Spatial Sampling and Sensitivity of Random Forest Model in Gridded Population Modeling

Authors: Parmanand Sinha*, University of Louisville, Andrea Gaughan, University of Louisville, Forrest Stevens, University of Louisville
Topics: Population Geography, Spatial Analysis & Modeling, Geographic Information Science and Systems
Keywords: gridded population modeling, dasymetric modeling, spatial sampling, high-resolution population
Session Type: Paper
Day: 4/14/2018
Start / End Time: 10:00 AM / 11:40 AM
Room: Napoleon A3, Sheraton, 3rd Floor
Presentation File: No File Uploaded


Rarely is spatially-explicit sampled data made up of completely uncorrelated observations. In gridded population modeling, an added level of complexity exists, as oftentimes, gridded outputs are trained at a different level, such as a census unit (i.e. source unit), than the one for which they are created such as the pixel or grid cell (i.e. target unit). In most of the cases, the actual distribution of population for the target unit is unknown. Typically, a top-down, dasymetric approach of gridded population modeling utilizes coarse level data to redistribute population at a finer scale. We examine the effects of biased sampling across source to target scale in the context of a regression-weighted, dasymetric mapping approach. The modeling methodology is similar to the current method used by the WorldPop Project (www.worldpop.org). The objective of this study is two-fold. First, we examined the different levels of bias introduced to a model through various sampling frames and levels of spatial correlation when applied to the training data. Second, we investigated its effect on prediction across the scales and the output of the dasymetric redistribution process. These results have been benchmarked against simple random sampling and geographically stratified random sampling. Samples with lower spatial autocorrelation than the true population indicate underprediction at census scale and higher deviation at pixel scale, while samples with spatial autocorrelation higher than actual level indicate overprediction at census scale but comparatively less deviation at the pixel level. We will discuss these findings and its impact on dasymetric redistribution in detail.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login