A Supervised Neural Linear Feature Extractor for Remotely Sensed Data
Genong
Yu, Indiana State University, gyu@indstate.edu
(Presenting)
Ryan
R
Jensen, Indiana State University, r-jensen@indstate.edu
Paul
W
Mausel, Indiana State University, gemause@isugw.indstate.edu
Linear features in remotely sensed data are important for classification and visual interpretation. A supervised neural network linear feature extractor that uses both spatial and spectral characteristics of linear features in a remotely sensed image was developed and implemented in the object-oriented Amazon Information System (AIS) built for a NASA funded LBA ECO project. This study compared the supervised neural network linear feature extractor with conventional extractors, i.e. maximum likelihood classifier (MLC), Canny edge detector, and Hough transform line detector. The study compared the following three aspects: (1) multiple spatial resolution images, including IKONOS, ASTER and Landsat TM data, (2) two different types of linear features, rivers and roads, and (3) different post-processing algorithms, including the Hough transform. A supervised neural network linear feature detector was developed that combines the sensitivity of a neural network to spectral information at different scales with the ability of neural network to utilize local information through model-based line detection. Analysis of results indicate that the neural network approach used is superior to a conventional MLC-based edge detection or spectral classification in all three aspects investigated.
Keywords: linear feature, multiple layer perceptron, multiscale
Submetido por Genong Yu em 17-MAR-2004
Tema Científico do LBA: LC (Mudanças dos Usos da Terra e da Vegetação)