How Remotely Sensed Built Areas And Their Realizations Inform And Constrain Gridded Population Models

Many disciplines require consistent, comparable areal information about where people are. Gridded population estimates fill this need typically by disaggregating census data or predicting population counts from ancillary information within identical pixel sizes across variable extents. Remotely-sensed built areas are an extremely important form of such ancillary data, and how they inform and constrain these techniques has yet to be fully explored. This study assesses the effectiveness of three remotely-sensed, built area data sets (High Resolution Settlement Layer, Global Human Settlement Layer and the World Settlement Footprint) in producing gridded population estimates. Three mapping techniques are applied with those three data sets in Haiti, Malawi, Madagascar, Nepal, Rwanda, and Thailand. The modeling techniques include binary dasymetric redistribution, random forest with a dasymetric component, and a hybrid of the previous two. Results show that in five of the six countries the random forest and hybrid models were comparable, although there may be case-specific instances in which statistically complex models are not as robust as simply redistributing population counts into "built" grid cells. Relative contributions from different built area data to informing population modelling was discussed in relation to the spatial resolution, modeling techniques, and underlying built area vs. built-up definitions. Further investigation is also discussed regarding the inclusion of additional built datasets, ancillary inputs, and reference data.