Population Estimation Using Geographically Weighted Regression

In view of the variability in the accuracy of the land use/cover data derived from satellite data, population estimation using a global ordinary linear regression (OLS) model cannot deal with the spatial non-stationarity problem. This research applied a local geographically weighted regression (GWR) model using four land use area variables (high-density urban use, low-density urban use, cropland, and forest) for population estimation in the city of Atlanta, Georgia at the census tract level, and found that the four-class local GWR model could help improve the accuracy of population estimation over a four-class global OLS model by 28%. The spatial non-stationarity seemed to be affected by the core-periphery locations in the city. This research also confirmed that a regional regression approach based on census tracts aggregated by counties could minimize spatial non-stationarity, and produced more accurate population estimation than that for the study area as a whole using the global OLS model, thus pointing to a way to improve population estimation using the traditional global model. The GWR model also identified the parts of the city where spatial non-stationarity is least influential, and as a result, a global OLS model could be used for population estimation.