Semi-Automated Landscape Feature Extraction and Modeling

We have developed a semi-automated procedure for generating correctly located 3D tree objects from overhead imagery. Cross-platform software partitions arbitrarily large, geocorrected and geolocated imagery into manageable sub-images. The user manually selects tree areas from one or more of these sub-images. Samples are taken from these areas, and color statistics are computed. Tree areas are detected in subsequent images. Tree group blobs are then narrowed to lines using a special thinning algorithm which retains the topology of the blobs, and also stores the thickness of the parent blob. Maxima along these thinned tree groups are found, and used as individual tree locations within the tree group. Magnitudes of the local maxima are used to scale the radii of the tree objects. Grossly overlapping trees are culled based on a comparison of tree-tree distance to combined radii. Tree color is randomly selected based on the distribution of sample tree pixels, and height is estimated from tree radius. The final tree objects (perpendicular intersecting tree cutouts) are then inserted into a terrain database which can be navigated by VGIS, a high-resolution global terrain visualization system developed at Georgia Tech.