Authors: Shengyuan Zou*, University At Buffalo, Le Wang, University at Buffalo
Topics: Remote Sensing, Urban and Regional Planning
Keywords: vacant house, remote sensing
Session Type: Paper
Start / End Time: 8:00 AM / 9:40 AM
Room: Studio 10, Marriott, 2nd Floor
Presentation File: No File Uploaded
Vacant houses are essential in house price estimation, environmental security monitoring and population estimation in the urban area. Existing individual-level vacant house data, which are derived from field survey data by Census Bureau or United States Postal Service, are not released to the public, and might not up to date. Remotely sensed imagery provides a time-saving and economical way in exploring vacant houses and potential in detecting vacant houses. However, there is no significant sign or special spectral signal for vacant houses. In this article, we study on the feasibility to detect vacant houses from high spatial resolution(HSR) remotely sensed imagery. The whole proposed method includes three steps: “ground-truth” data collection, classification implementation and feature selection, which give us a clear procedure flow for vacant house detection from HSR images. A new building change detection method was developed to collect “ground truth” vacant house data from multi-temporal images. Then, a supervised classification algorithm was implemented to classify occupied houses and vacant houses from the single scene image. Based on the estimated error from the predicted result, we selected important features in vacant house detection and tested if they are significantly different between occupied houses and vacant houses. With selected features, the classification accuracy reached 89.77% in overall accuracy and 78.6% user’s accuracy for vacant houses, which shows a better result than the result of classification before feature selection. The final results show that the size of vacant house parcel and surrounding vegetation have significant features in vacant house detection.