Background & Summary

The United Nations (UN) projects that the global human population will grow by 2 billion between 2019 and 20501. Specifically, Latin America and the Caribbean has a total population of approximately 658 million, and is expected to increase by approximately 90 million by 20501.

The region has made important strides against infant and maternal mortality, communicable disease transmission, and incidence of noncommunicable disease in the last 10 years2, largely due to economic development, and the improved capacity and flexibility of healthcare systems3,4. However, the challenge to overcome inequalities of health outcomes derived from the intersection of determinants including socio-economic status, gender, and ethnicity at subnational levels is identified as a key step to universal health access, a key target of the UN Sustainable Development Goals (SDGs)1,2. Moreover, geographic access is a principal determinant of healthcare access, and is crucial to identifying inequities in subnational health status and access to healthcare5,6.

According to the UN Office for the Coordination of Humanitarian Affairs7, Latin America and the Caribbean is the second most disaster-prone region in the world, with 152 million people impacted by 1,205 disasters between 2000 and 2019. Hydrometeorological phenomena including flooding, storm surges, and hurricanes are the most common and destructive hazards in the region7, comprising 60% of all reported disasters during 2010–2016, at an estimated cost of US$278 million dollars2. Climate change operates as a ‘risk magnifier’, increasing the volatility and frequency of hazard events, which disproportionately impacts the populations of low- and lower middle-income countries8,9. Small-island territories and major coastal settlements are particularly threatened by sea-level rise8, with an estimated 30 million people living in low-lying areas (i.e. within the first 10 m of elevation) in the region10. Moreover, the region is exposed to significant seismic and volcanic activity11, due to its location along the ‘Ring of Fire’, a belt following the edge of the Pacific Ocean encountering 80% of the world’s volcanic and seismic events12. Between 2000 and 2019, 75 earthquakes occurred in the region, resulting in 226,000 deaths at a total damages cost of US$54 billion7.

Consequently, efforts towards a fuller and clearer understanding of the spatial distribution of population is crucial to a whole swathe of developmental goals. Amongst natural and man-made disaster scenarios there is a demand for high-resolution population estimates to support the accurate assessment of the scale of an event and the required relief13,14,15,16. Since such hazard events are highly unlikely to impact areas conforming to administrative units, detailed WorldPop gridded data is already regularly used to more precisely assess the size and characteristics of potentially affected population, typically age and sex17,18. Moreover, accurate population estimations are fundamental to nearly all public health intervention and planning efforts19,20. Regularly updated estimates facilitate an enhanced understanding of population size and distribution, improving the efficiency and effectiveness of targeted vaccination planning and delivery programmes21.

Therefore, significant work has been undertaken since the early 1990s to develop high-resolution gridded population datasets at global or continental scales22. Advancements in the spatial resolution and availability of geospatial data, statistical analysis approaches, and processing power have enabled the generation of more accurate datasets that describe changes in human population scale, composition, and distribution over time23. These advancements have facilitated the development of a wide range of openly available, large-scale gridded population datasets22,24,25,26,27,28,29,30,31,32. However, these datasets have been of limited value to governments due to the lack of alignment with their own population figures. Therefore, seeking to overcome this limitation and encourage the uptake of gridded population data, this project represents the first endeavour to use official population figures and boundaries to create gridded population data across an entire continental region.

WorldPop is an interdisciplinary applied research program that develops peer-reviewed research and methods for the construction of open and high-resolution geospatial data on population distribution, demographics, and dynamics. Within this framework, an open-access repository of high-resolution gridded population datasets for the Latin America and the Caribbean region has been generated, using official, finest-available population census-based figures and projections (Table 1) and national boundaries provided by National Statistic Offices (NSOs) from the region, alongside a suite of ancillary geospatial datasets relating to human population, including high-resolution settlement data. Using a Random Forest (RF) dasymetric modelling approach33, population count data and ancillary geospatial datasets for 40 countries (Tables 1, 2) were gathered, prepared, and processed to create gridded population datasets with a spatial resolution of 3 arc-seconds (approximately 100 m at the equator).

Table 1 Summary information of population count data and administrative boundary datasets used to produce the gridded population datasets.
Full size table
Table 2 Summary information of the default datasets and the derived default covariates used for RF model fitting and prediction.
Full size table

Methods

The methodology used to construct this data product, similarly to previous WorldPop products for the region28, implements a top-down approach to population disaggregation via a RF dasymetric modelling approach33. However, there are two marked differences underlying the data product presented herein: i) the use of official, finest-available census-based population figures and projections (Table 1) and administrative boundaries, and ii) the addition of high-resolution World Settlement Footprint 3D (WSF3D) data to the suite of RF-fitting covariates.

Random forest-based dasymetric population mapping approach

A RF algorithm was implemented to generate a gridded population density weighting layer at 3 arc-second resolution (approximately 100 m resolution at the equator); this prediction layer is then used to perform dasymetric disaggregation of population counts from administrative units into target grid cells at country level33. RF is a predictive, non-linear, and non-parametric ensemble learning approach that generates a large set of decision tree models and aggregates their predictions34. Decision trees are independently generated by bagging (i.e., by sampling the entire dataset with replacement)35, typically two thirds of samples are used to train the trees (known as the ‘bagged’ sample). Each node of each decision tree is split according to an iterative method in which, at each node, the optimal splitting method is used34. After all regression trees have been constructed, the outputs of all tree predictions are aggregated by calculating either their mode or average, contingent on whether the trees are utilised for classification or regression, to produce a final classification decision36. The remaining third of unsampled data, known as ‘out-of-bag’ (OOB), are used to perform the internal cross-validation technique to accurately estimate the prediction error of the RF model34; this is achieved by averaging all mean squared errors calculated using the OOB data. The RF approach is robust to overfitting34, and its predictive accuracy is not very sensitive to the three parameters to be specified for model fitting36, explicitly, (i) the number of observations in the terminal nodes of each tree, (ii) the number of trees in the forest, and (iii) the number of covariates to be randomly selected at each node.

The RF-based dasymetric population mapping approach developed by Stevens et al.33, has been used in this framework to produce gridded population distribution datasets for Latin American and Caribbean countries. This approach consists of using a RF algorithm to generate gridded population density estimates that are subsequently used, as a weighting layer, to dasymetrically disaggregate population counts from administrative units into grid cells37.

RF model fitting was undertaken by generating 500 trees, and assigning the number of observations in the terminal nodes equal to one. Following RF model fitting, population density was predicted using a reduced selection of covariates. For each target grid cell, the average of all decision tree predictions was designated to the cell as the estimated population density value. Where there were insufficient observations (i.e. insufficient administrative unit population counts) to fit a RF model for a given country, an additional country with similar characteristics was selected, and utilised to fit an appropriate RF model for predicting population density at the grid cell level38. Subsequently, in both scenarios, dasymetric disaggregation of the administrative unit-based population counts was undertaken using the population density weighting layer37, thereby generating two gridded population datasets of estimated number of people per grid cell.

All tasks described above were performed using the popRF package in R39. The popRF package functionalises the RF-informed dasymetric population modelling procedure33 within a single programming language framework, and is publicly available, open source, and environment agnostic39. This package has been parallelised where possible to achieve efficient prediction and geoprocessing over large extents, supporting functions that have applied utility beyond simply performing disaggregative population modelling39.

Data collection

For every country listed in Table 1, population counts were paired with their corresponding administrative unit boundaries within a GIS interface. Official and best available subnational population census-based figures and projections, and at the finest administrative unit level possible, alongside matching official administrative unit boundaries were provided by NSOs of the region with support from the UNFPA and OCHA. These input data are technically assessed by the UNFPA and subject-matter country experts, and are adopted as common baseline population data for use in disaster preparedness and operational humanitarian response. Further summary information regarding the input population data, including base-census year, and corresponding administrative unit boundaries is provided in Table 1.

Human population density is known to be highly influenced and correlated with a variety of environmental and physical factors, each of which can be credibly associated with and influence the spatial distribution of population23,30,40. These factors are classified into two distinct categories; firstly, continuous variables such as topographic elevation and slope41,42, climate43, and intensity of night-time lights44,45. Secondly, categorical variables notably including land cover type46,47 and the presence or absence of settlements and urban areas48, roads48, waterbodies and waterways49, and protected areas50. Therefore, the 12 most up-to-date global raster and vector datasets available at the time of production, were identified, collected, and processed into a uniform set of default covariates used for model fitting and prediction (Table 2).

The spatial variation of factors related to population distribution, such as night-light intensity, was measured using nightly day/night band (DNB) low-light imaging data collected by the Visible Infrared Imaging Radiometer Suite (VIIRS) aboard the Suomi National Polar Partnership (SNPP) satellite51,52. HydroSHEDS data53,54, derived from NASA’s Shuttle Radar Topography Mission (SRTM) elevation data55, was used to generate elevation and slope covariates. Specifically, the 3-arc second, void-filled digital elevation model (DEM) product was implemented56.

A global dataset of inland and ocean water was acquired from the European Space Agency’s (ESA) Climate Change Initiative (CCI) land cover project at a spatial resolution of 150m57,58. The data was built within the ESA-CCI project framework for the 2000–2012 period and enabled the generation of inland and ocean water masks. Global gridded land cover (LC) data for 2018 was obtained via the Copernicus Climate Change Service (C3S), using Intermediate Climate Data Records (ICDR) produced by the ESA-CCI project59,60,61. This data was used to identify different land cover types, and generate distance to land cover class covariates, whilst the permanent ice land cover class was incorporated with the ESA-CCI waterbodies dataset to generate a global mask of inland water and permanent ice. This global watermask was used to identify areas of non-human habitation due to the presence of waterbodies. The final stage of production for all covariates masked pixels identified as containing water, setting pixels to ‘No Data’ within these areas.

OpenStreetMap62 vector datasets were extracted for road and road intersection features via two distinct data repositories Geofabrik63 and BBBike64, respectively. Temperature and precipitation data, representative of the 1970–2000 period, were downloaded from WorldClim, version 2.1 climate data for 1970–200065. Moreover, a selection of pre-prepared covariates was extracted from the WorldPop open access gridded data archive to complete the set of covariates required for model fitting and prediction66. All data was available at 3 arc-second resolution, and had been already fully harmonised to support population distribution prediction applications27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,,67. These datasets included time-invariant covariates: distance to waterways, protected areas, and coastlines (Table 2).

The DLR’s World Settlement Footprint 3D (WSF3D) product was used to identify, quantify, and calculate distance to settlement in this research. The processing methodology of the WSF3D product is based on work presented by Esch et al.68. The WSF3D production approach is dependent on two predominant input datasets: (i) the 12 m spatial resolution TDX_DEM, and (ii) an updated version of the WSF imperviousness (WSF-Imp) dataset displaying the percent of impervious surface at ~10 m spatial resolution69,70 within the built-up area defined by the WSF2019 human settlement mask71,72. The DLR provided layers for each of the 40 countries specific to this research. A short description of each layer’s production process is denoted below68,73:

Building height (BH)

The ~450 m BH layer represents a spatial aggregation of the standard 90 m WSF3D BH layer, which was derived by measuring the height variations of vertical edges most likely related to building edges (BE) in the 12 m TDX-DEM within the settlement areas defined by the WSF-Imp layer. The height is reported in metres (m) in the final product.

Building area (BA)

The ~90 m BA layer is derived by firstly generating the Building Fraction (BF) layer, which measures percentage building coverage per ~90 m cell in a range of 0–100. This is produced by quantifying the built-up coverage at 12 m spatial resolution, derived from the joint analysis of the WSF-IMP, TDX-amplitude images (TDX-AMP), and BE. The BF is subsequently multiplied by the area of each ~90 m grid cell (~8100 m2 at the equator), thereby producing the BA. This area is reported in square metres (m2) in the final product.

Data processing

The population count data for each country was manually cleaned, processed, and harmonised to match to its corresponding official vector administrative unit dataset. The administrative and population count data was recoded, adding a ‘GID’ primary key field through which each row in the two datasets could be joined.

For each country (Table 1), the vector dataset representing its administrative units, used to match to the population counts, was cleaned and projected using the WGS 1984 geographic coordinate system; this system was selected to ensure uniformity across all generated covariate datasets. These datasets were then buffered by 100 km extent and rasterized at a resolution of 3 arc-seconds (approximately 100 m at the equator). These measures were taken to: (i) obtain a raster dataset of the study area to register and ensure uniformity across all raster covariates, (ii) produce a set of raster ‘distance to’ covariates that were unaffected by artificial boundary effects throughout spatial processing74, and (iii) conduct spatial processing on a buffered country-level basis, rather than on a global scale, to save processing time where necessary.

Default input covariates for the RF model were derived as follows. In most cases, raw datasets required specific cleaning and conversion methods to ensure format accessibility for further spatial processing. All raster variables representing continuous values (Table 2), were projected to WGS 1984 datum, resampled to 3 arc-second resolution, and matched to the rasterized study area. ‘No Data’ grid cells overlapping the rasterized buffered study area extent were filled with values of the nearest neighbour (using the Nibble tool available in ArcGIS 2.7.1). Finally, each covariate variable was extracted to the rasterized study area extent, maintaining uniformity of spatial extent, and resolution. All vector and raster datasets representing categorical variables were projected, rasterized, or resampled to 3 arc-second resolution, and matched to each rasterized buffered study area. Rasterized categorical variables were then converted into binary raster covariates, and subsequently utilised to generate continuous ‘distance to’ raster covariates (Table 2).

Bespoke measures were taken to prepare the land cover covariate variables. Similarly, to the aforementioned raster datasets representing categorical variables, the obtained C3S Global Land Cover data was projected, resampled to 3 arc-second resolution, and matched to the rasterized study area. The recoded global landcover dataset was then reclassified according to Sorichetta et al.28. Each land cover class was extracted and converted to a binary variable indicating presence/absence of the specified land cover class. Binary raster covariates were extracted to 100 km buffered study area raster datasets, and subsequently used to produce continuous ‘distance to’ raster covariates for each study area (Table 2). When a certain land cover class was completely absent, the covariate was disregarded for that specific country during RF model fitting and estimation, as on balance, the absence of the land cover type would not influence population distribution. The final land cover class (210) representing water and permanent ice cover distribution was disregarded in RF model fitting. Instead, the ESA’s waterbodies dataset was implemented as the ‘distance to’ water covariate variable, due to the improved spatial resolution it offers compared to the C3S Global Land Cover dataset (Table 2). Moreover, where available, waterbodies from the administrative unit boundary shapefiles were identified, rasterized, and incorporated into the waterbodies dataset, which was then processed using similar steps to the other raster datasets representing categorical variables, producing ‘distance to’ water covariate data for each country.

The distance to settlement covariate was prepared in the same way, generating a binary layer of building presence/absence from the WSF3D building area datasets; subsequently, ‘distance to’ settlement covariates were generated. The WSF3D building height data was prepared using a slightly differing methodology to the other continuous covariates; the settlement height data was extracted to the binary layer of WSF3D building presence/absence, instead of the official administrative boundaries. This measure ensures uniformity of building footprint delineation across all settlement covariates. ‘No Data’ grid cells in the building height layer overlapping the rasterized study area extent were filled with 0 values of the nearest neighbouring pixels. These building area and height datasets were multiplied using raster calculator to generate a building volume covariate.

Moreover, a bespoke road classification system was established and applied to the raw OpenStreetMap data, using the ‘fclass’ field. This classification system comprised three distinct classes: (i) Pedestrian access, (ii) Motor-vehicle access, and (iii) residential roads (Table 3). The application of this custom classification system aims to aid the improvement of population estimations, by providing enhanced covariate detail. Furthermore, vector point data representing road intersections was generated for each road class using ArcGIS’s Intersect tool. These vector data were used to generate ‘distance to’ covariates for road and road intersection features for all countries, matching the corresponding spatial resolution of 3 arc seconds Figure 1.

Fig. 1
figure 1

Schematic overview of the approach to generate gridded population estimates using the random forest (RF) model. For illustrative purposes, only a reduced set of considered covariates are shown here.

Full size image
Table 3 Reclassification of OSM roads and road intersection data.
Full size table

Vector road data were also used to produce road density covariate of corresponding spatial resolution raster. Road density is defined as the ratio of the length of the roads in the pixel to the land area of the pixel. Therefore, vector data of roads was intersected with a raster grid at a resolution of 3 arc seconds (approximately 100 m at the equator) to ensure that each pixel has exact information for the roads within this pixel. Figure 2 shows the example of road density in Colombia.

Fig. 2
figure 2

Road density in Bogotá, Colombia (3 arc second resolution).

Full size image

Furthermore, in order to estimate the road density within a grid cell/pixel neighbourhood, a non-parametric ‘kernel’ method was used. Using the kernel approximation, one can achieve a smoother density estimate, compared to that of a coarser distribution. Therefore, to investigate the effect of road density at different spatial scales, 4 bandwidths (500 m, 1000 m, 2000 m and 5000 m) were used for the kernel density estimations. Road intensity was calculated using Epanechnikov kernel function75. Figure 3 shows the example of road intensity in Colombia.

Fig. 3
figure 3

Road intensity (5 km bandwidth) in Bogotá, Colombia (3 arc second resolution).

Full size image

Random forest modelling scenarios

A set of modelling scenarios were devised to define the importance of covariate parameters for model fitting and prediction, as well as to enable the undertaking of a technical validation (Table 4). Specifically, the utility of WSF3D datasets when integrated into the RF modelling approach were to be assessed to assist the identification of the best final dataset for each country. Additionally, a simple areal-weighting (SAW) approach was generated as a comparison to assess the accuracy of RF-based dasymetric population modelling. These scenarios are detailed below (Table 4).

Table 4 Descriptions of modelled population distribution scenarios.
Full size table

Data Records

The high-resolution gridded population datasets detailed in this paper referring to the 40 countries listed in Table 1, are publicly and freely available through the WorldPop Data Repository76. The datasets can be downloaded as WinRAR Zip archives (win-rar.com) containing the population distribution datasets of the associated country for each of the five different RF modelling scenarios (Table 5).

Table 5 Name, description, and format of all files contained in each WinRAR zip archive related to the 40 countries listed in Table 1.
Full size table

Technical Validation

A technical validation framework was incorporated into the RF-modelling package, to ensure that the modelled population distribution outputs for each country and its administrative units were matching their population data input counterparts. However, as demonstrated by prior studies a ‘true-validation’ of gridded population datasets remains a significantly complex challenge due to the lack of high-resolution ground-truth data (i.e. population counts at the pixel level) required for an independent accuracy assessment of large-scale population models73.

Firstly, the technical validation framework calculates zonal sums in the RF-output population distribution Figure 4, and checks the total population per administrative unit for the RF-output distribution against the input population data (e.g. Figure 5). This ensures that the population total within each administrative unit for RF-model outputs, matches the population total within corresponding administrative units for population inputs data prior to the RF-modelling.

Fig. 4
figure 4

Estimated people per grid cell for 40 countries in Latin America and the Caribbean. Fitted using all base covariates including built area layers (for specific years see Table 1).

Full size image
Fig. 5
figure 5

Population distribution in Dominican Republic, 2020. (a) input count data at ‘finest-available’ administrative unit level, (b) modelling outputs following random forest fitting at 3 arc second resolution (approximately 100 m at the equator), RF model fitted according to Scenario 6 (Table 4).

Full size image

In addition to this primary technical check, the existing research in the field of large-scale population modelling has utilised a validation method that quantifies the internal accuracy of population distribution method, in terms of “how well and plausibly populations are distributed”77. This framework performs a selection of statistical analyses using the differences between population counts extracted from distributions modelled using a coarser level of administrative units (‘levelled-up’), and the population counts of the finest available administrative units (‘finest-available’), here, the official population count data69,73,78. To generate this coarser administrative level, population counts were aggregated for each country by merging together pairs of contiguous administrative units characterised by similar population density values; this method was chosen with the aim to merge pairs of low population density units together and pairs of high population density units together (Figure 6).

Fig. 6
figure 6

Comparison of ‘finest-available’, and the ‘levelled-up’ administrative units for Dominican Republic, 2020. ‘Finest-available’ units generated via aggregation of contiguous units.

Full size image

A subset of countries (Table 6), located in different parts of the LAC region, were selected to assess the increased accuracy of the RF-based dasymetric mapping approach with respect to a SAW approach assuming that the population of each administrative unit is evenly distributed within it; this subset is defined as countries with sufficient administrative units following aggregation (minimum of 25) to fit the RF model. Although it is possible to fit the RF model for a given country with fewer than 25 administrative units by pairing it with an additional country with similar characteristics, it was deemed that the influence of the additional, finer-resolution country object would distort the validation of the modelling approach. Therefore, these countries were omitted from the subset.

Table 6 Prediction accuracy of the RF model used to generate dasymetric weighting layers compared to the ‘levelled-up’ mapping approach.
Full size table

Model validation

The OOB error estimate (Table 6) is calculated during RF model fitting, and serves a robust and unbiased metric of the model’s internal prediction accuracy34. However, the OOB error estimate cannot be understood as the prediction error at the grid cell level, given that the RF model is fitted at the finest-available administrative level but predicts at the grid cell level. Furthermore, it should not be considered as the prediction error at the administrative unit level, via totalling of all final grid cell values within each administrative unit, and comparing it to the observed population count of the equivalent administrative unit. Nevertheless, it is expected that a higher accuracy of predicted values at the administrative level, should be associated with higher accuracy of the final gridded population distribution datasets33.

Between ‘finest-available’ and ‘levelled-up’ modelling scenarios, the OOB error increased and the percentage of variance explained decreased for 10 countries amongst the subset: ABW, ARG, CHL, COL, CUB, DOM, GTM, NIC, PRY and SUR (Table 6). The most significant difference is noted for ABW in which the OOB error more than doubled, whilst percentage of variance explained reduced by almost 10% (Table 6). However, the degree of differences in OOB error and percentage of variance explained were much less significant for the remaining countries within the subset (Table 6). There are some examples of countries in which the ‘levelled-up’ scenario exhibited reduced OOB error values and higher percentage of variance explained, compared to the ‘finest-available’ modelling output; most notable amongst these are MEX and SLV (Table 6). The OOB error value for both MEX and SLV decreased by 0.03, whilst the percentage of variance explained increased by 1.7% and 2.8%, respectively (Table 6).

WSF3D quantitative assessment

For each country within this subset, a selection of spatial error metrics were identified and calculated to assess the accuracy and reported differences between the actual and the estimated values for each country’s ‘finest-available’ administrative unit; in this case the actual values are obtained from the input population count data at the ‘finest-available’ administrative unit level, whilst the estimated values are derived from Zonal Statistics sum calculations of each resultant RF modelling scenario output (Table 4) at the same ‘finest-available’ administrative level. For each country (Table 6) and each modelled RF-scenario (Table 4), the following error metrics are derived in Table 7.

Table 7 Descriptive metrics for accuracy assessment at the validation unit level for modelling scenarios (Table 4).
Full size table

For each country, four accuracy metrics were used to assess how well each RF modelling scenario distributed the population. Both the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) measure the absolute differences between the actual (popa) and estimated population (pope) counts of the L1-base units73,79. However, MAE is known to be more robust to outliers80, since RMSE penalises significant errors by squaring differences, compared to MAE which weights each error equally73. Conversely, the Mean Absolute Percentage Error (MAPE)  is the MAE adjusted to each level of analysis, calculated as MAE divided by the average population of each country78. Similarly, the RMSE is also expressed as a percentage of the mean population size of the administrative unit level via the root mean square percentage error (RMSPE). These metrics enable comparison across countries by omitting the bias caused by different population totals and number of administrative units; furthermore, ‘percentage error’ metrics help to determine if errors generated by different modelling layers are similar and systematic, or if different behaviours are observable across countries73.

Results, summarised in Figure 7, indicate that the high-resolution gridded population datasets produced under this project’s framework outperform their corresponding SAW-based outputs across almost all cross-sections of metrics, countries, and modelling scenarios. The first exception to this finding is El Salvador (SLV), in which the calculated RMSE value increases from 8159 to a maximum of 8779 between the SAW and Scenario 6 (Table 4) modelling approaches, respectively (Figure 7). The second exception is Guatemala (GTM), in which the calculated MAPE value of 17.05% for the SAW modelling approach is lower than Scenarios 1 and 3 (Table 4); nevertheless the remaining modelling scenarios for Guatemala (Scenarios 2, 4, and 5) are an improvement on the SAW-based outputs according to calculated MAPE values of 16.64, 17.00, and 16.96, respectively (Figure 7).

Fig. 7
figure 7

Accuracy assessment results for modelling population density of all scenarios (Table 4) for each country (ISO-3).

Full size image

Beyond these exceptions, according to the calculated accuracy assessment metrics (Figure 7), Scenario 6 (Table 4) is the best performing modelling method for 31 of the 68 country-accuracy metric combinations. Moreover, Scenarios 3 and 5 are the next best performing modelling methods, registering the best accuracy metric result for 14 and 13 of the 68 country-accuracy metric combinations, respectively; as discussed above, the SAW-based modelling approach was found to be the best performing modelling scenario in only one case (SLV-MAE). These findings highlight a number of concepts, including (i) the importance of the building area covariate to RF-model fitting, (ii) the value of integrating all building covariates to RF-model fitting, and (iii) the increased accuracy of the RF dasymetric disaggregation approach compared to a SAW-based disaggregation.

Usage Notes

In particular, the presented gridded datasets provide improved spatial detail of the residential population distribution at sub-administrative unit level comparatively to most publicly-available (i) administrative unit-level official and non-official estimates or projections, which implicitly rely on the assumption that the population is homogeneously distributed within each units, and (ii) gridded population datasets, which are based on non-official estimates or projections. This is achieved via the disaggregation of the most recent and finest administrative unit-level official population projections, produced by 40 NSOs and processed with support by the UNFPA Regional Office for Latin America and the Caribbean, UNFPA’s Population and Development Branch and the Information Management Branch of the UN Office for the Coordination of Humanitarian Affairs (OCHA).

Furthermore, these gridded population distribution datasets represent a consistent and comparable format, as well as a scalable framework, providing flexibility in (i) summarisation to any spatial area of interest (e.g., areas impacted by natural and/or man-made hazards which may not correspond to predefined artificial administrative boundaries), and (ii) analysis and data integration (e.g., GIS and remote sensing data, such as locations of healthcare facilities and CO2 emissions, respectively). Thereby, they can be effectively considered for  planning and supporting interventions and applications (e.g., planning for elections, assessing exposure to natural hazards, and measuring demand for services), measuring progress (e.g., measuring and monitoring the SDGs and their indicators), and performing analyses (e.g., predicting response variables intrinsically dependent on the population distribution, and modelling epidemic spreads).

However, it is important to note that there are also a number of limitations, caveats, and assumptions inherent in the modelling approach used to produce the gridded population datasets, that should be considered before using them. For consistency, all datasets were produced using a fixed number of ancillary covariates available for all countries, and thus only a limited number of factors, potentially related to population presence and densities in each country, have been considered overall. For this reason, which represents a trade-off in the production of generalizable models, the accuracy of the gridded population datasets for some of the countries could be improved by considering additional, locally-specific factors that could help to increase the percentage of variance explained by the corresponding RF model.

Other limitations are represented by (i) the fact that the spatial detail of the administrative unit-level population projections was not the same for all countries (refer to the “Unit level” column in Table 1), with the use of smaller administrative units for a given countries translating into higher accuracy of the corresponding gridded population dataset, and (ii) the fact that, because of the lack of enough administrative units to fit a country specific RF model, the gridded population datasets for a number of countries and islands have been produced using RF models referring to another country or a set of countries, and “Grouped Islands”, respectively (refer to the “Modelled with” column in Table 1)38.

Additionally, it may be worth to reflect on the fact that the official administrative unit-level population census-based figures and projections, used as inputs to the RF model, may or may not have captured effects of potential rapid onset events responsible for abrupt fluctuations of population numbers at the administrative unit level (e.g., forced displacements due to natural disasters). Similarly, the gridded population datasets produced using them do not account for seasonal or intra-annual population mobility between administrative units.

Upon aggregation, gridded population datasets constructed using this disaggregation methodology are proved to be more accurate at representing human population distribution compared to those produced using an equal-area approach28,33. The reliability of the data product is unknown at the grid-cell level, therefore it is recommended that population datasets be aggregated before use instead of at the grid-cell level77. Furthermore, it is important to highlight that gridded population datasets give end-users the flexibility to aggregate population according to different boundaries and/or areas (i.e. boundaries and/or areas that do not align with the administrative unit boundaries of the input population data).

Furthermore, it is critical to consider that these gridded population datasets represent modelling outputs generated using a number of ancillary covariates and thus, to avoid circularity, they should not be used to make predictions about, or explore relationships with, any of the factors included in the model (e.g., correlating population distribution with settlement distribution). If there is such need, ideally, the modelling process should be re-run using the publicly and freely available WorldPop-RF code (https://github.com/wpgp/popRF) with the covariate of interest being removed to avoid issues relating to endogeneity.

Finally, it is also important to note that most of the considered ancillary covariates are derived from modelling outputs themselves, and thus they have a degree of uncertainty that carries over into the gridded population datasets.