Proximate sensing using georeferenced community contributed photo collections

Abstract

Volunteered geographic information such as that available in blogs, wikis, social networking sites, and community contributed photo collections is enabling new applications. This work investigates the use of georeferenced images from a popular photo sharing site for proximate sensing. In particular, we use computer vision and machine learning techniques to perform land cover classification based on the content of the georeferenced images. We evaluate the results using a ground truth dataset from the National Land Cover Database. We demonstrate that our approach can achieve upwards of 75% classification accuracy in a completely automated fashion.