%0 Conference Proceedings
%4 sid.inpe.br/marte2/2017/10.27.12.50.46
%2 sid.inpe.br/marte2/2017/10.27.12.50.47
%@isbn 978-85-17-00088-1
%F 59456
%T Efeito das estratégias de treinamento na exatidão dos modelos preditivos em classificação de imagens
%D 2017
%A Zeferino, Leiliane Bozzi,
%A Souza, Ligia Faria Tavares de,
%A Fernandes Filho, Elpidio Inácio,
%A Oliveira, Teógenes Senna,
%@electronicmailaddress leiliane.zeferino@ufv.br
%E Gherardi, Douglas Francisco Marcolino,
%E Aragão, Luiz Eduardo Oliveira e Cruz de,
%B Simpósio Brasileiro de Sensoriamento Remoto, 18 (SBSR)
%C Santos
%8 28-31 maio 2017
%I Instituto Nacional de Pesquisas Espaciais (INPE)
%J São José dos Campos
%P 2844-2851
%S Anais
%1 Instituto Nacional de Pesquisas Espaciais (INPE)
%X This paper aims to evaluate the effect of training samples as points and polygons related to four different approaches by the Gradient Boosting Machine (GBM) and Random Forest (RF) classifiers on the accuracy of the supervised classification. The process of image classification used 11 predictive co-variables, six were spectral co-variable, and the remaining 5 were related to topography, climate, geology and pedology. The treatments were: 1 separation of training and validation data without considering the origin polygon; 2 separation of training and validation sample considering the origin polygons; 3 random choice of a pixel from each polygon to represent this; and 4 median and mode statistics calculated to each numeric and categorical co-variable, respectively. The results concluded that the treatments with the samples as points, treatments 3 and 4, presented the lowest average. Those with the sample as polygons diverged among itself, and treatment 1 presented the highest average. The validation with external data made the approaches became similar, where treatment 1 and 2, the same way as treatments 3 and 4, did not show any statistical difference, which can be concluded that this is necessary to the generation of more reliable predictive models. The effect of the sample collection in polygons to latter obtain a point, as approached by treatment 4, showed to be superior when compared to those that used the information as a whole polygon, even presenting lowest kappa index values.
%9 Classificação e mineração de dados
%@language pt
%3 59456.pdf