Characterizing urban subpixel composition using spectral mixture analysis - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 116

constraints imposed below. Only three- and four-endmember
models were tested, because visual inspection of the high spatial
resolution aerial photographs indicated that very few Landsat
pixels in the area of interest had fewer than three basic spectral
components (i.e., every 30-m pixel was expected to contain
at least two different material types and shade). Non-shade
endmember fractions were constrained between − 0.05 and 1.05,
a standard choice in the literature to account for sensor noise and
measurement uncertainty (e.g., Dennison and Roberts, 2003).
The maximum shade fraction allowed was 0.50 (water bodies
therefore remained unmodeled). Only models with an RMS
error less than 6.24 DNs (equivalent to approximately 2.5%
reflectance) were considered.
Model selection rules : For each level of complexity (three-
or four-endmember), the model that met all the constraints
and had the lowest RMS error was selected. Three-endmember
models were selected only when no four-endmember models
met the criteria; i.e., more complex models were favored over
simpler models. This decision was based on evaluation of the
average spectral complexity of this urban landscape by visually
inspecting the aerial photographs overlaid with a 30-m grid, as
well as testing different model selection rules and evaluating the
resulting fraction and RMS images.
Results : Shade normalized fractions for each date are presented
as RGB (NPV, GV, Soil) composites in Fig. 8.7(a) (April, leaf-
off image) and Fig. 8.7(b) (August, leaf-on image). Because
impervious fractions are not included in the RGB display of
the image, variations in brightness are related to the presence
or absence of impervious surfaces (i.e., dark areas have high
impervious fractions, and bright areas have low impervious
fractions). It should be noted that the relative abundance of
soil in the April image most likely indicates confusion between
soil and bright impervious spectra. Because the purpose of
this analysis was solely to assess abundance of vegetation, no
attempt was made to further refine the soil and impervious
endmembers.
Another way to display fraction images that facilitates visual
interpretation, especially to assess change between two images, is
tobin the fraction values intomeaningful categories. For example,
green vegetation fractions have been binned into categories in
Fig. 8.7c (April, leaf-off image) and Fig. 8.7d (August, leaf-
on image), where values of the bins were selected to highlight
change between the two dates and to enable direct comparison
of vegetation fractions.
Estimation of percent tree cover and percent grass cover was
based on the simple assumption that the green vegetation mea-
sured in April predominantly corresponded to grass cover, and
the green vegetation that measured in August was a combination
of tree and grass cover. Therefore, the crudest estimate of tree
cover would be to subtract the estimated April GV fraction from
the estimated August GV fraction. However, pixel-to-pixel com-
parisons can be problematic for several reasons, as discussed in
Section 8.2.6. In addition, while the reference dataset was derived
from very fine resolution imagery, it was only reported at the
neighborhood level; as a result, preliminary accuracy assessment
of theMESMA-generated estimates of tree coverwas conducted at
the neighborhood scale of analysis. To estimate percent tree cover
at the neighborhood level, the mean values of shade-normalized,
green vegetation cover were calculated for each neighborhood for
each date. Percent tree cover was calculated as percent green veg-
etation cover estimated from the August image, minus percent
grass cover estimated from the April image.

The estimates of zonal averages of percent tree were compared
to those derived from the Quickbird imagery for 47 neighbor-
hoods in South Denver. Modeled estimates of tree cover agreed
with reference measures within ± 5% for 35 neighborhoods, and
within ± 10% for 44 neighborhoods. Only three neighborhood
estimates disagreed by 10% or more. A scatterplot of modeled
versus reference tree-cover fractions by neighborhood is pre-
sented in Fig. 8.8a, with the 1:1 line plotted for reference. The
Pearson's correlation coefficient between the two datasets is 0.78,
while the slope of the best-fit line is 0.641. While ideally the slope
of the best-fit line would be closer to one, the MAE was 3.4 and
the bias was 0.7, indicating a high average agreement between

modeled and reference fraction when aggregated to the neigh-
borhood level. Additional information is provided by plotting
the residuals (modeled - observed fractions) as a function of the
reference fractions; the best-fit line is superimposed to indicate
general trends of over- and under-estimation (Fig. 8.8b). This
plot indicates that tree fractions tend to be over-estimated by
MESMA for very small fractions (positive residuals) and under-
estimated for larger fractions (negative residuals). While some of
the disagreement is certainly due to oversimplified assumptions
used by the analysis as discussed below, a portion of the disagree-
ment could also be due to the discrepancy between the dates of
reference data collection (2005 and 2006) and image collection
(2002 and 2003).
This study involved several obvious oversimplifications of
urban vegetation, ignoring potentially confounding factors, such
as grass cover that was obscured by tree canopy in summer
imagery and the presence of coniferous tree canopy in the early
spring imagery. In addition, accuracy assessment to date has
only included tree canopy cover and neglected assessment of the
grass cover estimates. Despite these limitations, estimates of tree
cover aggregated to a neighborhood level were within ∼ 10%
agreement of estimates derived from very high spatial resolu-
tion imagery, indicating the potential of MESMA to provide a
relatively quick and effective method for discriminating the two
plant functional types in urban environments in the temperate
West of North America. Because this methodology is concep-
tually simple and can be applied to data that are regularly and
freely available, assessments of urban vegetation cover can be
applied on a regular basis with very low cost. Such analysis can
support urban resource managers, as well as researchers inves-
tigating relationships between urban vegetation and local and
regional environmental processes. The ''synoptic view'' offered
by moderate resolution remote sensing data can complement
analyses of finer spatial resolution imagery and in situ measure-
ments that cannot be supported as frequently because of high
cost (Small, 2001).

−

Conclusions

This chapter has summarized the implementation of MESMA
to quantify the V-I-S components of urban land cover. The
analysis involves compiling a regional library of the spectra that
best represent the diversity of urban materials. The spectral
response of each pixel in the urban landscape is ''unmixed'' to
determine the fraction of each material component present. The
number and type of materials can vary on a per-pixel basis to
accommodate the high spectral and spatial variability of urban
land cover. The sub-pixel fractions of specific materials are