Satellite Imagery

The methods used here to model population distribution in Africa are adapted from previous work undertaken for East Africa [15] (link)–[17] (link). The methods were modified for ease of replication and to facilitate the incorporation of new data. Full details on population distribution modelling methods are presented in Text S1. Tables summarizing country-level input data are available on the AfriPop website: www.afripop.org.
Recent work showed that GlobCover was the global land cover dataset that, combined with detailed settlement extents, produced the most accurate population distribution data in an African context [17] (link). The GlobCover dataset was modified to accommodate the more detailed settlement extents obtained from satellite imagery and geolocated points. The GlobCover dataset was first resampled to 100 m spatial resolution, and the urban class – which typically overestimates settlement extent size [15] (link), [17] (link) – was removed and the surrounding classes expanded equally to fill the remaining space. The more detailed settlement extents were then overlaid onto the ‘urban class deprived’ land cover map and land covers beneath were replaced to produce a refined land cover map focussed on detailed and precise mapping of human settlements.
Human population census data, official population size estimates and corresponding administrative unit boundaries at the highest level available from the most recent available censuses were acquired for each African country. High resolution census data were available for three countries in Africa: Ghana, Swaziland and Kenya. Kenya data were also available at enumeration area level (finer than level 5) for 58 of the 69 Kenyan districts. Also obtained was a population density map of Namibia at 1 km spatial resolution (for details on the Namibia density map, see description on www.afripop.org). A table summarizing the spatial resolution, year and source of all data used is available on www.afripop.org.
The modelling method distinguishes urban and rural populations in the redistribution of populations. Major settlements have population numbers already derived and validated and this makes up 38% of the total African population. The remaining 62% rural population was redistributed using land cover-based weightings. The refined land cover data and fine resolution population data from Ghana, Kenya, Namibia and Swaziland were used to define per land cover class population densities (i.e. the average number of people per 100×100 m pixel), following approaches previously outlined [15] (link), [17] (link). These land cover specific population densities were then used as weights to redistribute the rural populations within administrative units in the remaining African countries. The population sizes at the national level for each dataset were projected forward to 2010 with rural and urban growth rates estimated by the UN Population Division [18] . The GRUMP urban extents (available online at: http://sedac.ciesin.columbia.edu/gpw) were used to distinguish between urban and rural areas.

To understand how the lives of Mosul residents have changed in the year following the liberation from ISIS control, we conducted this household survey in September 2018. We started with the same 40 clusters identified for the 2016/7 Mosul household survey [15 ]. For that survey, the residential administrative units or neighborhoods of Mosul were listed. These residential administrative units would commonly contain around 400 households. From this listing, 40 neighborhoods were selected as clusters, 15 in west Mosul and 25 in east Mosul, a ratio approximating the pre-ISIS population distribution of Mosul. In each cluster, a random point was selected using satellite imagery. The survey began with the household nearest to the random start point and moved to the nearest subsequent households until 30 households had been interviewed. A household was defined as a group of people living together, eating from a common kitchen and living in a structure with a separate entrance from the street. The sampling process is described more fully elsewhere [3 ]. For the one-year follow up survey reported here, 30 households were selected from each of 12 clusters randomly chosen from the original 25 in east Mosul, and from eight clusters of the previous 15 west Mosul clusters. A total of 600 households from 20 clusters were interviewed in September 2018. No households included in the 2016/17 survey were sampled again in this 2018 follow-on survey. Data from completed forms were computer entered in SPSS in Baghdad and preliminary analysis conducted there. Further analysis using Stata1 was conducted in Baltimore.
The questionnaire was adapted from the 2016/17 survey. Interviewers began with a household listing of persons in the household on or after July 2017. Demographic and educational characteristics were collected including births, deaths, injuries and marriages. Household and economic characteristics were recorded, as well as employment and economic indicators.
The interview team was composed of four persons with doctoral degrees in community medicine and who had extensive experience in community surveys. All were originally from Mosul. Some interviewers had participated in the 2016/17 survey. Two days of training was provided.

Population data for Burundi, Kenya and Rwanda were adjusted forward [10] (link) to estimated 2002 levels using inter-censal growth rates to match the most recent census data used and the majority of the satellite imagery. Three approaches to the creation of gridded population distribution maps were tested. Firstly (EApop1), the census data was simply areal-weighted [10] (link) to a 100 m spatial resolution grid. Secondly (EApop2), the satellite derived settlement map for each country was degraded to the same 100 m grid, and census counts within an administrative unit were then allocated to the grid squares classified as settlement. Administrative units not containing any grid squares that were classified as settlements had their population counts simply areal weighted. Finally (EApop3), the satellite imagery derived settlement maps were degraded to 100 m spatial resolution and ‘burned’ into the Africover land cover layers to create a refined land cover map for the region. Where settlement extent was mapped as smaller in the settlement map than the ‘urban area’ or ‘rural settlement’ classes in Africover, the surrounding land covers were grown to infill the gaps. This refined land cover layer and Kenyan enumeration area census data were then used to define per land cover class population densities (Table S1), were then used as weights to distribute the census data across the entire region to create a population map.
The accuracies of the various population mapping procedures were tested principally using the enumeration area level census data for 50 Kenyan districts (figure 2). Additionally, to provide a finer resolution measure of settlement population mapping accuracy, the Africover settlement extents with assigned populations were used. For each 100 m gridded population distribution map produced, the population numbers falling within each enumeration area and settlement polygon were extracted and compared against the actual population figures, with overall and district-specific root mean square errors (RMSEs) calculated. To explore the effectiveness of the population mapping procedures in the absence of high resolution census data, for Kenya the map production process was repeated using census data at administrative levels 4, 3, 2, 1 and 0 (national). Finally, the maps of Kenya from existing gridded population products (APD, GRUMP, GPW3 and Landscan 2005) were adjusted to 2002 [10] (link), areal weighted to a 100 m grid and compared to the enumeration area census data to obtain estimates of their accuracy relative to the approaches outlined in this paper.

Defining urban agglomerations is challenging and often depends on considerations and parameters (18 (link)–20 (no link found, link), 50 ). Africapolis applies the same definition for an urban agglomeration at a continental level (i.e., an agglomeration of at least 10,000 inhabitants and buildings less than 200 m apart), enabling it to analyze and compare cities between different countries (31 ). Buildings’ locations were extracted from the recently launched Google Open Buildings dataset, https://sites.research.google/open-buildings/. The buildings’ footprints were obtained using a deep learning model with high-resolution satellite imagery (50 cm pixel size) (27 ). We extract the buildings’ footprints located within Africapolis polygons and keep the attributes of the building’s center location (latitude and longitude), the confidence score, and the building footprint area. We assign each building the unique identifier of the Africapolis urban agglomeration to which it belongs. This way, we obtained the location and attributes of the buildings’ footprint of 6,849 African urban agglomerations (Fig. 4). We computed street network metrics (29 , 51 ) and terrain metrics (30 ), such as the difference in elevation between the highest and lowest point, the average slope, and the average height within each Africapolis urban agglomeration polygon (SI Appendix, Appendix A).