We developed a systematic yet flexible approach to model the structure of urban deprivation by integrating socioeconomic and land cover characteristics. Current operational definitions of urban slums do not adequately address the dynamic and heterogeneous nature of slum communities. Our modeling approach generates pixel-level information on levels of deprivation, which reduces the spatial uncertainty in aggregate census-level maps and incorporates natural heterogeneity within urban slums. We found that urban slums were not discrete entities with homogeneously shared characteristics of deprivation and marginalization, but were rather part of an urban continuum. These findings highlight the complexity inherent in these communities and challenge efforts to apply rigid or simplistic classification schemes that aim to draw simple correlations between, for instance, changes in land cover characteristics and improvements in human health outcomes .
The results of the canonical correlation analysis confirmed our hypothesis that land cover variables are associated with socioeconomic variables. This association is particularly strong in the first dimension of the canonical correlation analysis, which captures a gradient along both land cover and socioeconomic metrics. The second dimension provides a more nuanced differentiation of slum settlements by identifying communities that lack basic infrastructure, are often newly squatted, and represent heavily marginalized portions of the city. Because these characteristics were not strongly associated to land cover, socioeconomic data weigh most strongly in this dimension. By combining the second dimension with the first dimension, we were able to identify and map the most marginalized areas, which may not have been identified from land cover or socioeconomic variables alone.
Our approach provides a crucially needed improvement in the spatial resolution of deprivation mapping that highlights areas of maximum marginalization. By identifying land cover proxies for urban deprivation, we were able to increase the map resolution from the census tract level to the 30 m pixel level, thus capturing within slum heterogeneity and reducing the spatial uncertainty of census-level maps. This higher spatial resolution provides an essential tool for programs aiming at improving access to services to the most marginalized populations. In particular, access to health care could be significantly improved through targeted immunization campaigns and triage of services and hospitals, improved access to education, better design of refuse collection and transportation, and development of community wide initiatives to increase leisure and exercise space [5, 22–28]. The more specific targeting interventions achieved by disaggregating census tract data to the pixel level would improve upon previous census level mapping approaches, which may not adequately capture internal heterogeneity and may overestimate the size and scope of urban slums [14–16].
Urban slum settlements within Salvador are not discrete entities with homogenous characteristics, but are rather part of an urban continuum of deprivation levels, as illustrated in the canonical plot (‘canonical space’). In geographic space, slum communities are more clearly defined in the eastern parts of the city than in western Salvador, where there is a larger range of deprivation levels. The representation of the census tracts in canonical space provides a snapshot of the urban structure in Salvador in the early 2000s. Urban slums evolve from squatter settlements to de facto, well-established communities as they gradually receive infrastructure [2, 28]. The canonical modeling approach distinguishes slum communities at different stages of their development and evolution, as illustrated by the position of the four Salvador communities in the canonical plot (Figure 5). By examining changes in this structure over time, we can describe community transitions as development or public health policies and interventions are implemented.
In addition to applications to city-wide infrastructure projects, the proposed approach can serve as a tool to guide policies aimed at changing the built and social environments of poor/informal areas of cities. The proposed approach can be used by municipal/national policy makers, urban planners and public health officials to guide interventions and measure the social and physical impacts of policies focused on access to health, transport, housing, social inclusion and environmental/climate change vulnerability. By tracking a community’s progression across canonical space we can quantitatively monitor the impacts of these policy interventions and assess the evolution of urban slums. Furthermore, our approach generates socioeconomic and land cover metrics that can be used as parameters or to validate mechanistic models of urban expansion.
Despite our approach’s flexibility, the constraints of the canonical correlation analysis and the resolution of our data limited the accuracy of the overall model. Unlike other techniques currently used to map urban slums [9–14, 16, 29] our method does not require expensive very high resolution imagery or specialized object-oriented classification software. Accuracy could be improved by using higher resolution data when available, which can easily be incorporated in this approach. High-resolution imagery would likely only be affordable for smaller area studies of neighborhoods of particular interest. While we were able to accurately describe socioeconomic characteristics of urban slums, some potentially important variables such as insecure access to tenure were not included in this model. There were also outlier communities with low infrastructure despite higher levels of income. These areas are likely squatter settlements among higher income census tracts, and overall account for less than 0.014% of the total census tracts considered. The selections of variables from the Brazilian census used in the canonical correlation analysis were also constrained by the linearity requirement.
While this approach was applied specifically to Salvador, Brazil, it is flexible enough to incorporate a wide variety of data sets that may be critical to other regions of interest. Our objective was to develop a low-cost method/model that might be applied in other urban environments where there are data collection challenges/gaps, which is the norm for most urban informal settlements in the global south. Our approach does require census data, which could limit its generalizability to other regions. However, the quality of census data throughout the world is increasing in major cities like Delhi, Accra, and in cities throughout Brazil. The specific associations between the socioeconomic and land cover variables, while specific to Salvador, can with further validation be used in other urban slum environments, particularly in Brazil. While we based our analysis loosely off the UN-HABITAT definition, our approach allows for the incorporation of various socioeconomic or land cover variables that can be modified based on site-specific attributes, such as insecure access to tenure or thatched rooftops instead of corrugated roofs as a proxy for poor quality of housing.