A knowledge-based approach to mapping roads from aerial imagery using a GIS database
thesisposted on 2023-05-26, 20:38 authored by Forghani, Ali
Conventional image classification approaches may be inadequate for extraction of complex and spectrally heterogeneous land use classes from remotely sensed imagery. The integration of spatial data with remotely sensed data has the potential to improve significantly the reliability of feature classification. Thus it is informative to use contextual and textural information in the classification process. This thesis describes a methodology developed to integrate GIS and aerial imagery in a manner that allows it to be used in a knowledge-based analysis system. Using a trial site and aerial photography, the methodology was implemented and tests indicate the technique works well in mapping of roads when roads pass through a rural area where the contrast is high, but fails in urban areas where the roads are confused with man-made structures. Also, a supervised multispectral image classification of the trial site using colour aerial photography was carried out to compare the performance of a supervised multispectral image analysis with the decision tree analysis to map out roads over the trial site. A classification accuracy assessment shows that the overall classification accuracy was marginally lower than the decision tree analysis. The GIS data used in the knowledge-base included a DTM and land use covers. For this research, part of the data was already available in digital format. In practice, it may be that a DTM and land use classification would need to be created from aerial photography or satellite imagery. It is in this context that the methodology developed here is most likely to improve significantly attribute-based classification. The GIS database included geometrically rectified aerial photography, roads, land use, drainage pattern, field and vegetation boundaries, DTM, and edge detection data. A program was developed for semi-automatic linear feature detection using different edge detectors, in which the process is followed by morphological operations. The extracted edges (lines) were used as a GIS layer in a later step of the methodology. Grid raster-based processing was undertaken to build a multi-source database in the GIS to be used for knowledge-based analysis. The multi-layer database was interfaced with decision tree software for creation of a classification tree. The independent data set comprised six variables, representing the contextual, textural, and geometrical characteristics of the knowledge-based data. In the process of decision tree analysis, the input data was recursively partitioned into mutually clustered, exhaustive subsets which define the best response variable. The resulting classification tree was used to generate generic rules for implementation of an expert system. The developed expert system was used to map out the spatial distribution of the grid data to show areas with roads (presence) and their background (absence). The output of this model is encouraging when applied over homogeneous rural scenes, but there are difficulties over heterogeneous urban areas. The results show that a framework of roads in a rural site mapped by this knowledge-based technique closely concurred with visual interpretation. This research devised a general approach to solving problems of road identification. This approach can serve as a model for practitioners who are trying to do practical work in this field. By generating a hybrid system which locates many different databases and integrates many different sources of knowledge in attempting to identify a specific (man-made) geographic feature, and by utilising current artificial intelligence (AI) techniques to perform the classification, this research provides an early example of the techniques which will be in more general use in the areas of GIS and remote sensing in the future. The methodology developed here is costly and data-intensive. Since the technique investigated in this research requires a large number of data sets to be built, construction of the data is relatively expensive over large areas. The initial costs involved in configuring a knowledge-base, such as the methodology developed in this study, are high, and this may not be justifiable in a production environment.
Rights statementCopyright 1997 the Author - The University is continuing to endeavour to trace the copyright owner(s) and in the meantime this item has been reproduced here in good faith. We would be pleased to hear from the copyright owner(s). Thesis (Ph.D.)--University of Tasmania, 1998. Includes bibliographical references