Gridded Population Maps Informed by Different Built Settlement Products

Abstract

The spatial distribution of humans on the earth is critical knowledge that informs many disciplines and is available in a spatially explicit manner through gridded population techniques. While many approaches exist to produce specialized gridded population maps, little has been done to explore how remotely sensed, built-area datasets might be used to dasymetrically constrain these estimates. This study presents the effectiveness of three different high-resolution built area datasets for producing gridded population estimates through the dasymetric disaggregation of census counts in Haiti, Malawi, Madagascar, Nepal, Rwanda, and Thailand. Modeling techniques include a binary dasymetric redistribution, a random forest with a dasymetric component, and a hybrid of the previous two. The relative merits of these approaches and the data are discussed with regards to studying human populations and related spatially explicit phenomena. Results showed that the accuracy of random forest and hybrid models was comparable in five of six countries.

Keywords: binary dasymetric; built areas; geographic information systems; geography; gridded population distribution; random forest; regression; remote sensing.

Figure 4

An example of the three primary model types and the rasters they produce for Kigali, Rwanda. Pictured built area extent on models 1 and 3 is the combination layer described in Section 3.1.2.

Figure 1

Census unit aggregation procedure in which 1/3 of the finest available units are randomly selected independent of spatial size or any other stratification and merged with its neighbor with the longest shared border until the target 2/3 census count is reached.

Figure 2

Workflow for generating the population distribution maps.

Figure 3

Model enumeration and visual representation of feature overlays used to produce output datasets by means of dasymetric redistribution. Ordered by increasing complexity.

Figure 5

Box plots of global variable importance presented as mean squared error for each covariate class. The median is represented by the black bar, while the whiskers represent the min/max values within 1.5× inter-quartile range. Variables sourced in Table 4.