Characterizing urban subpixel composition using spectral mixture analysis - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 111

8.3.1 Evolution of urbanmorphology
on a tropical forest frontier

TABLE 8.1 Allowed models by generalized material class.

2-emb (15) 3-emb (27) 4-emb (108)

Urban morphology (i.e., the structure and composition of a
city) is the physical manifestation of complex social and eco-
nomic forces, in the context of a specific natural environment
(Moudon, 1997; Rashed et al ., 2003, 2005). A first step in
understanding connections between urban landscape change
and socio-economic drivers is to characterize urban land cover
in terms of the composition and distribution of its bio-physical
components (Rashed et al ., 2005; Powell and Roberts, 2008). The
objective of this case study was to characterize trajectories of
intraurban and periurban landscapes on a tropical forest frontier
in terms of V-I-S components. This study demonstrates for the
first time the possibility of applying the same endmember library
and same model rules across multiple scenes and through time,
implying that modeled fractions are directly comparable from
one date to another. Additionally, this case study illustrates how
assumptions concerning the unidirectional trajectory of urban
change through time can inform spectral library construction
and model selection.
Data: The full analysis involved ten settlements in the state
of Rondonia, in the Southwest Brazilian Amazon, a location
of government-directed settlement programs initiated in the
early 1970s (Fig. 8.1a). The region is a unique environment
to study urbanizing landscapes because the history of urban
development essentially coincides with the Landsat data archive.
The dominant land-cover types in this region are primary forest
(dark green regions in the false color display of the Landsat
sub-image, Fig. 8.4a), pasture (pink areas), and second-growth
forest (brighter green areas). The region of analysis was covered
by five Landsat TM/ETM + scenes. Four dates were analyzed
for each scene, corresponding as closely as possible to the years
1985, 1990, 1995, and 2000. Two settlements are highlighted in
this chapter: Buritis (Settlement #1 in Fig. 8.1a) and Seringueiras
(Settlement #2 in Fig. 8.1a).
A single date for each scene was converted to surface
reflectance using a radiative transfer model, and all other dates
were intercalibrated to the reflectance images using relative
radiometric calibration (Roberts et al ., 1998a; Furby and Camp-
bell, 2001). Subsequent data processing occurred in reflectance
space. An area of approximately 20

NPV + shade NPV + GV + shade NPV + Soil + Imp + shade
GV + shade

NPV + Soil + shade GV + Soil + Imp + shade

Soil + shade GV + Soil + shade

Imp + shade

Numbers in parentheses indicate the total number of models generated for
all permutations of endmembers included in the final MESMA library.
NPV = non-photosynthetic vegetation, GV = green vegetation,

Imp = impervious.

spectra that could be visually classified with high confidence as
belonging to one of the four materials of interest.
The final MESMA library was selected in a two-step process.
First, candidate endmembers were ranked in terms of how
representative they were of other endmembers in the library
based on EAR. Second, low-EAR candidate endmembers were
ranked in terms of how well they represented materials in the
image by sequentially applying low-EAR endmembers to four-
endmember models and assessing model fraction accuracy using
reference data. Similarly, the appropriate number of endmembers
for each category was determined by iteratively adding low-EAR
endmembers to theMESMA library and assessing resultingmodel
accuracy based on the visual inspection of fraction and RMS
error images, as well as quantitative comparison of model results
and high-resolution reference data, following the procedure
summarized in Powell and Roberts (2008). The final MESMA
library consisted of 15 endmembers, in addition to photometric
shade: three GV, three NPV, six impervious surfaces, and three
soil (Fig. 8.3, Step 1).
Allowed models: After experimentation with different end-
member combinations concurrent with inspection of the high
spatial-resolution reference data, allowed models were specified
as those combinations presented in Table 8.1. All permutations
of two-, three-, and four-endmember models were tested for
each pixel. Inspection of reference data indicated that four-
endmember models were sufficient to capture the spectral
complexity of this landscape; therefore, five-endmember mod-
els were not tested. Impervious material spectra were excluded
from three-endmember models, because inspection of the high
spatial-resolution reference data revealed that 30-m samples of
built-up land cover in this region consisted almost exclusively
of three different material types, in addition to shade (e.g., see
Fig. 8.2a); the only exceptions were major roads, which were
sometimes modeled as a two-endmember model of impervi-
ous and shade. The tightness of model constraints minimized
confusion between two-endmember models. When impervious
materials were included in 3-endmember models, bright imper-
vious surfaces were confused with soil and/or NPV. By only
including impervious surface endmembers in the highest com-
plexity models, impervious fractions were restricted to urban
land cover; non-urban land-cover types were spectrally less com-
plex and therefore could be better modeled with simpler models
(i.e., 2- or 3-endmember models).
Model constraints: Candidate models were selected based on
the following criteria: (a) onlymodels withRMS error below2.5%

20 km centered on each
settlement in the study was clipped for analysis. A water-burn
mask was generated for each sub-scene and each date by applying
a threshold to band 7, and those pixels were excluded from fur-
ther analysis. Validation data were derived from high-resolution
aerial videography collected in 1999, which transected three
settlements in the study area (Hess et al ., 2002; Powell and
Roberts, 2008).
Endmember library: Endmembers were selected to represent
each V-I-S category. Vegetation was subdivided into green veg-
etation (GV) and non-photosynthetic (i.e., senesced) vegetation
(NPV), because these seasonal states of vegetation have distinctly
different spectral properties (e.g., Roberts et al ., 1998a). Candi-
date endmembers were collected froma combination of reference
and image endmembers; the former were selected from a spec-
tral library that included laboratory- and field-collected spectra.
Candidate image endmembers were identified by applying the
PPI algorithm to all ten samples from the year 2000 and selecting

×