Dasymetric mapping for population and sociodemographic data redistribution - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 192

For the fourth process, we used the Dasymetric Mapping
Extension (Sleeter and Gould, 2007) developed by the USGS for
ArcGIS, which is based on the three-class dasymetric mapping
process. We again used the NLCD 2001 as our ancillary layer, but
for this implementation we reclassified the data into three classes:
2

14.3.3 Areal interpolation

To demonstrate the redistribution of sociodemographic data,
we use the example of total population, and the population
of individuals living at or below the poverty level. We began by
mapping these variables using the binary dasymetric method with
LandPro 2001 as the ancillary data. Figure 14.5(a, b) depicts the
dasymetric estimates for total population density, and the density
of the population living in poverty overlaid with voting districts,
respectively. For areal interpolation, we used block groups as
the source zone and Gwinnett County voting districts ( n = 132)
as the target zone. We selected voting districts because they are
completely contained within the Gwinnett County boundary, and
Census blocks do not nest within the voting district boundaries
in all cases, therefore offering an applied example of how areal
interpolation can be useful. The areal reaggregation of the data
into the target zones was relatively straightforward: we used the
Zonal Statistic tool to compute data values (total population,
and individuals in poverty) at the voting district levels. The
areally-interpolated Gwinnett County voting district estimates
for the total population, and the population living in poverty
are depicted in Fig. 14.6(a, b). We also selected ZIP codes as
our target zone, because they are not completely contained
within the county boundary. By using areal interpolation we
obtained estimates for the portion of the ZIP codes that were
contained within Gwinnett County. The results are depicted
in Fig. 14.6(c, d) for total population and population living
in poverty.
For dasymetric mapping and areal interpolation using these
four implementations, we observed that accuracies depended
on the classification of our ancillary layers. When using NLCD
2001, it is very difficult to differentiate high-density office build-
ings, high-rise residential buildings, two-story office buildings,
small strip malls, residential apartments, and townhouses. When
using the LandPro 2001 data, institutional populations (e.g.,
university housing and correctional facilities) may get eliminated
if these areas are classified as nonresidential areas. It may be
necessary to obtain the location of such institutions and to man-
ually reclassify as residential. We also observed that the resulting
accuracies of the four implementations did not vary greatly. We
obtained slightly better RMS errors with a binary dasymetric
approach, using custom-written geoprocessing algorithms writ-
ten by one of the authors. These algorithms (Fig. 14.7), although
straightforward, required time to conceptualize and to imple-
ment in ArcGIS Model Builder. However, the geoprocessing
algorithms can be saved and reused for subsequent applications.
Therefore, researchers who anticipate the need for repeated dasy-
metric mapping may wish to consider such an approach. This
approach should be weighed against the ease of use of a cus-
tom dasymetric mapping tool such as the Dasymetric Mapping
Extension, produced and distributed (free of charge) by the
USGS.
We chose not to use more-complex dasymetric mapping
strategies, such as the one outlined by Mennis and Hultgren
(2006), simply to illustrate the ease of implementation of methods
that have been shown to be accurate and produce useful results.
Indeed, Langford and Higgs (2006) note that despite recent
efforts to refine dasymetric mapping with three-tier density
classifications, ''there is little evidence to date of any clear benefit
over the simpler two-tier, binary dasymetric method'' (Langford
and Higgs, pp. 297 - 8). We also support Langford's (2007)
argument that ''there is little evidence to suggest widespread

low-density residential area (NLCD class: Developed Low
Intensity [22]), 1 = high-density residential area (NLCD class:
Developed Medium Intensity [23]), (Developed Open Space
[21] and Developed High Intensity [24] may be appropriate to
include as residential area based on the particular study area) and
0 = non-residential area (all other NLCD classes). We used the
same block group layer as before for the 2000 population. We
initiated and ran the Dasymetric Mapping Extension, using the
default of 80% for the minimum threshold value of the source
unit that must be covered by an ancillary class. The population
density estimates that we derived using this tool are depicted in
Fig. 14.3d.
We compared our dasymetrically-derived population esti-
mates to the 2000 population data for Gwinnett County Census
blocks ( n = 5048). Because we used population data from the
block-group level for the dasymetric process, and because cen-
sus blocks nest completely within block groups, we can use the
block-level data to validate the block-group level estimates. We
used the Zonal Statistic tool to calculate block level population
from the four dasymetric population raster maps. Table 14.1 pro-
vides a comparison of the root mean square error (RMSE) - both
standard and normalized - from the four dasymetric implemen-
tations. The lowest RMSE was obtained by using the binary
method in conjunction with the LandPro 2001 ancillary data,
closely followed by the binary method using the parcel data. The
RMSE for the two methods (binary and N-class) that used the
NLCD 2001 data were close to one another and both higher than
the error associated with the LandPro and parcel data implemen-
tations. The normalized RMSE (essentially the RMSE divided by
the range of the observed population counts at the block level)
provides more of a relative measure of error (express in percent-
ages). We could not compute mean percentage errors because
1209 census blocks contained no population, and thus division
by zero would have been required in the calculation.
Figure 14.4 provides a map for the difference of census
2000 population and area interpolation estimates at block group
level.

=

TABLE 14.1 Comparison of root mean squared (RMS) errors for
four dasymetric mapping implementations for the Gwinnett County
study area.

Dasymetric Ancillary Root Normalized

method data means square root mean

error square

(population) error (%)

Binary NLCD 2001 115.3 2.01

Binary LandPro 2001 102.2 1.78

Binary Parcel 2001 105.5 1.84

N-Class with
D-M Extension

NLCD 2001 117.3 2.04