Dasymetric modelling of small-area population distribution using land cover and light emissions data

Despite the improvements made in census procedures over recent decades, the availability of detailed population data is limited. For many applications, including environmental and health analyses, methods are therefore needed to model population distribution at the small-area level. With the development of GIS and remote sensing techniques, the ability to develop such models has greatly improved. This paper describes a GIS-based approach using remotely sensed land cover and nighttime light emissions data to model population distribution at the land parcel level across the European Union. Light emission data from the DMSP satellites were first resampled and modelled using kriging and inverse distance weighting methods to provide a 200-m resolution light emissions map. This was then matched to CORINE land cover classes across the EU. Regression methods were used to derive models of relationships between census population counts (at NUTS 5 level) and land cover area and light emissions. Models were developed at both national and EU scale, using a range of different modelling strategies. Model performance, as indicated by the regression statistics, was seen to be good, with R-2 typically in the order of 0.8-0.9 and SEE ca. 4000 people. In southern countries, especially, incorporation of light emissions data was found to improve model performance considerably compared to models based only on land cover data. More detailed post hoe validation in Great Britain, using independent data on population at census tract (enumeration district and output area) and postcode level, for 1991 and 2001, showed that models gave good predictions of population at the 1 km level (R-2 > 0.9), but were less reliable at resolutions below ca. 500 m. Impending enhancements in the available land cover and light emissions data are expected to improve the capability of this modelling approach in the future. (c) 2007 Elsevier Inc. All rights reserved.