Application of spatial reasoning methods to the extraction of roads from high resolution satellite imagery

By the end of this decade the civilian remote sensing community will have access to timely satellite imagery at high spatial resolution (1-3 meters). This will open the door for new applications, such as detailed topographic mapping. These data can be considered to be in a 'transition' spatial resolution regime between conventional, low resolution satellite imagery and aerial photography. Correspondingly, information extraction methodologies to exploit these data must draw on both the data-driven, per-pixel processing of conventional satellite image analysis and the object-driven, image understanding technologies now being developed for aerial photography. This paper describes experiments to evaluate rule-based recognition algorithms for the purpose of automating planimetric feature extraction from these new satellite data. The overall feature extraction strategy is one drawn from image understanding namely spatial reasoning with a segmented rendition of the image. Recognition involves applying a set of evidence-accumulation attribute (inherent spatial/spectral and context) tests to selected segments in order to identify candidates which may form all or part of an object of interest. Conventional classification 'training' has been modifed to develop a novel approach to evidence weight quantification and to assess inter-test correlation, analogous to conventional covariance. A recognition system has been developed to recognize residential streets in imagery.