Abstracting GIS layers from hyperspectral imagery

The spectral-spatial relationship of materials in a hyperspectral image cube is exploited to partially automate the creation of geographic information system (GIS) layers. The topological neighborhood preservation property of the self-organizing map (SOM) is clustered into six (partially overlapping) neighborhoods that are mapped into the image domain to locate in-scene structures of similar material type. GIS layers are abstracted through spatial logical and morphological operations on the six image domain material maps and a novel road finding algorithm connects road segments under significant tree-occlusion resulting in a contiguous road network. It is assumed that specific knowledge of the scene (e.g. endmember spectra) is not available. The results are eight separate high-quality GIS layers (vegetation, trees, fields, buildings, major buildings, roadways, and parking areas) that follow the scene features of the hyperspectral image and are separately and automatically labeled. The material maps resulting from clustering the SOM have an 84.3% average accuracy, which increases to 93.9% after spatial processing into GIS layers.