Dasymetric mapping of population data using geographically-weighted random forest regression

Keywords: Population downscaling, landscape metrics, impervious cover percentage
Session Type: Virtual Paper
Day: Friday
Session Start / End Time: 4/9/2021 08:00 AM (Pacific Time (US & Canada)) - 4/9/2021 09:15 AM (Pacific Time (US & Canada))
Room: Virtual 9

Abstract

Population data are always aggregated at the census enumeration units. In U.S., the smallest population enumeration units are not static and changes from time to time. This discrepancy in the spatial resolution of population data would retard the analysis of population and its further functioning as a data input in other models. Numerous studies have been researched on the dasymetric mapping of population, which downscales population data from a coarser resolution to a finer resolution. The main goal of this study is to downscale population from census tract to block group by using geographically-weighted random forest regression and dasymetric mapping. The relationship between population and a series of predictors, including distance to city center, numerous landscape metrics, and impervious cover percentage data derived from Spectral Mixture Analysis (SMA) was established on the census tract level, and then applied to the block group level to predict population at block group level. This predicted population was further used as relative weights to redistribute population from census tract level to the corresponding block groups. Accuracy assessment was conducted by comparing the final predicted population at block group level with the actual population at block group level.