Social cartography and satellite-derived building coverage for post-census population estimates in difficult-to-access regions of Colombia

Abstract

Effective government services rely on accurate population numbers to allocate resources. In Colombia and globally, census enumeration is challenging in remote regions and where armed conflict is occurring. During census preparations, the Colombian National Administrative Department of Statistics conducted social cartography workshops where community representatives estimated numbers of dwellings and people throughout their regions. We repurposed this information combining it with remotely-sensed buildings and other geospatial data. We developed hierarchical Bayesian models to estimate building counts and population sizes, trained using nearby full-coverage census enumerations, and assessed using 10-fold cross-validation. We compared models to assess the relative contribution of community knowledge, remotely-sensed buildings, and their combination to model fit. The “community” model was unbiased but imprecise; the “satellite” model was more precise but biased; the “combination” model had the best overall accuracy. Results reaffirmed the power of remotely-sensed buildings for population estimation and highlighted the value of incorporating local knowledge.