Environmental Engineering Reference
In-Depth Information
Impervious
Soil
Vegetation
10
25
100
(a)
8
20
15
90
6
4
80
10
70
2
0
5
0
60
1985
1990
1995
2000
1985
1990
1995
2000
1985
1990
1995
2000
10
8
25
20
100
(b)
80
6
4
15
60
10
5
40
20
2
0
0
0
1985
1990
1995
2000
1985
1990
1995
2000
1985
1990
1995
2000
FIGURE 8.6
Generalized urban growth trajectories represented by percent area of 2000 urban extent covered by each V-I-S
component for (a) Seringueiras and (b) Buritis.
strategy to separate materials that are spectrally similar is to inte-
grate temporal information in the analysis. This strategy has been
most commonly applied to traditional land-cover classification;
for example, Yuan
et al
. (2005) distinguished between forests and
agriculture in a peri-urban environment using two images col-
lected at different times in the crop cycle - the first in late spring
before crop green-up and the second in late summer shortly
before harvest. Kuemmerle, Roder and Hill, (2006) separated
percent herbaceous and percent woody cover in a Mediterranean
shrubland by comparing SMA results applied to two dates - the
first image corresponded to the end of the rainy season when all
vegetation was green, and the second image corresponded to the
end of the dry season when grasses were senesced while shrubs
remained green.
A similar methodology was tested here to separate the lawn
and tree components of urban vegetation. MESMAwas applied to
two dates in the growing season to separate vegetation fractional
cover fromother materials in the urban environment. Vegetation
types were separated based on two simple assumptions concern-
ing the phenology (i.e., seasonal cycle) of urban vegetation in
this climate: (a) grass cover can be quantified in early spring,
leaf-off conditions, because deciduous trees and shrubs do not
leaf out until approximately 6-8 weeks after lawns green up, and
(b) tree canopy can be estimated by comparing green vegetation
measured in late summer (i.e., full-canopy conditions) to green
vegetation measured in early spring.
Data
:TwoLandsatETM
+
images (P33/R33) were acquired
to compare different seasonal states of urban vegetation - the
first from early spring, leaf-off conditions (16 April 2003), and
the second from late summer, full-canopy conditions (18 August
2002). Only the southern half of Denver was included for this
pilot study because the city is located at the boundary of two
Landsat scenes. Aerial photographs collected in 2005 by the
National Agriculture Imagery Program (NAIP) were used to
guide endmember selection, SMA model specification, and SMA
model constraints and selection rules. The reference dataset for
accuracy assessment was based on a fine-resolution inventory
of tree-canopy cover for the city and county of Denver, gen-
erated in 2006 using object-oriented analysis of pan-sharpened
QuickBird imagery (approximately 61-cm spatial resolution);
however, the publically available estimates of tree-canopy cover
were aggregated to the neighborhood level (NCDC, 2006).
Endmember library
: Image endmembers were selected from
the 2003 Landsat image (leaf-off conditions); candidate end-
members were identified by applying PPI to the image subset
and verifying the material composition of selected pixels using
the aerial photographs. To determine which candidate end-
members were most representative of materials in the image,
two-endmember models (i.e., bright endmember
+
shade) were
tested using each candidate endmember to unmix the scene, and
only those endmembers that modeled at least a portion of the
scene well were selected for the final MESMA library. Library
spectra were radiometrically calibrated to the 2002 image using
relative radiometric calibration (Roberts
et al
., 1998a; Furby and
Cambell, 2001). The same endmember library was applied to
both dates, and all MESMA processing occurred in DN-space.
The final spectral library consisted of nine endmembers, in addi-
tion to non-zero shade: one green vegetation endmember, two
non-photosynthetic vegetation, two soil, two dark impervious
endmembers, and two bright impervious endmembers. A proxy
for shade was selected from a deep water body. The selection of a
single GV endmember for the MESMA library may seem incon-
sistent, as the goal of this study was to discriminate between two
different types of vegetation. However, while the mixture of basic
spectral components (e.g., the proportion of GV, NPV, Shade)
was expected to vary between grass cover and tree canopy cover,
the basic spectral component representing green vegetation was
assumed to be the same for both vegetation types, especially given
the coarse spectral resolution of Landsat data (Smith
et al
., 1990).
Allowed models and model constraints
: All models included
the green vegetation endmember; any pixel that did not include
green vegetation was presumed to remain unmodeled given the
Search WWH ::
Custom Search