MAP: Abandoned Property Detection Using Multifaceted Urban Service Data

Abandoned property detection is a tedious and labor-intensive task for field inspectors of local governments. Recent research employs computer vision-based approaches to identify abandoned properties, using Google Street View [1], mobile sensing images [2], and remote sensing images [3]. However, such approaches rely on large-scale up-to-date image datasets of city properties, which are usually unavailable and very costly to obtain. In this work, we propose a novel approach to abandoned property detection by combining urban service data, such as utility payment records, service requests, and complaints reported by residents, which are available from digital city management systems. Since such data potentially reflect abandoned properties from different perspectives, such as financial status, maintenance, and environment, we design mAP: a multi-view framework with an intra-view to inter-view encoder architecture to fuse multifaceted service data for abandoned property detection. Such service data are inherently spatiotemporally imbalanced due to the diversity of the community. To address these issues, a Bayesian Network is designed to assess each report with the socioeconomic factors of the local community and infer its value as evidence for an abandoned property. We evaluate our solution with extensive experiments using datasets collected from the City of Newark, New Jersey. mAP outperforms state-of-the-art solutions by 1.94 to 13.42 times in F1 score and by 21% in recall rate.