RESUMO
BACKGROUND: This is the third paper in a 3-paper series evaluating alternative models for rapidly estimating neighborhood populations using limited survey data, augmented with aerial imagery. METHODS: Bayesian methods were used to sample the large solution space of candidate regression models for estimating population density. RESULTS: We accurately estimated the population densities and counts of 20 neighborhoods in the city of Bo, Sierra Leone, using statistical measures derived from Landsat multi-band satellite imagery. The best regression model proposed estimated the latter with an absolute median proportional error of 8.0%, while the total population of the 20 neighborhoods was estimated with an error of less than 1.0%. We also compare our results with those obtained using an empirical Bayes approach. CONCLUSIONS: Our approach provides a rapid and effective method for constructing predictive models for population densities and counts utilizing remote sensing imagery. Our results, including cross-validation analysis, suggest that masking non-urban areas in the Landsat section images prior to computing the candidate covariate regressors should further improve model generality.