Lessons from a human-in-the-loop machine learning approach for identifying vacant, abandoned, and deteriorated properties in Savannah, Georgia

Aug 4, 2026

Addressing strategies for managing vacant, abandoned, and deteriorated (VAD) properties is essential for sustaining healthy communities, yet reliable identification remains challenging due to limited standardized definitions and data. This work develops VADecide, a human-in-the-loop machine learning model, and applies it to parcel-level data in Savannah, Georgia. Compared with training-only machine learning, VADecide improves prediction accuracy, yields more dependable identification against field surveys and USPS vacancy records, and clarifies systematic differences between human- and machine-generated results for urban planning.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages21

Reading Time53 min

AccessFree to download

Introduction

Challenges of identifying VAD properties

Research objectives and HITLML workflow

Partnering with Savannah stakeholders and data labeling

Model comparison: VADecide vs classic ML and city workflow

Lessons from a human-in-the-loop machine learning approach for identifying vacant, abandoned, and deteriorated properties in Savannah, Georgia Xiaofan Liang 1*, Brian Brainerd2, Tara Hicks3, Clio Andris4 1University of Michigan – Ann Arbor, Ann Arbor, United States 2City of Savannah, Savannah, United States 3City of Savannah, Savannah, United States 4Georgia Institute of Technology, Atlanta, United States * Corresponding Author, 2000 Bonisteel Blvd, Ann Arbor, MI, 48109, [xfliang@umich.edu](xfliang@umich.edu) Abstract Addressing strategies for managing vacant, abandoned, and deteriorated (VAD) properties is important for maintaining healthy communities. Yet, the process of identifying these properties can be difficult. Here, we create a human-in-the-loop machine learning (HITLML) model called VADecide and apply it to a parcel-level case study in Savannah, Georgia. The results show a higher prediction accuracy than was achieved when using a machine learning model without human input in the training. The HITLML approach also reveals differences between machine vs. human-generated results. Our findings contribute to knowledge about the advantages and challenges of HITLML in urban planning. Introduction Over time, neglected structures and land in communities degrade into unusable or unsafe infrastructure. This process adversely affects public health and welfare and can increase crime and lower property values. To protect and support growth in communities, local jurisdictions try to identify these vacant, abandoned, and deteriorated (VAD) properties. However, identifying VAD properties, i.e., recording their location, is not a simple task: there is no national database nor a standardized definition with which to detect VAD properties, and as such, cities typically rely on block-by-block field surveys to count VAD properties (Mallach 2018) . Often, housing planners drive around the neighborhoods to visually identify and record these properties, or they rely on calls from neighbors to the civic hotline to report problems with properties in their communities 1. Some cities have used techniques such as spatial decision support systems (SDSSs) and machine learning (ML) models to find VAD properties (Hill et al. 2003; Appel et al. 2014; Hillenbrand 2016; Reyes et al. 2016) but reports of these efforts emphasize the technology itself rather than the combination of human, technology, and data. Minimizing or ignoring the human aspect of these decision-making statistical models is dangerous because the 1 Brian Brainerd, Discussion with authors, April, 2021. models can provide suggestions that could counter a practitioner’s commonsense or what is known to be helpful for a community. Accordingly, this research has two objectives. The first objective is to design a workflow that integrates human expertise into a machine learning model to efficiently identify VAD properties from a dataset of parcels with parcel attributes. The outcome of this objective is a human-in-the-loop machine learning (HITLML) model that we call VADecide. The second objective is to measure how VADecide improves or does not improve upon (a) a simple, classic machine learning (ML) model using existing datasets (i.e., the “ML model” in this paper) and (b) the city’s current workflow based on their expert knowledge. An improved model will be more efficient (requiring fewer training samples and giving more predictions), more accurate (yielding a higher prediction accuracy rate), more reliable (recalling a similar set of properties that were previously discovered using field surveys and U.S. Post Office (USPS) vacancy records), and more robust (capturing all features of VAD properties) . To meet these research objectives, we partnered with the City of Savannah, Georgia, in collaboration with four housing officials from Savannah’s Housing and Neighborhood Services Department (HNSD) and the Chatham County / City of Savannah Land Bank Authority (LBA) .

FAQs

Why is identifying vacant, abandoned, and deteriorated (VAD) properties difficult?

There is no national database and no standardized definition for detecting VAD properties, so cities often rely on block-by-block field surveys and local reporting rather than consistent centralized data.

What is VADecide in this study?

VADecide is a human-in-the-loop machine learning model that integrates expert labeling from housing officials and outputs a parcel-level classification of 'VAD' or 'not VAD'.

How does VADecide perform compared with a classic machine learning model and the city’s workflow?

VADecide achieves similar prediction accuracy to the classic ML model but is more reliable, with higher agreement when validated using field surveys and USPS vacancy data, and it identifies VAD candidates across more neighborhoods than the city’s workflow.

Aug 4, 2026

Upload by Theodora