UNEP Logo CIESIN logo

African Population Database Documentation [ Back | Contents | Next ]
 

Part II: Raster data

One objective of improving the boundary and population data for Africa as described in the previous sections was to develop new population distribution surfaces. The global demography project at NCGIA produced a gridded data set for the whole world which was constructed using a smoothing technique that has the property of preserving population totals within each administrative unit. The raster surfaces based on the approach outlined in the following section were constructed using an alternative interpolation method. This method preserves population totals in each district as well and incorporates additional information on settlements, transport infrastructure and other features important in determining population distribution. The conversion of population data from a vector or polygon representation to raster format has the added advantage that the data can be more easily combined with many spatially referenced physical data sets which are most often stored in a gridded format. This facilitates the use of these data in research and policy analysis and will hopefully contribute towards an increasingly integrated approach to the study of problems related to population, the environment, economics and culture as advocated, among others, by Cohen (1995). The approach outlined here as well as alternative approaches to spatial population modeling are discussed in more detail in Deichmann (1996b).

II.1. Gridding approach

The basic assumption upon which the construction of population distribution raster grids for Africa is based is that population densities are strongly correlated with accessibility. Accessibility is most generally defined as the relative opportunity of interaction and contact. These opportunities are the largest where people are concentrated and transport infrastructure is well developed. Within any given area, we therefore expect a larger share of the known total population to live in more accessible regions compared to areas that are less well connected to major urban centers.

Summary description of the method

The method for the development of population raster grids consists of the following steps. The most important input into the model is information about the transportation network consisting of roads, railroads and navigable rivers. The second main component is information on urban centers. Data on the location and size of as many towns and cities as can be identified are collected, and these settlements are linked to the transport network. This information is then used to compute a simple measure of accessibility for each node in the network. The measure is the so-called population potential which is the sum of the population of towns in the vicinity of the current node weighted by a function of distance, whereby network distances rather than straight-line distances are used. The following figure illustrates the computation of the accessibility index for a single node.

Graphic of the relationship between nodes, towns and the transport networks in the accessibility model

The computed accessibility estimates for each node are subsequently interpolated onto a regular raster surface. Raster data on inland water bodies (lakes and glaciers), protected areas and altitude are then used to adjust the accessibility surface heuristically. Finally, the population totals estimated for each administrative unit (as described in the first part of this documentation) are distributed in proportion to the accessibility index measures estimated for each grid cell. The resulting population counts in each pixel can then be converted to densities for further analysis and mapping. Each of these steps will now be described in more detail.

Construction of the transportation network

There are few data sources that provide consistent, geographically referenced base data layers for large areas such as an entire continent. The transportation infrastructure data for this project was constructed using the following data sets: roads and railways from the Digital Chart of the World (DCW) and major navigable waterways from the World Boundary Databank II (WBDII). WBDII originated at the U.S. Central Intelligence Agency and an Arc/Info version is available from the Environmental Systems Research Institute (ESRI). The nominal scale of WBDII is 1:3 million. The scale of the DCW base maps (the Operational Navigational Charts) is 1:1 million. Since we also used DCW for the international boundaries in the administrative unit data layers and since WBDII and DCW appear to share common ancestors, a good fit exists between the individual data layers.

The road dataset was improved by including road type characteristics from 1:4,000,000 paper maps (Michelin, 2004). Roads were identified as follows.

Road type Total length in km (and as a % of all roads) Estimated travel speed (kmph)
Primary (Motorway/Major road) 132,000 (11%) 50
Secondary (All weather/Improved) 282,000 (22%) 35
Tertiary (Partially improved/Earth roads) 839,000 (67%) 25
Other network types
Railroads 80,722 50
Navigable rivers 93,722 20
Auxillary arcs 2,771,705 5

A brief technical discussion is now required to clarify the structure of the transportation data. After merging the individual components of the transport network into one data layer there are still no connections between the individual components (e.g., railroads and rivers). To allow the model to choose the most efficient means of transport at any point in the network, the intersections between the individual transport layers need to be found. This is a standard GIS operation that results in a well-structured data layer of arcs (or links) representing roads, railroads or rivers. These are connected by nodes which are intersections of two or more arcs of different or similar types. Nodes, of course also represent the end of an unconnected arc.

The program used for calculating accessibility produces an estimate for each node in the network. The problem in an application where the network is sparse in many regions is that no values are derived for areas that are not connected to the network. In the Africa application this applies to large areas, for example, in the Northern and Southern African desert regions. Also, WBDII and DCW only include fairly important transport features that are relevant at a cartographic scale of 1:1 or 1:3 million. One solution is to calculate the accessibility index for the center of each grid cell of the subsequently generated output raster. From each grid cell, the distance to the closest transport feature could be calculated and added to the network distances to the closest towns. This approach was used by Geertman and van Eck (1995).

However, this approach is not realistic where the closest access point to the transport network is at a location which is actually far away from urban centers. Another network access location may be further away from the grid cell initially, but better connected to major towns. To evaluate different options of network access for each grid cell would be impractical, and we therefore chose a different approach. In areas where the transport network is sparse, auxiliary arcs were added which could be thought of as "feeder roads". Essentially, this implies that people who may be living in these remote areas are using trails or tracks to get to the main transport network first and then continue their travel to the nearest city along the fastest routes. The algorithm automatically determines which network access is optimal in minimizing overall travel times.

It would be straightforward to use simple network distance for the calculation of accessibility. However, different arcs representing various transportation modes are associated with quite different travel speeds. For example, a kilometer travel on a paved road will take much less time than the same distance on a river. Instead of simple distance, we therefore used cumulative travel time as the weight in the accessibility calculation. Each arc in the resulting complete transportation network is associated with an estimate of average travel speed that is thought to be possible. Primary roads were assumed to allow for a travel time of 50kmph, secondary roads were assigned a speed of 35kmph, and tertiary roads were assigned a speed of 25kmph. 50kmph were used for railroads, and 20kmph for navigable rivers, and 5kmph for the auxiliary network access routes. For each arc, we calculated the real-world distance in kilometers.

However, all data layers are referenced in geographic (latitude/longitude) coordinates and no map projection is able to represent real-world distances in all directions with sufficient accuracy for large regions. We therefore calculated the correct length of each arc as the sum of the great-circle distances of all individual segments that make up the arc between two nodes. The time it takes to traverse each section of the transport network is then simply its length in km divided by the travel speed associated with the specific mode of transport.

Setting up urban data

The accessibility index is the sum of the population totals of the towns in the vicinity of the current location weighted by the network travel time ("distance") to those towns. Data on the location and size of urban centers were provided by CIESIN and accounted for 3,997 cities with population estimates for year 2000.

In areas where the points were sparsely distributed (e.g. Somalia) we added a further 132 auxillary populated places based on gazeteer information and the Michelin 1:4 million road maps and assigned them nominal populations of 1,000.

Towns need to be connected to the transport network to enable the accessibility calculation algorithm to find the closest towns for each node in the network. Each settlement was therefore assigned to the network node closest to its recorded location.

Run accessibility calculation

For the actual accessibility calculation we used a stand-alone program written in the C programming language. This program reads the entire network definition which consists of (a) the identifiers for each node and the population size of the town that corresponds to the node - zero in most cases, indicating that no town is located at the node -, and (b) the identifiers of the two nodes that define each arc and the travel time required to traverse the arc.

A further option of the program that allows for considering the direction of travel along an arc was not used. This implies that there are no "one-way streets" and that travel time is the same regardless of which way one travels. This assumption could be relaxed since, for example, travel speeds are lower up-river than down-river, but the added gain in realism will not compensate for the additional effort required in defining these details. Also, no further assumptions are made about modal choice. In moving through the network, an imaginary traveler may change his or her means of transport at will. This is unrealistic since a switch, say from road travel to a train and on to a boat, are all associated with delays. Even so, in order to keep the model simple (and run-times manageable) we did not introduce a penalty for switching the transport mode. A modification relevant to an application in a regional setting was made, however. For any arc that crosses an international boundary, the travel time was increased by 30 minutes reflecting delays in border crossings. This added travel time could be varied depending on the relations between two neighboring countries. This would either require subjective judgment or very detailed information on the permeability of international borders.

For each node in the network, the program now finds the network path to each of a specified number of towns that results in the lowest overall travel time. In the initial program specification, all towns reached within a user-defined specified travel time (e.g., 5 hours) were determined. However, in areas where towns are sparsely distributed and the number of nodes and arcs is large, this resulted in unacceptably long run-times. Instead, we modified the program to find the closest four towns or less if fewer than four towns were accessible within a more generous threshold travel time. This also makes the index somewhat more comparable across large areas, since the previous specification resulted in the accessibility index for some densely urbanized areas to be based on fifty or more towns, while other regions would only contain two or three.

For the shortest path calculation the program uses the standard Dijkstra algorithm. The program section used for this search consists of a modified version of a fast implementation of this algorithm developed by Tom Cova, a transportation GIS specialist at NCGIA. The Dijkstra algorithm evaluates the network structure around the current location starting from the center and reaching out further and further. For applications in which only one origin-destination pair is of interest, this is inefficient and various modifications have been suggested to speed up the search. In this application, in contrast, the interest is in finding the shortest path to all towns within the vicinity and the Dikstra's "shortcoming" is actually a bonus. The slightly modified algorithm thus "collects" towns as it ventures out from the originating node. Once four towns have been found and the program has determined that all additional connected arcs will not lead to a town that is closer than those already found, the search is terminated and the town populations and travel times are passed to a program section that calculates the accessibility measure.

This measure is the sum of the town populations weighted by a negative exponential function of travel time ("distance"), i.e.

Negative exponential model of accessibility

where Vi is the accessibility estimate for node i, Pk is the population of town k, dik is the travel time/distance between node i and town k, and is the distance to the point of inflection in the distance decay function. This parameter was set to one hour in this case which means that the influence of a town one hour away decreases to about 60 percent, and a town two hours away will only contribute 14 percent of its total population to the accessibility index. Rather than using total urban population, we applied a square root transformation to the population figures, implying that each additional person living in a city has an increasingly lower influence. This transformation avoids an exaggerated influence of very large mega-cities while being less of an equalizer than the more common log-transformation.

Interpolation

The accessibility index that is available for each of the nodes in the network needs to be converted into a regular raster grid. We used a simple inverse distance interpolation procedure that resulted in a relatively smooth surface. A problem with this technique is that interpolated values will not fall outside the range of the values recorded at the neighboring node locations. In analogy to interpolating elevation data: if recorded values are available only for locations on the slope of a mountain but not at the peak, the interpolated value for the summit location will be underestimated. Conversely in our application, if values are recorded only for network nodes, but not for areas that are remote from transport routes (e.g., deserts), then using the neighboring node values for interpolation will overestimate the accessibility for the remote location.

Yet, experiments with other interpolation procedures did not result in satisfactory results. Thin plate spline interpolation may be more appealing theoretically since it would allow values at interpolated locations to fall below (or above) those that are recorded at neighboring locations if the overall tension surface suggests a corresponding trend. However, the values estimated for some locations were clearly out of the range of what would be reasonable. Given the large number of nodes introduced in remote areas by adding the auxiliary access routes, we considerthe simple inverse distance interpolation to be sufficiently accurate.

Adjustment of the accessibility measure

Three additional data sets were used to adjust the resulting accessibility index grid: inland water bodies, protected areas, and elevation. Lake areas were masked and grid cells that fell onto a glacier were assigned an accessibility value of zero. This information was derived from the DCW lakes layer.

GIS data layers on protected areas were obtained from the World Conservation Monitoring Center (WCMC). Unfortunately, little information about each protected area was available besides its name, such that it was impossible to relate, for example, protection status to an estimate of how much the areas may still be used and inhabited by people. We reduced the accessibility index for grid cells that fell into national parks to 20 percent of the original value and for areas falling into forest reserves to 50 percent. These values are subjectively determined to allow for the fact that the protection of protected areas is not always perfect. Since most of these parks are in remote region, the change in predicted population densities that would be introduced by varying the adjustment factors should be small.

In contrast to the Asia population surfaces, no adjustment for elevation was made in the Africa data sets. Only few areas in Africa are uninhabitable due to high altitude. These areas (e.g., Mount Kenya or Kilimanjaro) are also protected areas and have thus been considered already.

Distribution of population

The distribution of the population total available for each administrative unit over the grid cells that fall into that unit is straightforward. The accessibility values estimated for each grid cell serve as weights to distribute population proportionately. First the grid cells in the accessibility index are summed within each district. Each value is divided by the corresponding district sum such that the resulting weights sum to one within each administrative unit. Multiplying each cell value by the total population yields the estimated number of people residing in each grid cell. The standardization of the accessibility index implies that the absolute magnitudes of the predicted access values are unimportant - only the variation within the administrative unit determines population densities within each district.

Again, we have to take account of the fact that all GIS data layers and raster grids are referenced in latitude/longitude coordinates. This means that grid cells further away from the equator represent a smaller real-world area than grid cells close to it. For example, a 2.5 minute grid cell has a real-world area of 10.8 square km at 60 degrees latitude, of 18.6 square km at 30 degrees and of 21.4 square km at the equator. We therefore weighted the accessibility index value for each grid cell by the actual area of the grid cell before standardizing the values within each district.

Because only the relative magnitudes of the accessibility index are important in distributing total population, and since most administrative units are fairly small, the error introduced by the distortions of the geographic coordinate system will usually be insignificant. However, in areas where the available resolution of the administrative units is fairly low, the difference in the actual areas of grid cells within a district that are located further away from the equator compared to those closer to the equator can be relatively large. For example, in Asia which reaches into much higher latitudes than Africa, the resulting difference in predicted population densities using undadjusted and adjusted accessibility values reached up to eight people per square km. From the grid cells of total population, population density images are created by dividing the population counts estimated for each grid cell by the real-world area in square km of that cell.

Evaluating the accuracy of this interpolation method is difficult in the absence of very high resolution population data (e.g., by enumeration areas) that could be used as a benchmark. For the Asian database a simple experiment was conducted in which state level population figures for India were interpolated. The total population allocated to each district could then be compared to the actual district figures. The differences are acceptable in relatively homogeneous regions but are obviously quite large in areas where population distribution is very scattered such as in high mountain or desert regions. The same results could be expected for Africa. The model will work better, the more detailed the administrative data, the more urban population figures are available, and the more homogeneous the population is distributed.


UNEP Logo UN Logo United Nations Environment Programme
Global Resource Information Database

Division of Early Warning & Assessment - North America

Page Modified: September 2004