Automatic extraction of man-made objects from high-resolution satellite imagery by information fusion

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] Recently available high-resolution commercial satellite imagery provides an important new data source for remote sensing applications. Automated feature extraction (AFE) techniques can assist human analysts by rapidly locating geospatial information and have the potential to significantly reduce the amount of time to process and analyze geospatial data. In this research, we have designed and developed systems for automatic extraction of man-made objects (roads, buildings and vehicles) from high-resolution satellite imagery. We conclude that AFE can be greatly enriched and improved by multiinformation fusion and/or multi-cue integration. For road extraction and building extraction respectively, multiple detectors were developed and the extraction performance was greatly improved using multi-detector fusion from different information sources. For vehicle detection, a GIS road vector layer was used to incorporate contextual information and an implicit vehicle model including spectral and spatial characteristics was learned by a morphological shared-weight neural network. An important characteristic of our research on road and building extraction is that our extraction strategies are fully automated with only a few preset parameters. Compared with related research in these areas, the performance evaluations of our extraction systems are among the highest statistical values reported in literature thus far.