A predictive modelling technique for human population distribution and abundance estimation using remote-sensing and geospatial data in a rural mountainous area in Kenya

Abstract

This study presents a predictive modelling technique to map population distribution and abundance for rural areas in Africa. Prediction models were created using a generalized regression analysis and spatial prediction (GRASP) method that uses the generalized additive model (GAM) regression technique. Dwelling unit presence–absence was mapped from airborne images covering 98 km² (30% of the study area) and used as a response variable. Remote-sensing-based (reflectance, texture and land cover) and geospatial (topography, climate and distance) data were used as predictors. For the rest of the study area (228 km²; 70%), GAM models were extrapolated, and prediction maps constructed. Model performance was measured as explanatory power (adj.D ², adjusted deviance change), predictive power (area under the receiver operator curve, AUC) and kappa value (κ). GAM models explained 19–31% of the variation in dwelling-unit occurrence and 28–47% of the variation in human population abundance. The predictive power for population distribution GAM models was good (AUC of 0.80–0.86). This study shows that for the prediction of dwelling-unit distribution and for human population abundance, the best modelling performance was achieved using combined geospatial- and remote-sensing-based predictor variables. The best predictors for modelling the variability in human population distribution using combined predictors were angular second moment image-texture measurement, precipitation, mean elevation, surface reflectance for Satellite Pour l'Observation de la Terre (SPOT) red and near-infrared (NIR) bands, correlation image-texture measurement and distance to roads, respectively. The population-abundance modelling result was compared with two existing global population datasets: Gridded Population of the World version 3 (GPWv3) and LandScan 2005. The result showed that for regional and local-scale population-estimation probability, models created using remotely sensed and geospatial data were superior compared to GPWv3 or LandScan 2005 data products. Population models had high correlation with Kenyan population census data for 1999 in mountainous sub-locations and low correlation for sub-locations that also extended into the lowlands.

Acknowledgements

The research was financially supported by the Academy of Finland funded TAITATOO project (http://www.helsinki.fi/science/taita). Thanks to Hanna Piepponen, who assisted in interpreting and digitizing the dwelling-unit points and to Milla Lötjönen and Antero Keskinen, who created the airborne digital image mosaics. Two anonymous reviewers greatly improved an earlier version of the manuscript.

Notes

*This paper came from a workshop entitled ‘Potentialities and Limitations in the Use of Remote Sensing for Detecting and Monitoring Environmental Change in the Horn of Africa’. The workshop was organized by Somalia Water and Land Information Management (SWALIM) between 13th and 14th June 2007 at Holiday Inn in Nairobi, Kenya. SWALIM is a project of the UN-FAO in Somalia (www.faoswalim.org).