Authors: Cheng Fu*, University of Zurich, Xiaopeng Song, University of Maryland - College Park, Kathleen Stewart, University of Maryland - College Park
Topics: Geographic Information Science and Systems, Remote Sensing
Keywords: Twitter; Social sensing; Machine learning; Land Use; Activity patterns
Session Type: Paper
Start / End Time: 5:00 PM / 6:40 PM
Room: Congressional B, Omni, West
Presentation File: No File Uploaded
Land use structure is a key component for understanding the complexity of urban systems as it provides insights into how people use space, as well as a snapshot of urban dynamics. This paper integrates socially-sensed activity data with remotely sensed data to infer urban land use change using machine learning. We conducted a case study in the Washington D.C.-Baltimore metropolitan area to identify residential and non-residential lands and to map their expansion between 1986 and 2008. The proposed approach employed a satellite-based impervious surface cover product to derive physical signatures of land use, as well as georeferenced tweets to derive activity signatures of land use. A Random Forests classification model was employed to differentiate residential and non-residential lands based on the physical and activity signatures. Model assessment showed that the proposed classification workflow could classify residential and non-residential land uses at an accuracy of 81%. Using the temporal information from historical satellite data, the study also reconstructed the temporal trajectory of development for different land use types. Results indicate that the proposed approach is capable of mapping detailed land use and change in an urban region, which represents a new and viable way forward for land use surveying over megacities and other large regions.