,
%A Bueno, Elivelton,
%A Rodrigues, Vagner J. S.,
%A Martins, Wellington S.,
%@affiliation Universidade Federal de Goiás (UFG)
%@affiliation Universidade Federal de Goiás (UFG)
%@affiliation Universidade Federal de Goiás (UFG)
%@affiliation Universidade Federal de Goiás (UFG)
%@affiliation Universidade Federal de Goiás (UFG)
%@electronicmailaddress savio.teles@gogeo.io
%@electronicmailaddress marcelo.cardoso@gogeo.io
%@electronicmailaddress elivelton.bueno@gogeo.io
%@electronicmailaddress vagner@gogeo.io
%@electronicmailaddress wellington@inf.ufg.br
%E Lisboa Filho, Jugurta,
%E Monteiro, Antonio Miguel Vieira,
%B Simpósio Brasileiro de Geoinformática, 20 (GEOINFO)
%C São José dos Campos
%8 11 -13 nov. 2019
%I Instituto Nacional de Pesquisas Espaciais (INPE)
%J São José dos Campos
%S Anais do 20º Simpósio Brasileiro de Geoinformática
%1 Instituto Nacional de Pesquisas Espaciais (INPE)
%K geoinformatica.
%X The automatic classification of remote sensing time series has become essential to identify the rapid and frequent changes that the earths surface has been undergoing. This work investigates the accuracy of land use and land cover classification with remote sensing time series when distance based metafeatures are added to existing features of some classifiers. The distance based meta-features presented are generated by comparing all time series of the region being studied to every time series patterns previously calculated for that region. This is a very costly operation that was made viable through the use of parallel processing. Although expensive, this operation is advantageous because the meta-features generated can be later used as input to any classifier. The experimental work conducted showed promising results when using the distance based meta-feature strategy. The proposed strategy was able to increase from 78% to 93,8% the classification accuracy of the KNN algorithm, and from 92,3% to 93,8% the accuracy of a state-of-art SVM-based algorithm proposed recently. These results indicate that distance-based meta-features allow revealing unknown data characteristics, potentially increasing classification accuracy.
%@language pt
%3 135-146.pdf