Automated Footprint Generation from Geotags with Kernel Density Estimation and Support Vector Machines

Abstract

A well-known problematic issue for the gazetteer services that many geospatial applications depend on is the lack of spatial footprints of imprecise regions. We present an automated method of footprint generation based on the statistical evaluation of a set of points, which are assumed to lie in the region. Two statistical methods, Kernel Density Estimation and Support Vector Machines (SVMs), are applied and compared for this task. The overall approach is evaluated using precise regions, and the results obtained from the two classes of methods are evaluated by means of statistical classification measures showing a slight superiority of SVMs. Finally, a priori choices for the input parameters of the methods are derived from the results and footprints of imprecise regions are generated in a completely automated process. The input dataset is acquired from georeferenced photographs freely available on the web.

Keywords:

• gazetteer
• placename
• footprint
• imprecise region
• kernel density estimation
• support vector machines
• geotag

Notes

¹Flickr (http://www.flickr.com) alone has more than 60 million geotagged photos online as of April 2008 (Catt, 2008).

²A photo of a Black Forest cake with the tag “black forest” would be an example where place semantics are falsely assigned to the tag by our assumption.

³Photos tagged with the English name and the local name (or names) qualified for the set.

⁴Albania, Belarus, Croatia, France, Germany, Greece, Ireland, Italy, Luxembourg, Switzerland, and Ukraine

⁵A caveat: The precision does not become 1, if the highest point density is located outside R*.

⁶To understand the shapes of the graphs consider the “behavior” of R ^KDE _S (b) for varying bias values as described before.

⁷This must be attributed to our choice of R*, where the coastline marks the footprint's boundary. Other choices for R* are imaginable, for instance including the sea between the mainland and the islands.

⁸In the case of the Alps, “Alps”, “Alpi”, “Alpes”, “Alpen”, and “Alpe” are cosidered the “same name” in this sense.