Keywords

1 Introduction

Global sustainable development, adoption of the United Nations (UN) Sustainable Development Goals (SDGs) (UN-Habitat 2015), and international development agreements have triggered a much-needed data revolution, with countries and institutions around the world recognising the critical role of geospatial data for evidence-based policymaking. Increasingly, high-quality geospatial datasets are becoming an essential source of information to guide social, economic, and environmental policies at global, regional, national, and subnational scales (Nilsson et al. 2016).

Among the wide variety of geospatial datasets needed to create credible measures of sustainable development (e.g. data on land use, land cover, risks, or climate indicators), some of the most spatially necessary datasets are those describing the spatial distribution of the urban form (Benza et al. 2016) and the human population. Global population datasets must ensure to include all populations, with particular attention to urban slums, which exhibit distinct patterns from the rest of the city and are often not accounted for in census data in most low- and middle-income countries (LMICs). While open population databases are widely used, including censuses, surveys, and gridded datasets (e.g. WorldPop), they systematically and dramatically underrepresent slum populations in LMICs (Thomson et al. 2020). Global population models come with high uncertainties in LMIC cities, e.g. the underestimation of the urban population living in slums (Kuffer et al. 2022).

Slums are characterised by sub-standard services, lack of open space, and high building density. Deprivation is exacerbated by high urban built-up density and overcrowding of households that are much higher than in urban formal areas. In addition, slum dwellers commonly lack land tenure security, so they are often not recognised by local authorities and are not included in official population counts (i.e. in national censuses). This means that population data collected through traditional methods may not fully reflect the number of people living in slums. This problem is compounded by the fact that slum populations are often highly mobile (people move in and out of neighbourhoods frequently) and with diverse population patterns according to social strata, which can result in multiple families residing within the same household. The presence of such diversity in slum populations makes it challenging to estimate their numbers accurately based on limited observations or extrapolation methods.

It should be noted, as shown in Fig. 14.1, that not only is there a morphological distinction between slums and formal areas, but there is variation in size and physiognomy between and within slums (Georganos et al. 2021). Therefore, the extrapolation of slum data needs to be carefully considered. To address this problem, more targeted and innovative approaches to data collection in LMIC, especially in slums, are needed. This may involve the use of new technologies, such as satellite imagery and geospatial data, to identify slum populations. It may also require closer collaboration between local authorities, researchers, and community organisations to ensure that slum populations are adequately represented in population data.

Fig. 14.1
16 enlarged satellite images of slums at scales of 100 meters. Mathare North, Pumwai, Mukuru, Korogocho, Mathare Central, Kariobangi, Waruku, and Kibera each have 2 images.

Physiognomic diversity between and within slums in Nairobi city

Full size image

Choosing the appropriate spatial unit for modelling urban slum population depends on several factors, including the scale and geographic extent of the study, the level of detail required for the analysis, and the available input data and the protection of privacy. To accurately model the population of slums within global population layers, the spatial unit chosen needs to match the intra-slum level patterns of the population density. This will enable a more accurate representation of the population distribution and help to address any potential biases in the mapping process. Therefore, administrative units may not be appropriate as they often contain very diverse areas and potentially include one or several slums. The choice of the spatial unit should also be driven by the availability and quality of data. For example, census tracts may be ideal for modelling urban populations at larger scales, in case of high-quality census data availability. However, in the case of LMIC cities, census data are not reliable data sources for population estimation in slums (Carr-Hill 2013) because they are often outdated regarding the high pace of their population growth, and their smallest administrative unit is too small. Therefore, considering the fine level of detail required in this study, fine grid units would be optimal (Fig. 14.2). It should be noted that the selection of appropriate spatial units involves adapting to the diversity within each grid to accurately represent the population distribution at a finer resolution.

Fig. 14.2
A satellite image of a slum with borders marked for various spatial units. Block level and district levels are in the order of increasing administrative divisions, while plot level and building level are in the order of increasing physiognomy from the central grid level marked.

Spatial units diagram: from administrative to more physiognomic

Full size image

The process of gridded population mapping requires understanding the integration approaches between diverse geo-datasets, from source units into the target units (Leyk et al. 2019). Previous research in different fields has shown that spatial aggregation (in top-down approaches) influences urban modelling results (Zhang and Kukadia 2005; Duque et al. 2018; Weigand et al. 2019). This common challenge is called the modifiable areal unit problem (MAUP) (Gehlke and Biehi 1934; Openshaw 1984; Wong 2009) and describes how results change as the spatial aggregation of data changes. A good approach to evaluating the influence of the aggregation level is to conduct a sensitivity analysis, testing different spatial units to see which one provides the best results and insights for the research question at hand.

As shown in Fig. 14.3, several global open-gridded population datasets are available, and the modelling process involved in their creation is variable (for further information, visit the POPGRID website (2020)). The first distinction is whether they are top-down or bottom-up approaches. Top-down approaches to distributing spatial population patterns have been shown to grossly underestimate the slum population, as they are based on the disaggregation of census data into smaller units (Thomson et al. 2020). On the other hand, bottom-up approaches are typically more accurate, employing commonly in situ survey data with a set of covariates in statistical and machine learning approaches (Boo et al. 2020). A distinction is also made between approaches that do and do not constrain population redistribution using settlement extents. A constrained approach produces improved population distributions (Linard et al. 2011). Finally, the degree of modelling, from un-modelled to highly modelled, indicates the size of the covariate set used to predict the population.

Fig. 14.3
A diagram for the global population. Constrained and unconstrained datasets, spatial data, census, survey data, grid size, and unmodeled, light and highly modeled are indicated. G P W, G H S L, H R S L, LAND SCAN, and W P G are under top-down census. World pop and G R i D 3 are under micro-survey.

Global population datasets diagram

Full size image

When creating spatial population data, it is necessary to disaggregate or aggregate data into a common spatial unit. Disaggregation is typically employed when the source data is census or administrative data, and the target grid cell is a smaller spatial unit than the source units. It has been shown that the incorporation of ancillary urban data, named constrained, results in better accuracies than unconstrained modelling when disaggregating from census data. As shown in Table 14.1, there are several disaggregation methods to allocate population distribution into smaller units. One common method is the areal weighting method, an unconstrained approach, which assumes that the population is uniformly redistributed from the source units to the target cells that overlap with the source units. Although this assumption is an oversimplification due to non-uniform population distributions, the method is computationally efficient and can create spatially explicit and globally consistent population estimates. An example of this approach is the gridded population of the world (GPW) product. When using ancillary data, the population redistribution is achieved through areal interpolation, referred to as the dasymetric mapping method. All dasymetric mapping techniques rely on the relationships between population (provided by the input census data) and ancillary data (such as land cover) that can be utilised to reallocate the population to finer spatial units with greater accuracy. The disaggregation method applied in traditional dasymetric approaches varies, ranging from binary dasymetric refinement to more complex weighting schemes or hybrid methods. As Leyk et al. (2019) stressed, they differ in the way relationships between population and ancillary data are derived (e.g. presence/absence based, empirically derived, or optimised) to determine weights for different locations to guide the disaggregation of population totals.

Table 14.1 Disaggregation methods
Full size table

Efforts have been made to develop models that are not dependent on census data and rely on bottom-up approaches. However, these attempts have not been applied on a global scale. For example, the GRID3 project has developed population estimates for LMICs that are independent of census data (Grid3 2023). Recently, a method combining household survey data with building footprints has been suggested as a possible census-independent approach (Thomson et al. 2020; Boo et al. 2022), the most promising method for creating benchmark models for slum populations. It is noted that attempts to develop bottom-up population estimations use a random sampling scheme of survey locations, and while this methodology suggests that sample surveys can provide unbiased estimates of the total population size of an aggregated population (Cottam et al. 1957), the precision of these estimates will depend on the level of aggregation, the distribution of population within the aggregate units, and the extent to which the sample is representative of the different urban physiognomies. As stated by Georganos et al. (2021), the latter differ vastly within and between slums.

In this study, we aim to propose a novel method to produce more accurate gridded population estimates that can be used as a benchmark to evaluate other open population datasets and validate and/or highlight existing uncertainties. Data was collected in situ in collaboration with slum dwellers through a novel data collection, as all the grid size was surveyed, using citizen science, i.e. slum dwellers. The authors hypothesise that leveraging bottom-up earth observation (EO) methods can improve slum population estimates compared to existing censuses and open-gridded population datasets.

2 Data

2.1 Onekana Population Database

The study covers six slum areas in Nairobi city, namely Kibera, Mukuru, Waruku, Pumwani, Korogocho, and Mathare. The study area was divided into 2000 uniformly sized grid cells, each measuring 1 hectare (100 m × 100 m). Out of these, 117 cells were randomly selected and comprehensively surveyed, resulting in a total of 10,550 surveys conducted (Fig.14.4). Throughout the data collection process, certain grid cells had to be excluded from the survey due to safety concerns reported by citizens. These cells were predominantly situated close to rivers, areas with known drug or alcohol consumption, or fenced-off areas with restricted access.

Fig. 14.4
A satellite image of Nairobi with the boundaries of census sub-locations marked. Mathare, Korogocho, Pumwani, Mukuru, and Kibera are divided into grids with few shaded survey cells. An inset enlarged view of western Mukuru has survey cells marked.

Nairobi case study. Census sub-locations, total grids corresponding to slums, and surveyed grid cells are depicted

Full size image

2.2 Open Geo-Datasets

Interpretable urban variables in the form of open geo-data were collected from two sources, namely satellite imagery and cooperative citizen databases such as OpenStreetMap. Variables derived from satellite imagery were produced in prior research studies utilising very high-resolution WorldView-3 images (0.30 cm) of 2019 (Abascal et al. 2022; Georganos et al. 2021; Wang et al. 2023). They consist of proportions of land cover classes and morphological metrics aggregated at the grid cell level. Additional morphological metrics were computed from the Google Open Building layer, which uses satellite imagery of 0.50 cm resolution (Sirko et al. 2021). Finally, road and river data were collected from OpenStreetMap (OSM) contributed by community members in 2021.

2.3 Open Population Datasets

Table 14.2 presents the population datasets selected for comparison, which were chosen based on two criteria: spatial resolution finer than or equal to 100 m by 100 m and data availability for Nairobi city.

Table 14.2 Population geospatial datasets for Nairobi with spatial resolution finer than or equal to 100 m × 100 m used in this research
Full size table

3 Methodology

Figure 14.5 illustrates the general workflow of the study. The first step involves designing and collecting field data, which are then subjected to cleaning and imputation to estimate the population of non-responding households in the grid. Subsequently, the data obtained from surveying 117 grids are used to train and validate the model using advanced spatial aggregation techniques. This results in the creation of the Onekana population dataset, which provides predicted population data for the remaining slum grids. Finally, this dataset is compared with more recent and disaggregated open population datasets described in Table 14.1.

Fig. 14.5
A block diagram of the workflow. Citizen science field data is highly modeled with R S data and open Geo-data into a bottom-up SLUMAP POP dataset. Census disaggregates, and top-down and bottom-up Open POP datasets aggregate with SLUMS uncertainties comparison indicated.

Overall workflow

Full size image

3.1 ONEKANA Population Database

Field Data Collection

To model population, we assume households living in similar physical housing and environmental conditions have similar demographic characteristics, and therefore similar population numbers (Pearce et al. 2010). To gather population data on the ground, slum dwellers participated in a citizen science process. The survey and its objectives were explained to them, and they were involved in designing the work plan and determining their remuneration. They also helped to convey the purpose of the surveys to the respondents to build trust and achieve a high response rate. However, reaching some households was not possible due to the unstable employment situation and unpredictable work schedules of some residents.

The data collection process involves two main steps: (A) map exploration and plot delineation and (B) a field survey conducted using a mobile app. Due to the complicated physical layout of the slums, it was not feasible to survey each dwelling individually as originally planned. Instead, maps were created from very-high-resolution (VHR) satellite image chips and OSM data to identify landmarks and serve as orientation guides within the slums. Surveyors then delineated plots, which are fenced compounds of grouped dwellings with a single entrance. The field survey also consists of two phases: First, counting the number of households in a plot and second, interviewing the inhabitants. Geo-data surveys were conducted using a mobile app created on the Firebase platform for this research (Fig. 14.6).

Fig. 14.6
A process flow for field data collection. In step a, a person divides a satellite map into grids numbered 1 to 18. In step b, A location is indicated in a mobile app, while a household survey form has 3 questions for plot number, h h per plot, and how many people per h h with a submit data button.

Field data collection process. (a) Locating the assigned grid cell and drawing plot boundaries. (b) Carrying out household surveys and filling out the form on the mobile app (grid id, plot id, household number, and population number)

Full size image

Aggregation Methods

To evaluate the sensitivity of slum population patterns and assess the aggregation error, the missing data were filled in using two different aggregation methods. It should be noted that the total number of dwellings per plot and per grid cell was known, as they were counted in the field. The two methods used were (1) grid aggregation method: Population numbers were aggregated at the grid cell level, and the missing population number for a household was assumed to be equal to the mean household population number in the grid cell (2) Plot to grid aggregation method: Population numbers were aggregated at the plot level, and the missing population number for a household was assumed to be equal to the mean household population number in the plot. A first R-squared (R2) test measurement was performed to evaluate the proportion of variance and strength of the relationship between variables.

Population Modelling

Slum population density is predicted for the six biggest slums within the city of Nairobi. As a modelling approach, we employed the random forest (RF) algorithm, where population numbers were predicted from geospatial covariates, such as land cover and building footprints and morphological metrics, for each slum separately, at the grid level. Additionally, to create a more predictive and generalisable slum population model, we used a spatial variant of RF, geographical random forest (Georganos et al. 2019) in which data from all slums were used simultaneously. We evaluate the results both by visual inspection and by validating our predictions against a set of surveys left out of the training stage by computing performance indicators such as root mean square error (RMSE) and mean absolute error (MAE).

Comparison of Slum Population Estimates

To assess the cell-level accuracy of the four existing gridded population datasets considered (Table 14.2), each cell-level estimate (e.g. WorldPop) was compared to the “ground true” population count obtained from the exhaustive household surveys. The reference grid used in the comparison was the Onekana grid, 100 by 100 m. The HRSL dataset is aggregated using the areal weighting method, as we know the population distribution within each grid cell could be diverse (see Fig. 14.1). The areal-weighted interpolation is a more sophisticated method compared to simple spatial aggregation as it considers the spatial distribution of the population within each grid cell. The Census has been disaggregated into the reference grid by the dasymetric mapping method. This method considers the spatial distribution of the census data and uses ancillary data, such as land use or land cover, to redistribute the census data within each grid cell based on the spatial characteristics of the ancillary data. The other 100 by 100 m open population dataset (i.e. WorldPop and GRID3), as there was a spatial mismatch between the cells being compared, were converted to the reference grid using the inverse distance weighting spatial interpolation method. This method estimates the values of unsampled locations based on the values of nearby sampled locations. This method ensured all compared cells were on the same grid and minimised errors due to spatial mismatches between data sets.

4 Results

4.1 ONEKANA Population Database

This study employed two distinct aggregation methods to evaluate the impact on modelling outcomes: the grid aggregation method and the plot-to-grid aggregation method. Figure 14.7 shows the predicted versus observed population for four slums, when modelling the population for each slum separately, revealing that the plot-to-grid aggregation method yielded a remarkable outcome, with an R2 value approaching 0.90 for the major slums except for Kibera. However, the results were less satisfactory (R2 ranging from 0.20 to 0.25) when all the slums were modelled together. The finding that the model demonstrated high accuracy when modelling each slum individually, but not when modelling all slums together, provides support for the hypothesis that population patterns vary across slums.

Fig. 14.7
8 scatterplots of predictions versus observations. Left. The grid aggregation method has increasing trends for Korogocha and Pumwani with R = 0.45 and 0.18, and decreasing trends for Mukuru and Kibera with R = 0.05 and 0.13. Right. The plot-to-grid aggregation has increasing trends for all 4 slums.

Population modelling per slum (with a distinct RF model for each slum), with two different aggregation methods

Full size image

After the aggregation method, the final modelling was determined through a rigorous testing process, and the model was subsequently refined by incorporating a geographical component. The utilisation of the geographical random forest (GRF) method produced the best model and yielded enhanced results, as evidenced in Table 14.3.

Table 14.3 Performance indicators for the ONEKANA population estimate
Full size table

4.2 Slum Population Comparison

Compared to existing population estimates, our results show a mean estimate per grid of almost 600 people, whereas others are less than half of that. As shown in Table 14.4, among the existing open population layers, the best estimate is GRID3 and the worst are both Census and the WorldPop database. This could be explained as when adding survey data (e.g. GRID3) or disaggregating the spatial area and model with exhaustive Remote Sensing data (e.g. HRSL) results improve quantitatively. Despite the improvement, it is clear, e.g. in Fig. 14.8, how the existing databases respond to a spatial distribution of the population such as the census data, as it has been the main data to train the models.

Table 14.4 Population estimate comparison in Nairobi slums
Full size table
Fig. 14.8
5 satellite images of Mukuru slum with the spatial variation of population indicated by gradient shades for counts 100 to 1500. A sub-location is marked with an enlarged view on the top. ONEKANA has the highest estimate followed by census, H R S L, work pop, and GRIS 3

Comparison of population estimates in Mukuru Slum, with ONEKANA (population counts from HH surveys) on top

Full size image

ONEKANA population account for more than one million people, whereas the census estimates less than half a million. These results highlight the need to produce reliable population datasets to have a more reliable and complete understanding of the urban population, accounting for the slum population and “Leave No One Behind”.

5 Conclusion

To reduce urban poverty, including upgrading and planning slum areas and providing slum dwellers with services, it is necessary to improve statistics on the urban population. For this purpose, the extent, nature, and location of slums are needed. Making the population of urban slums visible will help design suitable urban policies, such as the provision of local services. According to our findings, the use of satellite imagery combined with in situ data collected by citizen science allows us to create generalisable, gridded estimates of slum populations. The R2 of the overall model (i.e. when modelling all slums together) is 0.48, although it is promising that when isolating the modelling of individual slums, in some cases the R2 reaches 0.9. This indicates that there is a relationship between the urban characteristics of slums and population distribution patterns. Our work provides insights as to how urban population models should tackle slum areas, as there is currently a lack of ad hoc approaches. The knowledge gained will contribute to a better understanding of the evolution of sub-Saharan African cities, enhancing evidence-based policymaking and ensuring sustainable urban growth.