Deep Learning Visual Methods for Identifying Abandoned Houses

Jim DeLisle
University of Missouri–Kansas City

Hye-Sung Han
University of Kansas

Duy H. Ho
Yugyung Lee
Brent Never
Ye Wang
University of Missouri–Kansas City

The opinions expressed in this article are those of the authors; they do not necessarily reflect the views and policies of HUD or the U.S. Government.

Housing abandonment contributes to neighborhood decline and disinvestment. Abandonment has plagued large metropolitan areas for decades, yet quantifying the scope and impact of abandonment has proven costly and elusive. This study introduces an innovative approach to detect and measure abandoned houses using technology innovations without requiring significant resource commitments. It presents a system of detecting abandoned houses leveraging deep learning models for image classification, building an ensemble model that considers both global and local contexts to identify abandoned structures. This study takes imagery and structure data from multiple sources and uses transfer learning in a three-stage ensemble approach to identify abandoned houses. Four deep learning models are constructed for this study: the ResNet-50 model, an incremental knowledge model, a hybrid approach, and a check model. Results from the different models are compared and analyzed to identify the visual characteristics of houses that improve or degrade each method’s accuracy. The methodology presented herein is scalable and could be applied in other neighborhoods and communities. The data generated by this method will empower communities and cities to design more effective strategies to address housing abandonment.