REU Supplement for CAREER: Learning and Using Models of Geo-Themporal Apprearance

Billions of geotagged and time-stamped images are publicly available via the Internet, providing a rich record of the appearance of materials, objects, and scenes across the globe. These images are a largely untapped resource that could be used to improve our understanding of the world and how it changes over time. The foundation of the proposal is research to extract useful information from this imagery and fuse it into high-resolution global models that capture geo-temporal trends. We show how these models can be applied to: improve performance on traditional computer vision tasks and address previously untenable tasks; enable interactive exploration of semantic properties of the world by novice and expert users; and make geotagged imagery a more usable and navigable resource for scientific research and education. To enable this, new methods are needed to extract, organize, utilize, and visualize knowledge from geotagged images. This proposal presents a unified research, education, outreach, and collaboration plan that fills many important gaps in this area. The work proposed through this REU funding request will focus on applying the concepts proposed in the CAREER proposal to specific problems. The project will use machine learning models to find relationships between the appearance of location, both from ground-level imagery and satellite imagery, and some geo-spatial semantic or quantitative properties of the scene. Funds for