Estimating the Spatial Extents of Geospatial Objects Using Hierarchical Models

Abstract

The goal of this work is to estimate the spatial extents of complex geospatial objects such as high schools and golf courses. Gazetteers are deficient in that they currently specify the spatial extents of these objects using a single latitude/longitude point. We propose a framework that uses readily available high resolution overhead imagery to estimate the boundaries of known object instances in order to update the gazetteers. Key to our approach is a hierarchical object model with three levels. The lowest level characterizes an object using local invariant features; an intermediate, latent level characterizes the land-use/land-cover (LULC) classes that constitute an object; and, the top level models an object as a distribution over these classes. We evaluate our approach using a manually labeled ground truth dataset of four object types: high schools, golf courses, mobile home parks, and Costco shopping centers.