Modelling housing unit density from land cover metrics: a Midwestern US example

Abstract

Geographic applications frequently require the gathering and analysis of socioeconomic data. For many nations, these data are normally collected through a census. However, during the intercensal period (5–10 years), these data lose their currency and must be updated. The objective of this project was to estimate housing unit density from Landsat ETM+ imagery in the Terre Haute, IN, USA, region. Modelling was done for 1945 census blocks in the study area containing 30 972 housing units. Landtype, as represented by six cluster classes, was used as the primary surrogate for housing unit density. The percentage of each landtype within the census blocks was calculated. Other landscape metrics representing landtype patch dominance and diversity were also calculated on a per-block basis. Housing unit density within the census block was then modelled as a function of those percentages and metrics using discriminant analysis and multiple regression. The simple correlation between the observed and modelled housing unit density was 0.79. The mean residual error produced by the model was 0.37 housing units per hectare.

Keywords:

• housing unit density estimation
• population estimation
• landscape metrics

Notes

1. There are 2550 census blocks with an average size of about 14 hectares in the Terre Haute study area. These 2550 blocks are aggregated by the census bureau into 50 block groups.

2. The term ‘pixel-based estimation’ has been defined various ways. We follow the convention of Lo (2003).

3. rho hereafter. Because of the skewed housing unit density distribution (Figure 2), the application of nonparametric descriptives was indicated.

4. To avoid the tedious repetition of the phrase ‘housing units per hectare’, the units will be implicitly understood hereafter whenever housing unit density is quantified.