DBPapers
DOI: 10.5593/SGEM2016/B22/S10.120

LINE STRUCTURE-BASED TRAINING DATA EXTRACTION FOR BUILDING CLASSIFICATION OF HIGH RESOLUTION SATELLITE IMAGES

J. Yeom, Y. Kim
Thursday 8 September 2016 by Libadmin2016

References: 16th International Multidisciplinary Scientific GeoConference SGEM 2016, www.sgem.org, SGEM2016 Conference Proceedings, ISBN 978-619-7105-59-9 / ISSN 1314-2704, June 28 - July 6, 2016, Book2 Vol. 2, 939-946 pp

ABSTRACT
Training data for building classification are generally extracted through the use of
auxiliary data or manual interpretation. However, geometric coincidence between
satellite images and auxiliary data is not often guaranteed. In addition, human manual
operations are quite subjective and labor-intensive. Therefore, training data should be
extracted objectively and automatically to ensure computational efficiency. In this
study, lines, which are one of the representative features in high resolution satellite
images, were used for the determination of building training data. Line features were
extracted using regular grid-based Hough transform. In particular, lines with orthogonal
geometry were selectively used to define corner regions. Building training data were
finally refined based on the spectral characteristics inside and outside of corner regions.
Classification was conducted using a Support Vector Machine classifier. The results of
the building classification were then compared with those of a Geographic Information
System-based classification. The accuracy of the proposed method was greater than
80% and came close to the Geographic Information System-based result that used exact building boundary data. The proposed method, line structure-based training data
extraction, contributes to the automation and objectification of building classification.

Keywords: line structure, regular grid-based Hough transform, training data extraction,
building classification, high resolution satellite image