An object-based approach to detect road features for informal settlements near Sao Paulo, Brazil

Abstract

Continuous mapping efforts have been required to monitor intense urbanization processes of large cities of developing countries. The uncontrolled sprawl occurring in the vicinity of Sao Paulo, Brazil since the 70’s illustrates this scenario. Considering that urban sprawl causes changes to road networks, monitoring new roads as well as the changes along existing roads can provide significant information for urban management. Due to the lack of coverage in historical and accurate aerial image and map products, existing or even outdated image data are unavailable for planning urban land use with substantial relevance. The recent availability of high resolution satellite images, beginning with the IKONOS II in 1999, has enabled urban applications of Remote Sensing. Unfortunately, traditional techniques employed to detect land cover information based on per-pixel analysis have yielded unsatisfactory results in urban application of high resolution satellite images. In this sense, enhanced capabilities and successful applications of object-based classification have stimulated research to develop new methodologies to provide geoinformation. To this end, road extraction research has been formulated to segment object-primitives from images and to use the resultant information to devise enhancements to improve the road detection and classification process. This chapter reports the use of object-based image classification applied to road detection in informal settlements areas. An 11-bit IKONOS image was employed as the primary remote sensing data for classification. Principal components and semi-homogeneous segmented area products (segmentation products) were computed and used to define custom features. Auxiliary data were calculated from spectral information in combination with geometric information extracted from segments. Contextual information was also employed to support the implementation of a classification rule base. The classification rule base eliminated vegetated areas and then considered impervious surface and bare soil areas, as well as the width, length, asymmetry and the neighborhood relationship for the objects to detect road features. Comparisons between the automatic approach results and manually extracted road feature areas delivered insight regarding omission and commission error by area counting as well as metrics employed to determine completeness and correctness of extracted road features by linear correspondence analysis attest to the efficiency of the methodology. Results indicate that the methodology produces significant information and offers improvements over traditional pixel-based methods of road extraction and classification.Keywordsremote sensingclassificationurbansprawlhigh resolution imagelinear correspondence analysis