%0 Conference Proceedings
%4 dpi.inpe.br/marte/2011/06.27.14.01
%2 dpi.inpe.br/marte/2011/06.27.14.01.32
%@isbn 978-85-17-00056-0 (Internet)
%@isbn 978-85-17-00057-7 (DVD)
%T Uso das redes neurais de função de base radial e Growing Neural Gas na classificação de imagens de sensoriamento remoto
%D 2011
%A Lima, Alexandre Gomes de,
%A Guerreiro, Ana Maria Guimarães,
%@affiliation Instituto Federal de Educação, Ciência e Tecnologia do RN - IFRN/DIETINF
%@affiliation Universidade Federal do Rio Grande do Norte – UFRN/DCA
%@electronicmailaddress alexandre.lima@ifrn.edu.br
%@electronicmailaddress anamaria@dca.ufrn.br
%E Epiphanio, José Carlos Neves,
%E Galvão, Lênio Soares,
%B Simpósio Brasileiro de Sensoriamento Remoto, 15 (SBSR).
%C Curitiba
%8 30 abr. - 5 maio 2011
%I Instituto Nacional de Pesquisas Espaciais (INPE)
%J São José dos Campos
%P 7247-7254
%S Anais
%1 Instituto Nacional de Pesquisas Espaciais (INPE)
%K remote sensing, image processing, artificial neural network, growing neural gas, sensoriamento remoto, processamento de imagens, rede neural artificial.
%X Several works describe the use of artificial neural networks in remote sensing applications. However, its relatively scarce the amount of publications about these applications involving the Growing Neural Gas (GNG) networks. This work describes a hybrid neural classifier based on the radial base functions (RBF) and GNG neural networks. The unsupervised learning is performed by GNG network in order to determine centers and number of hidden neurons RBF network. The supervised learning is performed by pseudo inverse matrix algorithm in order to find RBF networks synaptic weights. The proposed classifier doesnt require the number of centers be specified in advance. This number starts in two and is successively increased by GNG network until a desired performance criterion be achieved. The classification of one multispectral ETM/Landsat7 image, bands 1, 2, 3 and 4, involving part of city of Natal-RN is performed for seven ground cover classes. The results achieved by the proposed classifier and maximum likelihood Bayesian classifier are compared through the confusion matrix, hit coefficient, Kappa coefficient and generated images. This set of data shows a slight superiority from proposed classifier. Further this satisfactory result, the neural classifier is important to eliminate the try-and-error procedure usually realized to find RBF neural networks centers.
%9 Processamento de Imagens
%@language pt
%3 p0398.pdf