Utilizing geo-referenced imagery for systematic social observation of neighborhood disorder

Highlights

• This study brings conceptual clarity in four ‘generations’ of research conducting SSO based on geo-referenced imagery.

• Cutting-edge technologies provide large, and largely unexplored, opportunities for studying neighborhood disorder.

• We propose steps forward to leverage ‘big primary data’ to enhance insights for policy, practice and research.

Abstract

Research methods in social science take advantage from broader trends such as digitalization and increasing computational power, however, this is an evolving explorative search. The main purpose of this article is to describe the methodological innovations in the collection and processing of geo-referenced imagery for the observation of neighborhood disorder. In this narrative review, attention is paid to advances in both the data sources and the data processing methods used. Neighborhood disorder is traditionally measured by means of survey methods and (systematic) (social) observations, but these methods have specific shortcomings, such as respectively the subjective measurement that does not deliver a valid measure of actual prevalence of disorderly phenomena and the intensive use of resources in terms of time and money. This has repercussions for (the interpretation of) the results based on these data. Today, scholars have innovative data sources and cutting-edge data processing methods at their disposal that can meet (some of) these shortcomings, but which have not yet been fully explored. In this article, the evolutions in the use of geo-referenced imagery for the observation of neighborhood disorder from the last 25 years are described with a focus on the empirical opportunities, and the methodological challenges and prospects. We conclude by outlining the road ahead: promising avenues for future research to exploit the full potential of ‘big primary data’.

Introduction

In this review article, an overview of the developments in the use of geo-referenced imagery for the observation of neighborhood disorder is provided. The main purpose of this study is to outline the methodological advances and to focus on the empirical opportunities, and methodological challenges and prospects. We specifically focus on the developments in the use of systematic social observation (hereafter: SSO). SSO is considered as a combination of qualitative and quantitative methods, i.e. “in-depth observational work championed by ethnographers and the large-scale, systematic, and generalizable approaches of quantitative social science” (Brunton-Smith, 2018, p. 294). Reiss (1971) described SSO as a method where:

… observation and recording are done according to explicit procedures which permit replication and that rules are followed which permit the use of the logic of scientific inference. The means of observation and recording, whether a person or some form of technology, must be independent of that which is observed and the effects of observing and measuring must be measurable. (p. 4).

The occurring developments regarding SSO are considered on two axes: the first one based on the type of data used (primary versus secondary), and the second axis based on the nature of the data processing methods (manual versus automated). Those two axes can be represented as a quadrant, resulting in four possible combinations which can also be situated in time (see Fig. 1). Primary data are data collected by researchers for specific research goals and are used in that context. Secondary data are data that are not collected for specific research goals but are used in the context of research or primary data used secondary by other researchers (Hox & Boeije, 2005). In other words: with secondary data, existing data are used retroactively to answer a specific research question. Manual coding means that the researchers are responsible for the processing of the data, with human effort in the coding process itself. In the case of SSO, this often involves the systematic coding of observed facts in which the researcher is interested (e.g. number of graffiti-tags or -pieces, number of units of illegal waste, number of cigarette ends). Automated coding means that this coding is executed computationally, without human effort in the coding process itself. More specifically, techniques that belong to the field of computer vision are used for this automatic coding.

It is not the intention to provide a systematic overview of the use of geo-referenced imagery. For this, we refer to systematic reviews of previous research in different fields (e.g., Aghaabbasi, Moeinaddini, Shah, & Asadi-Shekari, 2018; Rzotkiewicz, Pearson, Dougherty, Shortridge, & Wilson, 2018; Vandeviver, 2014). This narrative review is centered around a number of seminal studies, that give an impetus to or are illustrative for the observation of neighborhood disorder. The aim is to give a concise description of how SSO using geo-referenced imagery is being applied, from a quantitative point of view. After all, social observation also has qualitative applications, such as spatial video geo-narratives¹ (e.g., Bell, Phoenix, Lovell, & Wheeler, 2015; Curtis, Curtis, Porter, Jefferis, & Shook, 2016), but these will not be discussed further in this article. SSOs are leveraged across many disciplines, for example in public health (e.g., Bader et al., 2015), sociology (e.g., Hwang & Sampson, 2014), and geography (e.g., Conley, Stein, & Davis, 2014). In this article, we focus specifically on the application of SSO in the field of criminology, by applying it to the observation of social and physical neighborhood disorder, as initiated by Sampson and Raudenbush in the mid-90s.

In what follows, we will first focus on disorder and how this is traditionally measured (Section 2). Next, we will chronologically describe the considered developments, as visualized in Fig. 1. In the first phase, researchers used imagery that they collected themselves and then manually coded (Section 3). In the second phase, they started using secondary data that was being coded manually, this under the influence of an increase in the publicly available secondary imagery data (Section 4). In the third phase, researchers started applying automated coding (based on advances in computer sciences, in particular machine learning and more specifically computer vision) to the publicly available secondary imagery data (Section 5). In the last phase, researchers started the application of this automated coding on imagery data that they collected themselves (Section 6). In the concluding section of this article, we reflect on the relatively fast developments of this research domain and the practical implications of this (Section 7). Finally, we will also consider the promising avenues for future research (Section 8).