PE&RS February 2018 Full - page 105

Mapping Urban Change
Once the road removal and clean-up of the 2001 urban
footprint were complete (Figure 2b and 2c), we evaluated the
developed pixels added between NLCD 2001 and NLCD 2006.
Only conversions “to developed” were considered in this
step, since we assume changes in developed lands are unidi-
rectional. New roads that were built between 2001 and 2006
were manually deleted, along with oil and gas drilling pads
classified as developed. The process was repeated for NLCD
developed pixels added between 2006 and 2011. New roads
that were built between 2006 and 2011 were deleted, as well
as oil and gas drilling pads classified as developed. Different
procedures were used to create the 1992 urban footprint. A
composite of the disaggregated 2001 developed layer, NLCD
1992 developed (Vogelmann et al., 2001), NLCD Retrofit 1992
developed (Fry et al., 2009), and NWALT 1992 (Falcone, 2015)
was created by intersecting all layers and comparing pixels
mapped in all 1992 products with the developed pixels in
the 2001 product. Urban areas were back-classified to 1992
by retaining only those pixels from NLCD 1992, NLCD Retrofit
1992, or NWALT 1992 that were located within the 2001 urban
footprint. This process ensured that development followed a
unidirectional conversion process.
The urban (see Supplementary Material Section) and urban
change maps were summarized nationally for illustrative
purposes. Regional results were calculated by intersecting
the original NLCD maps and the final urban maps with Level
III EPA ecoregions (Omernik, 1987). Comparative maps and
statistics were compiled to highlight key differences between
each set of products.
2001 and 2006 Accuracy Assessment
The accuracy of the final versions of the 2001 and 2006
maps of the binary classification “urban” and “not urban”
was evaluated where rural roads were included in the “not
urban” class. The sample selected for the accuracy assessment
of the 2006 NLCD products (Wickham et al., 2013) provided
the sample locations for assessing the accuracy of the urban
maps. The NLCD sampling design was stratified geographically
by ten regions of the United States, and within each geograph-
ic stratum further stratified by land-cover change. For each
sampled pixel, reference labels were obtained for 2001 and
2006 (the interpreter did not know the map class label). The
original NLCD reference classification provided from Wickham
et al. (2013) for each of the 15,000 sample pixels included a
primary and alternate classification for each sample pixel.
Because “urban” is not one of the NLCD land-cover classes, the
NLCD sample pixels had to be re-interpreted to assess accuracy
of the urban class. Given that the urban class is a subset of the
NLCD developed class, the re-interpretation of the reference
class labeling could be limited to only those sample pixels
that had a primary or alternate label of “developed” in the
NLCD reference sample database.
For the 2001 reference classification we used circa-2001
Landsat imagery and 1998 to 2001 aerial imagery to deter-
mine if the pixel should be re-labeled from “developed” to
“not urban” if the pixel was a rural road or fell more than 500
meters outside of a designated urban area (US Census Bureau,
2016b). The 2006 reference sample pixels were similarly
re-examined using 2005 Landsat imagery and 2003 to 2007
aerial imagery to re-label sample pixels from “developed” to
“not urban” if the pixel was a rural road or was located well
outside of an urban area. Through this procedure we effective-
ly revised the definition of the remaining developed reference
sample pixels to correspond to our definition of “urban”. Our
reference data interpretation protocol mirrors that of the NLCD
accuracy assessment in that a 3-by-3 pixel window was ap-
plied to provide context for the class labeling (Wickham et al.,
2010). Accuracy estimates and standard errors were produced
using the same formulas reported by Wickham et al. (2013).
Results
Comparative Assessment of NLCD and Urban Maps
In the NLCD products for 2001, 2006, and 2011, the developed
class totaled 434,990 km2, 446,660 km2, and 454,290 km2,
respectively. The mapped developed area represents roughly
6 percent of the land area in the conterminous United States.
The change in map area of the NLCD developed class was
12,670 km2 for 2001 to 2006 and 7,630 km2 for 2006 to 2011.
The process of creating urban maps by removing rural roads
from the NLCD developed class resulted in the deletion of
230,360 km2, 230,670 km2, and 230,880 km2 from the NLCD
developed class area in 2001, 2006, and 2011, respectively.
The resulting urban maps included far less area of urban land,
totaling 173,580 km2 in 1992, 203,640 km2 in 2001, 215,990
km2 in 2006, and 223,410 km2 in 2011, or about 3 percent of
the land area in the conterminous United States (see Supple-
mentary Material). The annual rate of change in urban area
decreased over time. Specifically, the change in mapped area
of urban land was 3,339 km2 per year for 1992 to 2001 (30,050
km2 total), 2,472 km 2 per year for 2001 to 2006 (12,360 km2 to-
tal), and 1,484 km2 per year for 2006 to 2011 (7,420 km2 total).
Within Level III Ecoregions the total area of NLCD devel-
oped land differs substantially from the total area of urban
land (Figure 5; Table 2). For example, many of the ecoregions
with the largest developed footprint in NLCD 2011 also contain
a sizable urban footprint in the 2011 urban map. The top ten
ecoregions of total NLCD developed area and the top ten of
total urban area include ecoregions in the eastern and western
United States with major urban centers (Table 2): Ecoregion 75
includes major urban centers throughout northern and central
Florida including Jacksonville, Orlando, and St. Petersburg,
Ecoregion 45 includes Atlanta (Georgia), Charlotte (North
Carolina) and Raleigh-Durham (North Carolina), Ecoregion 6
includes the San Francisco-Bay Area and Los Angeles-San Di-
ego Area (California), and Ecoregion 67, which is the second
largest ecoregion in the United States, includes Mobile (Ala-
bama), Tallahassee (Florida), Fayetteville (North Carolina),
Richmond (Virginia), and substantial portions of Baltimore
(Maryland) and Washington (District of Columbia). Although
total areas and rankings differ between the two maps, seven
of the ecoregions with the largest developed footprint in
NLCD 2011 are also included in the urban map list. The most
notable differences between NLCD and the urban maps are that
Ecoregions 47 (Western Corn Belt), 27 (Central Great Plains),
and 25 (Western High Plains) are no longer among the top ten.
The smaller urban land area relative to the developed land
area across these ecoregions is the result of removal of rural
roads during editing. In the ecoregions removed from the top
10 of total developed area, the only major urban centers are
Omaha (Nebraska) and Des Moines (Iowa) in Ecoregion 47,
Wichita (Kansas) and Oklahoma City (Oklahoma) in Ecoregion
27, and Denver (Colorado) in Ecoregion 25.
2001 and 2006 Accuracy Assessment Results
Both the urban and NLCD maps for 2001 and 2006 exceeded
the overall accuracy benchmark of 85 percent (Anderson,
1976) for the binary classification of urban and not urban. The
overall accuracy of the urban classification improved follow-
ing removal of rural roads from the NLCD developed class,
primarily due to the substantial reduction in urban commis-
sion errors (Table 3). Over 230,000 km2 of area formerly classi-
fied as developed by the NLCD was removed to create the new
urban maps resulting in an increase in overall accuracy of 2.5
percent and a reduction in urban commission error by over
30 percent relative to the NLCD developed class. The improve-
ment in urban commission error was achieved at the small
expense of increasing the urban omission error by 4 percent
relative to the NLCD developed class (Table 3).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING February 2018 105