Skip to main content
Open Access Publications from the University of California

Proximate Sensing: Geographic Knowledge Discovery in On-line Photo Collections

  • Author(s): Leung, Chi Yan
  • Advisor(s): Newsam, Shawn
  • et al.

On-line photo sharing websites such as Flickr not only allow users to share their precious memories with others, they also act as a repository of all kinds of information carried by their photos and tags. As geo-tagged photos can be easily created with the help of global position systems (GPS), we contend that the hundreds of millions of these geo-referenced images being acquired by millions of citizen sensors are a valuable source of geographic information. The objective of this dissertation is to perform geographic knowledge discovery using community-contributed geo-referenced photo collections such as available at Flickr. We present a novel knowledge discovery paradigm termed proximate sensing and demonstrate how it can be used to perform land cover and land use classification using widely applied image features in computer vision as well as text associated with the photos.

For land cover classification, a case study is performed using a supervised classification framework. More than one million photos are collected from two on-line photo sharing websites over a 100x100 km study region in the United Kingdom. The study region is further divided into 10,000 sub-regions, where each sub-region is classified into developed or undeveloped regions by analyzing the ground-level photos. The classification results are then compared based on the image features used as well as the source of the photos.

A case study of land use classification is conducted to further validate the concept of proximate sensing. More than 16,000 images are collected from Flickr over two university campuses. The images are classified into academic, sports, and residential facilities; a land use map of each campus is generated according to the classification results. Furthermore, we explore the idea of extracting geographic information semantically by applying state-of-the art object and concept detectors directly to the photo collections. Maps of object distributions are generated according to the detection results of different object detectors. The spatial analysis performed on these object maps suggests that it is possible to extract useful geographic information using these object detectors, and an experiment of land use classification is conducted to validate this finding.

Main Content
Current View