Characterizing urban subpixel composition using spectral mixture analysis - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 107

Step 1: Build endmember library Step 3: Select the “best” model

for each pixel

M4 M8 M15 M1 M4 M3

M4 M3 M1 M9 M10 M3

M12 M18 M5 M1 M4 M9

M5 M12 M15 M15 M6 M14

M9 M7 M3 M13 M6 M14

Wavelength

Step 2: Test simple SMA models Step 4: Map V-I-S fractions

emb1 + shade

Vegetation Shade

2- emb

emb2 shade

+

Impervious RMS error

emb1 + emb2 + shade

3- emb

emb1 + emb3 + shade

Model
Complexity

Soil

emb1 + emb2 + emb3 + shade

emb1 + emb2 + emb4 + shade

Step 5:

Assess accuracy of fraction images

etc ....

No Accuracy

acceptable
?

Yes

Revisit Steps 1, 2, 3 END

FIGURE 8.3 Overview of the MESMA technique to map urban V-I-S components (adapted from Rashed et al ., 2003).

an imaging spectrometer; these are known as ''reference end-
members.'' Additionally, ''image endmembers'' may be derived
from the image itself or from other images. Reference endmem-
bers have the advantage that they are collected under controlled
conditions and that each spectrum is associated with a specific,
known material. Because they are collected in reflectance mode,
they are easily transported between image platforms and across
dates (Adams et al ., 1995; Roberts et al ., 1998a). In contrast, image
endmembers have the advantage that no additional data or field
campaigns are required and that spectra are collected at the same
scale as the data (Roberts et al ., 1998a; Rashed et al ., 2003). While
image endmembers are not as readily portable across imaging
platforms, they can be compared across dates and across adjacent

scenes if they are radiometrically intercalibrated (e.g., Roberts
et al ., 1998a). A spectral library may include both reference
and image endmembers, as long as all endmembers in the library
have been radiometrically intercalibrated (e.g., Powell et al ., 2007;
Powell and Roberts, 2008).
Endmembers selected for a spectral library should be spec-
trally ''pure,'' though the definition of purity will depend on
the scale of analysis (Smith et al ., 1990). An endmember should
also be representative of other spectra in its own material class
(Song, 2005). Ideally, an endmember will also be spectrally dis-
tinct from spectra in other material classes, though this may be
more or less true, as the primary goal of SMA is to model mate-
rials of interest, not to identify spectra that are mathematically