In this paper we present a novel method of road detection in aerial and satellite imaging. This structural method is based on the concept of profile, meant as a local one-dimensional cross-section (cast) of raster image. We acquire such profiles from the image at different orientation angles and extract from them features well discriminating road pixels from non-road pixels. In particular, we use feature definitions tailored to road characteristics (mostly elongation); these include, among others, mutual similarity of close and equally orientated profiles (road continuity) and symmetry. To improve precision of analysis, the method computes profiles using sub-pixel sampling. The further part of processing relies on machine learning, in particular on supervised learning from examples. The algorithm is given a training sample of pixels, for which the decision class assignment (road, non-road) is known. This information may be manually entered by a decision maker (expert) by marking image regions representing road fragments, or alternatively it may be retrieved from an appropriate module of a geographical information system. Given that information, the algorithm acquires the knowledge from training examples performing so-called inductive learning. That knowledge may be then used to classify the remaining image pixels, for which the decision class assignment is not known. Moreover, the knowledge may be inspected (and potentially corrected) by the decision maker, as it is expressed in readable form of a decision tree. The paper presents the algorithm in detail, describes its computer implementation, and demonstrates its application to an aerial image of urban area. The obtained results demonstrate good performance of the method and indicate usefulness of machine learning approach in analysis of aerial and satellite imagery.
@INPROCEEDINGS { KrawiecWyczalek06,
ABSTRACT = { In this paper we present a novel method of road detection in aerial and satellite imaging. This structural method is based on the concept of profile, meant as a local one-dimensional cross-section (cast) of raster image. We acquire such profiles from the image at different orientation angles and extract from them features well discriminating road pixels from non-road pixels. In particular, we use feature definitions tailored to road characteristics (mostly elongation); these include, among others, mutual similarity of close and equally orientated profiles (road continuity) and symmetry. To improve precision of analysis, the method computes profiles using sub-pixel sampling. The further part of processing relies on machine learning, in particular on supervised learning from examples. The algorithm is given a training sample of pixels, for which the decision class assignment (road, non-road) is known. This information may be manually entered by a decision maker (expert) by marking image regions representing road fragments, or alternatively it may be retrieved from an appropriate module of a geographical information system. Given that information, the algorithm acquires the knowledge from training examples performing so-called inductive learning. That knowledge may be then used to classify the remaining image pixels, for which the decision class assignment is not known. Moreover, the knowledge may be inspected (and potentially corrected) by the decision maker, as it is expressed in readable form of a decision tree. The paper presents the algorithm in detail, describes its computer implementation, and demonstrates its application to an aerial image of urban area. The obtained results demonstrate good performance of the method and indicate usefulness of machine learning approach in analysis of aerial and satellite imagery. },
ADDRESS = { Olsztyn },
AUTHOR = { Krzysztof Krawiec and Ireneusz Wycza{\l}ek },
BOOKTITLE = { Materia{\l}y Og{\'o}lnopolskiego Sympozjum Naukowego Opracowania cyfrowe w Fotogrametrii, Teledetekcji i GIS },
NOTE = { in Polish },
OWNER = { krawiec },
TIMESTAMP = { 2006.10.06 },
TITLE = { Supervised Road Detection Using Machine Learning Methodology },
YEAR = { 2006 },
}