3FR;01FEB99


SOCIAL SCIENCE COMPUTER REVIEW
Schweik, Green / SPECTRAL MIXTURE ANALYSIS

The Use of Spectral Mixture Analysis to
Study Human Incentives, Actions, and
Environmental Outcomes
CHARLES M. SCHWEIK

Indiana University
University of Massachusetts

GLEN M. GREEN

Indiana University

The social science community has given increasing attention to the analysis of multitemporal satellite
images in human dimensions of environmental change research. This article provides an overview of
how image-processing techniques, such as radiometric calibration and spectral mixture analysis, can
be applied. The application of these techniques is vital if the social science community wishes to
develop a robust research program that allows the accurate comparison of distant but similar locations.
Furthermore, by mapping the institutional landscape, the article demonstrates how social scientists can
begin to understand landscape change by carefully relating human incentives to actions and actions to
outcomes as measured by a set of multitemporal satellite images.

Keywords: global change, satellite images, institutional analysis, spectral mixture
analysis, image restoration, institutional landscape

INTRODUCTION

Environmental policy makers, public officials involved in land management, and
researchers studying the human dimensions of environmental change all hope to answer a
particular question, each from their particular perspectives: In what ways is the landscape
changing, and how are humans influencing those changes? Robust answers to this question
are critical if we are to (a) understand the broad role humans play in the realm of global
change as well as (b) guide officials at local and regional levels who are making real-life land
use planning decisions and prioritizing the use of scarce public funds in land management.

The National Aeronautics and Space Administration’s Landsat, or Land Monitoring Sat-
ellite system, has amassed an immense longitudinal archive of multispectral images for
every location in the United States since its initiation in 1972. These images provide a wealth
of information characterizing the country’s forest resources as well as identifying where land
cover is undergoing change. However, monitoring land cover using Landsat images has only
just recently begun to significantly penetrate disciplines beyond the physical sciences. Most
satellite-based case studies involve land-cover inventories, or classifications, at one point in

AUTHORS’ NOTE: Support from the National Science Foundation (Grant No. SBR-952-1918) is gratefully
acknowledged. We are grateful for the exceptional editing skills provided by Patty Dalecki at the Workshop in
Political Theory and Policy Analysis, Indiana University, Bloomington.

Social Science Computer Review, Vol. 17 No. 1, Spring 1999 40-63
© 1999 Sage Publications, Inc.

40



time where the satellite data associated with each picture element, or pixel, of the image is
assigned to one land-cover class. Programs such as the National Wetlands Inventory and
state-level gap analysis provide examples of such inventory programs (IGAP, 1998; Scott et
al., 1993). These studies typically employ classification schemes based on variations in
satellite-derived digital numbers, or DN values, which, unlike reflectance values, are not
physical units and thus have limited utility for comparison outside the specific image. There-
fore, a given classification from one image cannot be easily applied to another image at a dif-
ferent location or time.

Recent technological advances now enable researchers, natural resource policy makers,
and land managers to move beyond these traditional satellite image-based cross-sectional
land inventories. Gigabyte storage volumes, Pentium-level or higher processors, cheaper
random access memory, and more accessible geographic information system (GIS) and
remote-sensing software allow even users of desktop PC platforms to conduct satellite-based
image change analysis. Moreover, by first calibrating the image data so that picture element
(pixel) values represent surface reflectance and then applying a relatively new remote-
sensing analysis technique called spectral mixture analysis (SMA), the policy analyst or
social scientist can move beyond a statistically relative and somewhat abstract land classifi-
cation to one based on the physical measures of surface reflectance. Furthermore, SMA pro-
vides the ability to identify subpixel measures such as the percentage of land-cover compo-
nents (e.g., soil, vegetation, water) that make up a pixel.

These advances are important for two reasons: First, SMA’s subpixel analysis allows the
researcher to use more effectively the older Landsat Multispectral Scanner data with coarse
spatial resolution (59 × 79 square meters, roughly the size of a U.S. football field). Pixels of
this size will rarely be exclusively one land-cover type and are more likely to be composed of
mixtures of different types of land cover. Subpixel information is vital when working at this
spatial resolution. Second, by radiometrically calibrating images to surface reflectance val-
ues and applying SMA, the analyst can develop land-cover classification schemes that are
more apt to be replicated and directly applied to other Landsat images of comparable land-
cover regions or ecosystems. Such broad-based classifications can be used to develop land-
scape change maps, which in turn can be linked with other geographic parameters known to
drive human decision making and action. Efforts such as these are critical as we continue to
address issues of scaling in global change research as well as make more practical policy and
land management decisions.

This article provides an overview on how SMA might be applied to understanding land-
cover change in a forested region and how this change can be related to human activities. The
first part provides a discussion on the utility of the Landsat archive for this research agenda.
Part 2 raises some open questions and discusses in more detail the problems related to stan-
dard image classification procedures as they are currently used. Part 3 presents a more tech-
nical discussion on traditional analytic techniques and then provides a description of how
SMA works. Part 4 presents a brief example of one recent application of SMA in the context
of public forest management in Indiana. It shows how we can begin to develop longitudinal
studies of land-cover change based on surface reflectance values that can be replicated by
other researchers using other Landsat images in comparable locations. It also presents two
examples of how these analyses can be linked to human incentives and actions. The article
concludes with a summary of the importance of this for the human dimensions of environ-
mental change and global change research programs as well as for policy makers and land
managers.

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 41



THE UTILITY OF LANDSAT MSS AND TM IMAGES FOR
THE STUDY OF THE HUMAN DIMENSIONS
OF ENVIRONMENTAL CHANGE

The Landsat system, designed in the 1960s and first launched in the early 1970s, was
specifically developed for the observation of Earth’s terrestrial surface (Campbell, 1996,
p. 158). Tables 1 and 2 provide a summary of the Landsat system. Since 1972, six Landsat
satellites or platforms have been launched. Landsat 5 continues to be operational, working
well beyond its intended life span. Two primary sensor instruments exist on these platforms:
the Multispectral Scanner (MSS) and Thematic Mapper (TM). MSS instruments were
employed on Landsats 1 through 5. TM instruments were first included on Landsat 4, remain
on Landsat 5, and an enhanced version will be launched on Landsat 7. There are several justi-
fications for the use of MSS and TM imagery in studies concerning the human dimensions of
environmental change, and many of these reasons are directly associated with scale issues
discussed above.

First, both the MSS and TM data archives cover a broad spatial extent. Each individual
scene acquired by both Landsat sensors covers a “footprint” on the Earth approximately 185
kilometers wide. Furthermore, since Landsat 1’s launch in 1972, all of the Earth’s terrestrial
surface between 81º N and 81º S latitude has been subject to image acquisition (Campbell,
1996, p. 162).

Second, the Landsat data archive covers a relatively long temporal duration (more than 25
years). MSS technology began in 1972; TM technology began 10 years later in 1982.
Although this cumulative period may be short in terms of the history of humanity, the time
period covers a long enough temporal range to capture many of the human-induced changes
that have occurred in forested landscapes. This is especially true in locations where land-
cover change is rapidly occurring at present, such as in developing countries where tropical
forests are being converted to agriculture or grazing land.

Third, the MSS and Landsat TM instruments provide a reasonably high degree of spectral
resolution when compared to other remote-sensing platforms such as aerial photographs,
France’s SPOT, or India’s IRS satellites. The MSS has four sensors, or bands, each sensitive
to a different portion of the electromagnetic spectrum. As Table 2 shows, MSS Band 1
responds to light at visible green wavelengths (0.5-0.6 micrometers, or µm), Band 2
responds to light in the visible red portion of the spectrum (0.6-0.7 µm), and both Band 3 and
Band 4 to different portions of near-infrared wavelengths (0.7-0.8 µm and 0.8-1.1 µm,
respectively). Landsat TM data provide better spectral resolution, employing seven bands
with sensors collecting additional data from the mid-infrared and thermal regions of the elec-
tromagnetic spectrum.

Although other technologies such as the National Oceanic and Atmospheric Administra-
tion’s Advanced High-Resolution Radiometer or hyperspectral instruments such as the Air-
borne Visible/Infrared Imaging Spectrometer provide even better spectral resolution than
Landsat data do (Campbell, 1996; Verbyla, 1995), the MSS instrument probably provides
adequate data for addressing many of the questions being asked in the human dimensions of
global change research program. Many of the activities humans undertake on the physical
landscape are dramatic, involving vegetation-to-soil or soil-to-vegetation conversions.
These types of changes provide a clear contrast in land-cover reflectance spectra (see Figure 1)
that can be adequately captured using MSS band passes and most certainly with TM spectral
resolution.

Fourth, the temporal sampling of the Landsat MSS and TM systems are relatively high.
Location coverage repetition frequency was 18 days for Landsats 1, 2, and 3 and is 16 days

42 SOCIAL SCIENCE COMPUTER REVIEW



TA
B

L
E

1
L

a
n

d
s

a
t

S
a

te
ll

it
e

In
s

tr
u

m
e

n
t

S
p

e
c

if
ic

a
ti

o
n

s

S
a

te
lli

te
P

la
tfo

rm

L
a

n
d

sa
t

1
:

L
a

n
d

sa
t

2
:

L
a

n
d

sa
t

3
:

L
a

n
d

sa
t

4
:

L
a

n
d

sa
t

5
:

L
a

n
d

sa
t

6
:

L
a

n
d

sa
t

7
:

L
a

u
n

ch
e

d
L

a
u

n
ch

e
d

L
a

u
n

ch
e

d
L

a
u

n
ch

e
d

L
a

u
n

ch
e

d
(f

a
ile

d
to

(e
xp

e
ct

e
d

7
/2

3
/7

2
;

1
/2

2
/7

5
;

3
/5

/7
8

;
7

/1
6

/8
2

3
/1

/8
4

;
re

a
ch

o
rb

it)
la

u
n

ch
d

a
te

,
O

u
t

o
f

S
e

rv
ic

e
O

u
t

o
f

S
e

rv
ic

e
O

u
t

o
f

S
e

rv
ic

e
O

p
e

ra
tio

n
a

l
O

p
e

ra
tio

n
a

l
O

u
t

o
f

S
e

rv
ic

e
m

id
-1

9
9

9
)

In
st

ru
m

e
n

t
1

/6
/7

8
2

/2
5

/8
2

3
/3

1
/8

3
(s

ta
n

d
b
y

m
o

d
e

)

M
u

lti
sp

e
ct

ra
lS

ca
n

n
e

r
(M

S
S

)
X

X
X

X
X

—
T

h
e

m
a

tic
M

a
p

p
e

r
(T

M
)

X
X

—
X

(E
n

h
a

n
ce

d
T

M
+

)

S
O

U
R

C
E

:A
d

a
p

te
d

fr
o

m
C

a
m

p
b

e
ll

(1
9

9
6

).

TA
B

L
E

2
S

u
m

m
a

ry
o

f
In

s
tr

u
m

e
n

t
S

e
n

s
o

rs

In
st

ru
m

e
n

t
B

a
n

d
S

p
e

ct
ra

lS
e

n
si

tiv
ity

(m
ic

ro
m

e
te

rs
)

R
e

so
lu

tio
n

(s
q

u
a

re
m

e
te

rs
)

M
u

lti
sp

e
ct

ra
lS

ca
n

n
e

r
(M

S
S

)a
1

0
.5

-0
.6

(g
re

e
n

)
7

9
×

5
6

2
0

.6
-0

.7
(r

e
d

)
7

9
×

5
6

3
0

.7
-0

.8
(n

e
a

r-
in

fr
a

re
d

)
7

9
×

5
6

4
0

.8
-1

.1
(n

e
a

r-
in

fr
a

re
d

)
7

9
×

5
6

T
h

e
m

a
tic

M
a

p
p

e
r

(T
M

)
1

0
.4

5
-0

.5
2

(b
lu

e
)

2
8

.5
×

2
8

.5
2

0
.5

2
-0

.6
0

(g
re

e
n

)
2

8
.5

×
2

8
.5

3
0

.6
3

-0
.6

9
(r

e
d

)
2

8
.5

×
2

8
.5

4
0

.7
6

-0
.9

0
(n

e
a

r-
in

fr
a

re
d

)
2

8
.5

×
2

8
.5

5
1

.5
5

-1
.7

5
(m

id
-i
n

fr
a

re
d

)
2

8
.5

×
2

8
.5

6
1

0
.4

-1
2

.5
(t

h
e

rm
a

l)
2

8
.5

×
2

8
.5

7
2

.0
8

-2
.3

5
(m

id
-i
n

fr
a

re
d

)
2

8
.5

×
2

8
.5

E
n

h
a

n
ce

d
T

M
+

S
a

m
e

a
s

T
M

,
e
xc

e
p

t:
O

p
tic

a
lb

a
n

d
s

2
8

.5
×

2
8

.5
T

h
e

rm
a

l
6

0
×

6
0

P
a

n
ch

ro
m

a
tic

1
5

×
1

5

a
.I

n
L

a
n

d
sa

t
sa

te
lli

te
p

la
tf
o

rm
s

1
,

2
a

n
d

3
,

M
S

S
b

a
n

d
s

1
,

2
,

3
a

n
d

4
w

e
re

re
fe

rr
e

d
to

a
s

b
a

n
d

s
4

,
5

,
6

,
a

n
d

7
re

sp
e

ct
iv

e
ly

.

43



for Landsats 4 and 5. This revisit frequency is probably adequate for addressing many of the
most important human-induced processes related to landscape change. Although cloud
cover may restrict acquisitions for any given location within these latitudes, there is likely a
large temporal archive of good quality still available for sampling. In our south-central Indi-
ana study region alone, we have identified approximately 150 Landsat MSS and TM images
with low cloud cover taken during the 1972 to 1997 time period.

OPEN QUESTIONS RELATED TO THE UTILITY OF
LANDSAT MSS IMAGES FOR HUMAN DIMENSIONS
OF ENVIRONMENTAL CHANGE STUDIES

As the previous discussion indicates, there are ample reasons to consider the use of Land-
sat MSS and TM for studies relating human decision making and action to land-cover
change. TM images are certainly appropriate given their high spatial and radiometric resolu-
tions. However, as noted above, TM technology was first introduced in 1982, and about 16
years of inventory exists. But if we are concerned about the temporal grain of the phenomena
of interest (such as forest cutting) and are concerned that a 16-year sampling period may not
capture a specific human activity adequately, we may wish to expand the temporal duration
of the image data set we use. By employing Landsat MSS images, we extend the temporal
range by another 10 years back to 1972. For many change analyses, this added decade may be
vital. However, there is one significant limitation in applying MSS data for these types of
studies: the relatively poor spatial resolution provided.

Other satellite platforms described earlier, such as Landsat TM and SPOT, provide rela-
tively fine spatial resolution with 28.5 × 28.5 meter and 10 × 10 meter spatial resolutions,
respectively (Campbell, 1996). An MSS pixel is coarser, capturing roughly a 79 × 56 meter
area on the ground. This coarse spatial resolution limits the amount of information related to
human activities that MSS images can provide (Wilkie & Finn, 1996).

We therefore face a tradeoff. It is clear that for these types of questions, images with a
higher spatial resolution are favorable, such as ones created by Landsat TM or SPOT, but
MSS adds an important 10 years onto the time series. MSS images are also much less expen-
sive than TM images. From this perspective, the use of Landsat MSS is certainly helpful,

44 SOCIAL SCIENCE COMPUTER REVIEW

Figure 1: Typical Continuous Spectral Reflectance Curves for Vegetation, Soil, and Water and a
Hypothetical Mixture Spectra

SOURCE: Adapted from Richards (1986).



economically efficient, and could also be absolutely critical to adequately capture the tempo-
ral frequency of certain land-cover change phenomena.

TRADITIONAL ANALYSIS TECHNIQUES AND THE
PROMISE OF SPECTRAL MIXTURE TECHNIQUES

The use of another technique, first developed by geologists and planetary scientists,
called spectral mixture analysis, or SMA (Adams, Smith, & Gillespie, 1993; Adams, Smith,
& Johnson, 1986; Smith, Ustin, Adams, & Gillespie, 1990a, 1990b), has the potential to
make Landsat MSS-based (and TM-based) analyses more applicable to the study of fine spa-
tial scale human-induced land-cover disturbance. SMA, also referred to as linear unmixing
(ENVI, 1998), allows the analyst to move to a subpixel level of analysis and works with sur-
face reflectance values, not image DNs (see below). The following section provides a brief
summary of the more traditional image-analysis techniques and then describes SMA and
how it might be used to improve the utility of Landsat MSS and TM image analysis for stud-
ies related to human dimensions.

Traditional Image-Analysis Techniques

Image-analysis techniques in this context refer to methods for displaying and interpreting
band-to-band variations of multispectral satellite images. Generally, traditional image
analysis involves a suite of tools including single-band analysis, color composite generation,
band-to-band ratioing and vegetation indices, principal component analysis (PCA), and
classification. These methods are widely used in analyzing Landsat images (Adams et al.,
1993). For a more detailed description of each technique, see any standard remote-sensing
textbook (e.g., Campbell, 1996). Only a brief synopsis will be included here.

Single-band analysis involves the review of individual bands of an image separately,
using a gray-scale display. Color compositing techniques superimpose three bands together,
displaying each band of information using the three colors: blue, green, and red. For
instance, for MSS data, the visible green band is often displayed using blue, the visible red
band is displayed as green, and the second infrared band is displayed as red. This simulates
what would be generated if the image were taken using color infrared film (Lillesand &
Kiefer, 1994). One limitation in using color composites is that only three channels of the
information from a satellite image can be displayed at one time (Adams et al., 1993).

Band-to-band ratioing is another analysis technique that uses multispectral satellite
images for the study of land cover. Vegetation indices are one group of methods that involve
popular band-to-band ratios. Vegetation chlorophyll is  known to absorb visible light,
whereas a leaf’s mesophyll tissue strongly reflects near-infrared light (Campbell, 1996). As
the amount of vegetation increases within the boundaries of a pixel, surface reflectance in the
visible red decreases, whereas reflection in the near-infrared increases (Hall, 1994). Vegeta-
tion indices, such as the normalized difference vegetation index (NDVI), take advantage of
these properties. There are a number of indices that have been developed, but they all involve
mathematical ratios or differences of the digital values of different spectral bands to produce
a single value for each pixel.

NDVI and other such ratios are useful in that they can provide a measure of the vegetation
content within each pixel. By employing the characteristic spectral response of vegetation,
these indices allow the analyst to develop a measure of one component of each pixel. There are
a variety of studies that have used ratios as part of their analyses (e.g., Cohen, 1991; Sader &
Wynne, 1992; Tucker, VanPraet, Sharman, & Van Ittersum, 1985). One of the limitations of

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 45



ratios is that they generally do not take advantage of all the available information contained
in multispectral images, usually using only two bands (Adams et al., 1993).

Also, these indices may be influenced by many factors not associated with vegetation
itself (e.g., soil background and sensor differences) (Campbell, 1996). These ratios are also
sensitive to the particular atmospheric conditions on the day of acquisition because the
atmosphere can cause light in the visible portions of the spectrum to be preferentially scat-
tered. Many analyses may have employed vegetation ratios without removing these effects,
which may lead to false conclusions (Robinove, 1982). Atmospheric correction is advised to
remove the effects of atmospheric scattering (Teillet & Fedosejevs, 1995). Others emphasize
the importance of radiometric calibration, together with atmospheric correction, to convert
the image from raw DN values to surface reflectance before calculating these ratios to mini-
mize sensor drift problems and other factors (e.g., Green, Schweik, & Hanson, 1998; Price,
1987). Unfortunately, the procedures required to remove much of the variance caused by
nonvegetation effects are logistically difficult and not readily accessible to many researchers
outside the remote sensing community. This topic will be discussed more fully later.

PCA is often used to remove interband correlation that typically exists within multispec-
tral image data. PCA identifies linear combinations of the original band data of an image to
produce component images representing the axes of maximum variation (Campbell, 1996).
Often the purpose of PCA is to compress most of the information contained in a multiple-
band image into a new image with fewer components. These product images are then used in
place of the original band data (Lillesand & Kiefer, 1994). This procedure is sometimes help-
ful in preprocessing the data to remove instrument noise such as image striping (Idrissi,
1997). Some scholars are investigating the use of the higher components (e.g., Component 5
and higher in TM images) to identify patterns in human activity (B. L. Turner II, personal
communication, February 22, 1998). One benefit of PCA is that it uses all the information
provided in a multispectral image.

Classification methods are one of the most common analytic techniques applied to
multispectral images (Adams et al., 1995). This becomes quite apparent when one traverses
the remote-sensing literature of the 1970s, 1980s, and 1990s. Most texts related to remote
sensing devote significant effort (sometimes the majority of the discussion of analysis) to
image classification procedures (see, e.g., Campbell, 1996; Lillesand & Kiefer, 1994; Ver-
byla, 1995; Wilkie & Finn, 1996). Using a classification algorithm, these methods assign
individual pixels of a multispectral image to discrete categories (Drury, 1990). The goal of
this type of analysis is to greatly simplify continuous image data (7 bands of 8-bit data for
TM) using quantitative techniques for identification of spectrally similar land-cover classes
within the image (Lillesand & Kiefer, 1994).

Two types of classification procedures are popular and work in conjunction with one
another. Unsupervised classification procedures group or cluster the multispectral values of
the image into distinct classes (e.g., water, soil, and vegetation) based solely on the image
statistics and produce a new raster (grid) map displaying the class designations within the
image (Hall, 1994). Supervised classification involves the use of ground inventories to
guide, or “train,” the computer to assign appropriate clusters of data to certain land-cover
classes (Hall, 1994). Various algorithm strategies (e.g., maximum likelihood) have been
employed to classify pixels not assigned to training areas (Campbell, 1996). Many important
human dimensions of environmental change studies use image classification as their domi-
nant image-analysis method (to name a few: Brondizio, Moran, Mausel, & Wu, 1994; Lee &
Marsh, 1995; Moran, Brondizio, & Mausel, 1994; Sader, 1995; Westman, Strong, & Wilcox,
1989). Others have combined analytic strategies by employing both classification and band-
to-band ratios (e.g., Lee & Marsh, 1995; Wolter, Mladenoff, Host, & Crow, 1995).

46 SOCIAL SCIENCE COMPUTER REVIEW



One great advantage of classification methods is that they use all the information (bands)
provided by multispectral images. However, there are two great limitations of classification
procedures. First, only one land-cover category or class can be assigned to each pixel or cell
of an image matrix. Yet, in many instances, the land represented by one 79 × 56 meter MSS
pixel will be a mixture of multiple land-cover types. Second, most classifications employed
are usually image dependent. Typically they are based on raw DNs and have not been con-
verted to physical measures such as surface reflectance values (see the below discussion on
image restoration). Nonsurface sources of variability, such as sensor, illumination, and
atmospheric differences, remain within the image values and may influence the way the data
are classified. Furthermore, because of scene dependence in classification, an expansion of
geographic coverage within one individual scene may reveal a new land-cover type that may
then alter classification groupings. For these reasons, classifications, based on satellite DNs
without image restoration, may not be transportable to other geographic locations.

Spectral Mixture Analysis

Efforts have been made to address these limitations in classification techniques (e.g.,
Jackson, 1983). SMA is a promising technique developed from the efforts of Earth and plane-
tary scientists (Adams & Adams, 1984; Adams et al., 1986; Bryan, Finger, & Chayes, 1969;
Kornder & Carpenter, 1984; Smith et al., 1990a, 1990b; Ustin, Smith, & Adams, 1993).

Certain land-cover types (e.g., soil, water, and vegetation) have been shown to exhibit
characteristic patterns of reflectance with wavelengths across the electromagnetic spectrum.
Figure 1 shows typical reflectance spectra of general land cover derived from Landsat image
data, sometimes referred to as stick spectra. Some landscapes may exhibit discrete and
homogeneous land-cover types such as a lake. In instances such as these, pixel data may
exhibit more “pure” forms of spectral responses (e.g., 100% water), and classification is an
appropriate analytic technique.

In reality, however, land-cover surfaces are often composed of a variety of mixes of mate-
rials, and the distinction between them (especially from space) is not so clear. When a sensor
observes a pixel composed of mixed types of surfaces, the reflectance spectra produced will
not match any pure spectra like the examples shown in Figure 1. Rather, the pixel spectral
response will exhibit a shape over the electromagnetic wavelength that is some mixture of
spectra of those land-cover types present (Campbell, 1996). For example, a pixel composed
of both soil and vegetation will have a spectral response that is some combination of the gen-
eral soil and vegetation spectra (see the hypothetical mixture example shown in Figure 1).
The objective of SMA is to identify primary spectral contributions within each pixel (Adams
et al., 1993, p. 149). It provides a means for determining relative abundances of land-cover
materials present in any given pixel based on the spectral characteristic of the material
(ENVI, 1996).

SMA involves two main steps. The first step is to define a set of pure spectra for selected
land-cover material, often referred to as endmembers. Endmembers can be identified using
either (a) libraries of known spectra collected with a spectrometer in the field or in a labora-
tory, (b) libraries of known spectra from previous SMA studies, or (c) spectrally pure or
“extreme” pixels identified within the images being analyzed. Most applications of SMA
will use the third option because libraries of field-collected endmember spectra are rare, and
field spectrometers are expensive and not readily available to researchers. The empirical por-
tion of this article uses this third option.

SMA endmembers are selected by testing whether a given pixel’s spectra can be
described as a mixture of pixel spectra previously analyzed or whether it is itself another end-
member (Adams et al., 1993). A simple hypothetical illustration is provided in Figure 2.

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 47



Figure 2A shows the spectra for vegetation, soil, and shade. The vertical gray bars repre-
sent the wavelength response for Bands 1 and 2. Figure 2B plots the reflectance values of
various mixed pixels from an image for Band 1 and Band 2. Pixels composed of entirely one
component—vegetation, soil, or shade—would plot at the corners of the triangle presented
in Figure 2B. Pixel purity functions help to identify those pixels in an image that lie at the
extremes of the pixel value distribution in n-dimensional space (ENVI, 1996). The number
of endmembers chosen must be one less than the number of bands available in the multispec-
tral image (ENVI, 1996). For MSS images with four bands, this means that up to three end-
members can be identified. For TM images, six endmembers can be identified if all seven
TM bands are used; however, because Landsat TM’s thermal band represents a fundamen-
tally different wavelength, SMA analyses typically use only the six optical TM bands and
therefore five endmembers (see, e.g., Adams et al., 1995). Once endmembers are identified,
their reflectance values associated with image bands can be documented, presented in a form
similar to Figure 2A, and stored in a spectral library for use in this and future studies.

The second step in SMA is to estimate, for each pixel, the abundance of each endmember
contained within it by applying a linear mixing equation (Adams et al., 1993; ENVI, 1998).
The general form of this equation, in matrix form, is as follows:

R
[mxl] = EM[mxn] * X

[nxl] + e[mxl]

where,
R = original image values calibrated to surface reflectance,
EM = endmember spectra matrix,
X = vector of unknown abundances or fractions,
m = number of image bands used,
n = number of endmembers used (must be m – 1 or less), and
e = root mean-squared (RMS) error in the fit of the model. It represents the sum of the squared

residuals over all m bands.
This SMA equation is used to convert the existing image spectra values for each pixel into

endmember fraction matrices. One fraction image is produced for each endmember along
with the RMS error matrix. This procedure requires the fraction values produced in matrix X
to be positive and sum to unity (Adams et al., 1993; ENVI, 1998). These fraction images can
be created for each time point in a multitemporal set of satellite images all using the same
endmember spectra.

48 SOCIAL SCIENCE COMPUTER REVIEW

Figure 2: Graphic Representing Hypothetical Image Data for Spectral Mixture Analysis
SOURCE: Adapted from Adams, Smith, and Gillespie (1993, p. 152).



To date, very few studies have applied SMA toward questions related to the human
dimensions of environmental change. But it has great potential over the more traditional
image-analysis techniques for several reasons. First, like classification and PCA, SMA uses
all the information (bands) of a multispectral image.

Second, unlike many traditional classification methods, SMA is image independent.
SMA endmembers can be identified from image data, field, or lab inventories or from end-
member fraction libraries (Bateson & Curtiss, 1996). For this reason, time series or multiple
geographic location SMA fraction coverages are more readily comparable than are products
from classification products based on DNs.

Third, endmember fraction coverages represent physical properties of the landscape
based on physical units: surface reflectance values. Classification maps often represent
abstract human language artifacts—forest or pasture, for example—that may have different
meanings to different people and are not directly tied to any measurable physical property.
Two areas may be called a pasture but may be composed of very different vegetation types
that exhibit very different spectral qualities. Analyses conducted using endmembers are
based on specific spectral qualities of identifiable materials and not on human abstractions.
They can therefore be more easily applied to other geographic areas.

Fourth, endmember fraction images are more appropriate for analysis of physical land-
scapes exhibiting a high degree of continuously varying land cover, such as most forested
areas. This is especially important for studies of more continuous landscapes where changes
in light reflectance are gradational, such as heavily forested areas. Although classification is
quite useful in physical landscapes with discrete land forms, it is not as easily applied to land
cover that is more continuous in nature and less easily distinguishable.

Fifth, and perhaps most important, SMA provides the spectral data in terms of multiple
endmember fraction coverages and not as a single pixel classification, hence allowing a more
detailed analysis of pixel contents (Adams et al., 1993). Because of this, some refer to the
mixing approach as a subpixel analysis (ENVI, 1998). From this we can get a better idea of
what mixtures of land-cover features comprise any individual pixel. The same amount of
data is used in SMA and in classification techniques, but the employment of SMA endmem-
bers and the production of endmember fraction images allow for a more detailed analysis of
individual pixel contents. This is important even for satellite images with relatively good
spatial resolution, such as Landsat TM, but becomes especially important when using coarse
resolution instruments such as Landsat MSS.

The Importance of Image Restoration Procedures

Given these justifications for SMA and that SMA has been around for more than a decade
(e.g., Adams & Adams, 1984), why has it not been more widely used? There are two answers
to this question.

SMA requires the identification of pure reflectance spectra from the image itself or by
using spectra libraries (Adams et al., 1993). The individual band DNs provided in the raw
images received from U.S. repositories, such as EROS and Space Imaging/EOSAT, do not
represent Earth surface reflectance values that can be immediately used. More than 15 years
ago, Robinove (1982) recognized and nicely summarized the problem that still is prevalent
today:

Landsat digital images are commonly analyzed by using the digital numbers for each pixel
recorded on computer-compatible magnetic tape. Although this procedure may be satisfactory
when only a single, internally consistent image is used, the procedure may produce incorrect

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 49



results if more than one image is used for analysis as in mosaics or temporal overlays. The digital
numbers . . . vary depending on the calibration of the multispectral scanner in each satellite at a
given time, the sun angle, the state of the atmosphere, the slope and aspect of the terrain, and sur-
face cover. (p. 781)

To develop reflectance spectra of endmembers from the image itself, or use an endmember
reference from a library, the image must have sensor, illumination, and atmospheric variation
removed from the original image DNs. In other words, the DNs must be converted to surface
reflectance. As Robinove suggests, this conversion, sometimes referred to as image restora-
tion (Green et al., 1998; Jensen, 1996), is vital for any SMA temporal or spatio-temporal
comparison.

SMA has not been used because it is logistically difficult to undertake full image restora-
tion. When image comparisons are to be made, after preprocessing steps,1 three main restora-
tion processes should be undertaken: radiometric calibration, atmospheric correction
(Chavez, 1989; Teillet & Fedosejevs, 1995), and radiometric rectification (Hall, Strebel,
Nickeson, & Goetz, 1991). Radiometric calibration involves converting image DNs to radi-
ance values and radiance to at-satellite reflectance2 values (Hill, 1991; Markham & Barker,
1986; Robinove, 1982). Atmospheric correction removes the effects of the atmosphere
(e.g., light scattering) thereby converting at-satellite reflectance values to surface reflec-
tance (Ballew, 1975; Chavez, 1989; Teillet & Fedosejevs, 1995). Radiometric rectification
normalizes a set of time-series image data to one given image. This ensures that time-series
images are directly comparable to one another.

Although the information on how to conduct radiometric calibration is available, the lit-
erature is scattered and not readily available to the environmental change research commu-
nity (Green et al., 1998). The procedures required differ by satellite platform and by sensor
type. A literature review of major remote-sensing journals revealed that over the past 10
years only a small proportion of researchers actually converted DNs to surface reflectance
values (Green et al., 1998). Classification of one time point or multiple time points (by apply-
ing the radiometric rectification procedures of Hall et al., 1991) can be conducted as long as
the researcher has no interest in applying the same classification to another geographic area
(Green et al., 1998). Because of this ability to normalize images taken at different times, but
of the same location, the technical difficulties in converting DNs to surface reflectance val-
ues may appear to many to be not worth the effort. This decision to leave Landsat images in
DN data space means that SMA cannot be considered a viable analysis option on these prod-
ucts. Over the past 2 years, we have worked to produce partially automated spreadsheets and
accompanying documentation that allow the user to convert any Landsat TM or MSS image
from DN representation to surface reflectance values (for more information on the availabil-
ity of these procedures, contact either one of the authors).

The second answer to why SMA is not more widely used is more simple: SMA computer
programs are largely unavailable. SMA itself is a fairly complicated process (see Adams
et al., 1993) and is largely prohibitive unless software is written to allow such an analysis to
be undertaken. At the point of this writing, surprisingly, SMA functionality is still not avail-
able as part of the general analysis tools in many of the most prominent image-processing
packages. We have been able to identify two software packages that have this functionality.
The first is research software called Impact, developed by Milton Smith, Steve Cothern, and
Bill Gustafson at the Department of Geology’s remote-sensing laboratory at the University
of Washington. The second is ENVI®, a commercial remote-sensing software package with
linear unmixing capability. With image processing and SMA now outlined, let us provide an
example of how SMA can be applied to the study of human management of a public forest in
south-central Indiana.

50 SOCIAL SCIENCE COMPUTER REVIEW



SPECTRAL MIXTURE ANALYSIS IN A PUBLIC
FOREST OF SOUTH-CENTRAL INDIANA

As an example of the use of SMA for understanding the human dimensions of environ-
mental change, we will focus the rest of our discussion on landscape change in one portion of
the Hoosier National Forest, the Pleasant Run unit, which is located in south-central Indiana.
The brief analysis that follows seeks to show how SMA can help identify differences in two
forested landscapes subject to different management histories. For a more complete analysis
of this area and other public forests in Indiana, see Schweik (1998).

Understanding the role humans play regarding landscape change in a particular region
requires an investigation of what actions humans have taken over time and how these actions
relate to the physical properties of land-cover reflectance. This involves examining the rela-
tionship between light reflectance and the physical land cover and how human activities alter
that land cover. However, merely linking human actions to physical outcomes is not enough.
If we are to truly understand human-environment relationships then we must understand
how the incentive structures humans face in their day-to-day activities alter their decision-
making processes and the actions they ultimately take (Ostrom, 1990, in press). If social sci-
entists wish to link human decision making to information captured by satellites, we must
relate incentives to human actions, actions to outcomes, and outcomes to surface reflectance
(Schweik, 1998).

To undertake such a change analysis using satellite images and SMA requires a series of
steps. First, we must carefully select the images we use in order to minimize weather and sea-
sonal effects. Second, it is important to undertake a theoretical consideration of how light
reflectance is related to the land cover we wish to study—in this case, a forested landscape.
This is useful toward the development of appropriate endmembers for fraction image genera-
tion. Third, theoretical considerations must also be given to the incentive structures humans
face that relate to the choice of actions they undertake in the region. This allows us to develop
expectations on how humans have altered the physical landscape. Fourth, image processing
is undertaken to convert the raw image DNs to surface reflectance values. Fifth, we must
identify appropriate endmember locations and develop SMA fraction images. Finally, sixth,
we can analyze SMA results and create other output products such as fraction-based classifi-
cations to assist in change analysis. We will briefly describe these steps below.

Image Selection

A crucial decision the analyst must make when undertaking change analysis using multi-
spectral satellite images is image selection. Over the 25-year Landsat life span, there are a
large number of images that might be selected for analysis for any particular location in the
United States or elsewhere. Careful image sampling is particularly important for it allows the
analyst to minimize unwanted sources of variability. We followed several steps in choosing
the images to analyze.

First, we considered the temporal frequency of the human activities of interest. For this
study, we are interested in public forest management. Forest-cutting, development, and pres-
ervation activities are most likely to be identified spectrally. Development and preservation
activities will begin at a particular time point and then will most likely remain throughout the
Landsat time series. Forest cutting in this region, however, will result in an initial pattern of
disturbance followed by regrowth and possible cutting again after 30 or more years. Given
that the Landsat system has been operational for more than 25 years, maximizing the tempo-
ral extent of the image time series can more appropriately sample the forest-cutting activity.

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 51



To capture the phenomenon of forest cutting throughout time, Landsat images are selected
early in the archives duration (1972) and late (1997) with at least one image in between.

Second, we considered the ramifications of seasonal effects on spectral response. For the
study of forest resources, a late summer scene when tree canopies are full is appropriate.
Gaps where cutting exposes litter and soil are more apparent when a tree canopy is in full leaf
cover than in fall or winter scenes. There is also a contrast between herbaceous growth and
forests in late summer when dry conditions promote herbaceous senescence. For these rea-
sons, we selected images acquired in September. The reader may also raise the concern of
diurnal variability in satellite images. Fortunately, Landsat’s sun synchronous orbit ensures
that all images are taken at nearly the same time each day.

Third, we considered the year-to-year climatic variability in the images. In the study done
by Schweik (1998), it became apparent that our selected 1985 image was dramatically differ-
ent from two other time points. One hypothesis is that the difference is attributed to a higher-
than-normal rainfall during that year. It is important, then, to select images that have had
similar rainfall and temperature patterns during the months or even the year prior to image
acquisition.

Finally, cost drives image selection. For this study, we used a georegistered time series of
Landsat MSS images made available at no cost through the North American Landscape
Characterization (NALC) program: 9/30/72, 9/1/85, and 9/28/92 (Lunetta, Lyon, Guindon, &
Elvidge, 1998; NALC, 1997). Fortunately, the images made available through this program
satisfied the first two selection criteria, but as noted above, the 1985 image exhibited signifi-
cant spectral differences that are attributed to the higher rainfall that year.

Theoretical Considerations Related to the
Spectral Properties of the Landscape

The analyst must next consider the spectrally important characteristics of the landscape
under investigation. The human actions we wished to study were silviculture and develop-
ment activities in a public forest. This meant considering the spectral characteristics of this
broadleaf forest region.

The major factors that affect reflectance from a Midwestern oak-hickory forest canopy
are leaf area and leaf optical properties that are a product of leaf morphology (Green, 1996).
It is well understood that leaf chlorophyll absorbs sunlight predominantly at red (.65 µm) and
blue (.45 µm) wavelengths for use in photosynthesis. Less green light (.55 µm) is absorbed
by photosynthetic plants, thereby making them look green to the human eye. However, the
reflectance spectra of vegetation spectra also reveals a significant peak at near-infrared
wavelengths (.7 µm through 1.1 µm), an effect that the human eye cannot distinguish. This
increase in reflectance is a response not to absorption by chlorophyll molecules but rather by
scattering of light as it is reflected within the deeper internal structure of a leaf’s spongy
mesophyll (Campbell, 1996; Knipling, 1970).

In studies of similar oak-hickory forests in Missouri (Green, 1988, 1996; Green & Arvid-
son, 1986; Green, Arvidson, Sultan, & Guinness, 1985), distinctly high, near-infrared leaf-
reflectance values were found to be associated with vegetation from xeric environments:
areas with soils that exhibit low water-retention capacities. Alternatively, these studies report
forest vegetation in wetter, mesic environments to exhibit lower leaf-reflectance values in
near-infrared. Green (1996) attributes this phenomenon to different morphological qualities
of the leaf internal structure as plants adapt to various environmental conditions: Xeromor-
phic leafs differ from mesomorphic ones by exhibiting a thicker outer cuticle layer (Kramer

52 SOCIAL SCIENCE COMPUTER REVIEW



& Kozolowski, 1960; Spurr & Barnes, 1992), a higher ratio of internal to external surface
(Kramer & Kozolowski, 1960), and a thicker leaf with more palisade layers (Green, 1988). It
is thought that these differences cause additional scattering of infrared light in xeromorphic
leaves.

How do these phenomena relate to human actions on deciduous forests? Xeromorphic
leaf structures are often produced as a result of topographic conditions and water availability
factors. They also appear to have a relationship with age of forest vegetation (Schweik,
1998). Water is gathered by the root system of trees (Spurr & Barnes, 1992), and younger
trees will exhibit less extensive root networks, which result in less access to soil moisture.
Moreover, the canopy of younger trees, in an open stand that has recently been cut, will be
exposed to more direct sunlight, thus promoting the development of more xeric leaf struc-
tures, than will similar species growing in more shaded conditions (Spurr & Barnes, 1992).

Schweik (1998) analyzed well-documented cutting areas in a state forest in Indiana and
found that in general, environmental conditions and age-class appear to have a relationship
with the amount of reflected near-infrared light. Xeric topographic environments and young
vegetated stands reflect higher amounts of near-infrared light than do older forest vegetation
or vegetation residing in more mesic environments. The choice of MSS imagery to maximize
this study’s temporal duration limits us to the selection of three endmembers (four bands
minus one). For this study, we chose endmembers as follows: one endmember represents
xeric (brighter near-infrared) vegetation, and one represents nonliving matter (exposed soil
and developed regions). Because of the three-endmember limitation when using MSS
images, the third endmember represents darker near-infrared vegetation and also shadowed
vegetation (shade) (Adams et al., 1995). This endmember may also pick up pine stands as
well because pine tends also to exhibit dark vegetation spectra. If we were using solely Land-
sat TM data, we would be able to employ five endmembers: xeric vegetation, mesic vegeta-
tion, soil, shade, and pine. See Adams et al. (1995) for one SMA example using TM imagery.

Theoretical Considerations About
Human Behavior in the Region

Similar theoretical considerations must be addressed concerning human behavior and
action in the region. Given that the forests studied are under public forest management, we
focused our investigation on the national forest property manager. Taking a “new-
institutionalist” approach to the study of these managers (e.g., see Eggertsson, 1990) and fol-
lowing Ostrom’s (1990, in press) work, we assume that property managers face a suite of
alternative actions that they and their employees might consider undertaking. They choose
between alternative activities based on the incentive structures—the costs and benefits—that
are established to encourage or discourage different types of activities. Broadly speaking, the
suite of activity choices in Indiana public forests can be categorized as three types: commod-
ity- (e.g., silviculture practices), recreation- (e.g., trail management), and preservation-
related activities (Koontz, 1997).

The costs and benefits associated with any particular activity are influenced by attributes
of the physical world, attributes of the human communities interested in the forest and the
institutional landscape governing the area. The physical attributes of the region sometimes
make the costs of undertaking an action so high that certain activities will not be considered,
such as building a road on an extremely steep slope. Other activities may have extremely high
costs because of the reaction or anticipated reaction of users of the public forest. Anticipated

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 53



protests from environmental groups and landowners adjacent to the property may also
increase the expected costs of particular activities such as clear-cutting of forest stands.

We define the institutional landscape as the set of rules that are established to determine
the spatial distribution of where activities can and cannot be conducted along with the moni-
toring and sanctioning mechanisms used to enforce these rules (Schweik, 1997). A property
manager’s selection of certain activities to undertake is also encouraged or discouraged by
the configuration of this institutional landscape. In the case of the Hoosier National Forest,
legislation, revenue rules and allocations, standard operating procedures, and planning
documents place significant constraints on where particular activities can be undertaken
(Koontz, 1997; Schweik, 1998), and many of these restrictions involve a geographic component.

A full institutional history of the Hoosier Pleasant Run unit is beyond the scope of this
article (see Schweik, 1998). But as an example, we will focus on two different regions of
Pleasant Run that have had different institutional histories over roughly the past 20 years.
These two units have different types of activities that are permitted, required, or prohibited
within their boundaries as specified in the Hoosier forest management plans. The two areas
of interest for this article are Area 5.1, designated wilderness in 1982, and Area 7.1, desig-
nated an intensive recreation area; their boundaries are identified in Figures 5 and 6.

From knowledge of the institutional histories and assumptions that the behavior of the
property managers will largely comply with the incentive structures created by these institu-
tional configurations, one can develop a set of expected and testable hypotheses related to
forest land-cover change.

Hypothesis 1: The area designated as wilderness (Area 5.1) will exhibit little evidence of human
disturbance and signs of natural vegetation regrowth. As the Deam Wilderness area ages, image
pixels are expected to exhibit declining percentages of the xeric endmember and increasing per-
centages of the mesic endmember over time.

Hypothesis 2: The area designated as intensive recreation (Area 7.1) will exhibit higher levels of
development activities and permanent soil exposure than will the wilderness area. Consistently
higher percentages of the soil endmember will be found in this management area over time.

Image Processing

Image-processing steps involved a series of procedures to minimize nonsources of radio-
metric variability, such as that produced by the satellite and atmosphere, and to ensure geo-
metric comparability. We first verified that the images had little or no variability produced by
differences in sensor sensitivity (striping) and sensor power surges (speckle). Georeferenc-
ing has been performed on all NALC program scenes. This is a crucial step in change analy-
sis and ensures comparability across the time series of images. The danger is that misregis-
tration between images may be mistaken for land-cover change and could cause significant
errors in change analysis results. The NALC program data sets have specifications of geo-
referencing error below one pixel (NALC, 1997). We performed additional georeferencing
verification by overlaying a very accurate road network GIS coverage over each image to
visually confirm the georeferencing of each image in the area of study. We then converted the
image DNs to at-satellite reflectance and minimized atmospheric effects using the
spreadsheet-based procedure that we developed (Green et al., 1998). We then normalized the
9/30/72 and the 9/1/85 image to the base 9/28/92 image using the radiometric rectification
procedure developed by Hall et al. (1991). This processing converted the DNs of each NALC
image to surface reflectance. These postprocessed images can now be readily compared.

54 SOCIAL SCIENCE COMPUTER REVIEW



Endmember Identification and SMA
Fraction Image Development

The next step is to identify candidate xeric vegetation, mesic vegetation, and soil end-
member pixels within the image region that exhibit extreme spectral qualities for their land-
cover type (recall Figure 2B). The 9/28/92 MSS image was searched for locations that appear
to be pure representations of these land-cover types. Xeric forests were identified in MSS
images by their high near-infrared reflectances. In Indiana, they tend to be found on hill sum-
mit areas and on southwest-facing slopes. These areas have higher near-infrared reflectances
and appear brighter in image composites. Mesic forest vegetation regions, on the other hand,
tend to reside in valley bottoms (where water availability is high) and exhibit relatively
darker near-infrared reflectance.

Searching for extreme pure endmember locations is an iterative process. First, we
removed water areas from consideration by identifying those pixels that exhibited low values
across all MSS bands. These locations were removed from consideration so that SMA did
not have to model them. Next, we used three-dimensional scatter plots of band data to iden-
tify extremely bright soil and xeric vegetation locations. Three extremely bright soil loca-
tions, four extremely bright xeric vegetation locations, and three dark mesic vegetation loca-
tions were identified using this technique. Using differential global positioning systems
technology with about five meters’ accuracy, we verified these locations in the field to make
sure they were indeed good examples of these endmember types. The spectra for all other
pixels within the Hoosier forest region, theoretically, should be explained by linear combina-
tions of these three endmember spectra.

Several SMA models were run combining various soil, xeric, and mesic/shade candi-
dates. SMA output for each MSS time point is four new images: three representing the frac-
tional percentage of each endmember for each pixel and the fourth representing the RMS
error or goodness of fit of each model. Areas not modeled well are identified by high values
in the RMS error image. The final model has to fit two selection criteria: (a) The spectral vari-
ability present is largely accounted for, and (b) the model produces endmember percentages
that fall between 0% and 100% (ENVI, 1998; Ustin et al., 1993, p. 345). Time-series analysis
requires a third criteria: The models must satisfy the first two criteria for all image dates. This
modeling process is complicated by limitations in MSS degrees of freedom and because the
mesic/shade endmember does not model both categories particularly well. The model that
satisfied these conditions the best used endmembers of a cleared agricultural field, a xeric
vegetation location residing on a summit and partially on a southwest-facing slope, and a
mesic vegetation location that resides in a valley and is somewhat shaded. Figure 3 displays
the stick spectra graphs for the final soil, xeric vegetation, and mesic vegetation endmembers
chosen. The resultant fraction and RMS images produced from this SMA model for one
region of the 9/28/92 time point are shown in Figure 4. Similar products were produced for
the other dates.

The landscape in Figure 4 highlights an example that will help the reader interpret SMA-
derived images. Circled in the image is a known area where a forest property manager office
exists. Note that this area is very bright in the soil fraction image, meaning that these pixels
have a high percentage of exposed soil. This area is also dark in the xeric endmember fraction
image and, therefore, exhibits a low percentage of this endmember. Finally, it is moderately
dark in the mesic/shade endmember, so some mesic vegetation and shadow also exist within
this area. This provides an example of how SMA provides the analyst with the capability to
conduct subpixel analysis—a capability not possible using traditional image classification

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 55



analysis techniques. By employing SMA, we have effectively moved our analysis from one
depicting changes in reflectance to one depicting changes in surface composition—a mea-
surement system much more accessible to the social science community. We now also have a
mechanism to evaluate how well this system is defined for each pixel (the RMS image).

Change Map Creation and Analysis

With the SMA fraction images developed, several types of analysis techniques can be
applied. Schweik (1998), for example, analyzed specific locations where known forest man-
agement activities have occurred and found that these SMA fraction images are sensitive to
various types of cutting and development activities. However, the hypotheses stated above
require an analysis at a broader spatial scale: the landscape.

Hypothesis 1 predicts that the Deam Wilderness will exhibit, overall, a general decline in
the percentage of the xeric endmember with time. Because the institutional landscape for this
area prohibits timber-cutting activities and the incentive structure strongly encourages prop-
erty managers to adhere to this mandate, it is expected that vegetation in this area designated
as wilderness will exhibit signs of aging across the time series. As trees get older, it is
expected that the cell structure of their leaves will become less xeromorphic as their root sys-
tems become more extensive, accessing more moisture, and as a greater proportion of their
leaves become protected from direct exposure to sunlight.

56 SOCIAL SCIENCE COMPUTER REVIEW

Figure 3: Stick Spectra of Chosen Endmembers: Soil, Bright (xeric) Vegetation, and Dark
(mesic) Vegetation and Shade (Landsat Multispectral Scanner Path 21, Row 33,
9/28/92)



A color composite image can be used qualitatively to test this hypothesis by assigning
each of the three xeric fraction images of each time-point image to blue, green, and red. In the
resulting image, different colors are associated with different temporal trajectories. The dif-
ficulty with this method is that the continuous nature of the fraction image data (and satellite
image data in general) makes interpretation of the subtle color changes difficult. This is
exactly why humans find classification maps so useful: They aggregate continuous data into
a few discrete categories, greatly simplifying analysis.

Recall the earlier argument about traditional classification approaches being image
dependent and not readily transferable to analyses of other images. Using SMA fraction
images, the analyst can now develop classifications that no longer are based on statistical
measures of a single image but are instead developed from endmember fractions, grounded
in physical measures of reflectance, and acquired at specific locations on the Earth’s surface.
For example, the analyst can now define what he or she considers a xeric forest based on
these percentages of certain materials. In our study, we define a xeric forest as one possessing
40% or higher of the xeric endmember and less than 40% of either the soil or mesic/shade
endmembers.

With these definitions established, the analyst can easily convert the continuous data of
the three endmember fractions for any image into a classification map based on endmember
values using software modeling functionality.3 Moreover, the same classification definitions
can be used for each image time point, something that cannot easily be accomplished using
traditional image classification procedures. Side-by-side image classification maps based on
SMA fractions for the Wilderness area are shown in Figure 5,4 and Figure 6 presents similar
maps for the intensive recreation area. To demonstrate the utility of these products, let us now
briefly address the hypotheses stated earlier and interpret these classification maps.

Hypothesis 1 predicted that given that Area 5.1 is treated as wilderness, the vegetation
here will exhibit signs of continued forest aging. In terms of SMA fraction images, it is
expected that the vegetation in the wilderness area will, in general, be more xeric in 1972 and

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 57

Figure 4: Spectral Mixture Analysis for a Portion of the Landsat Multispectral Scanner 9/28/92
Image

NOTE: Bright white = high percentage; dark = low percentage; RMS = root mean squared.



move more toward mesic conditions in 1992 after 20 years of root growth and increasing
moisture availability. Reviewing the side-by-side classification images in Figure 5, several
pieces of evidence tend to confirm this hypothesis.

First, the northern peninsula of the Deam Wilderness, circled and labeled A in both maps
of Figure 5, reveals a significant amount of exposed soil (light areas or gold in color) in the
1972 map, whereas in the 1992 map, it is replaced by a combination of the mesic vegeta-
tion/pine/shade class. A review of a September 1997 TM image reveals a significant pine
stand on this peninsula. Because pine is not indigenous to this area, this suggests that a major
planting initiative occurred in this area between 1972 and 1992, probably to protect soil ero-
sion into the adjacent Lake Monroe.

Second, the conversion of soil to vegetation is apparent in the area identified by B in Fig-
ure 5, which exhibits a pattern we would expect in an area left alone to regenerate. Scattered
soil-dominated pixels (gold) in 1972 have been replaced by more prevalent xeric spectra
(red), mesic/pine/shade spectra (dark green), or a combination of xeric/mesic vegetation
(bright green). Thus, as the Deam ages after its designation in 1982, forest vegetation appears
to be returning.

Third, the large circles labeled C on the Figure 5 maps reveal a significant area that exhib-
its what appears to be following patterns of natural regrowth. Many of the stands that exhib-
ited xeric (red) characteristics in 1972 remain xeric in 1992. We used a digital elevation
model for the region to identify locations that are naturally more xeric in nature (e.g., sum-
mits and southwest-facing slopes). These areas would be expected to remain xeric over time
because of the natural effects of their physical environment. But many other xeric-dominated

58 SOCIAL SCIENCE COMPUTER REVIEW

Figure 5: Change in the Hoosier National Forest Pleasant Run Unit Area 5.1 (Deam Wilderness)
NOTE: For a color version, visit the Web site at http://www.indiana.edu/~cipec/publications/
publications.html. MSS = multispectral scanner; SMA = spectral mixture analysis.



areas in 1972 have been replaced in 1992 with higher percentages of mesic vegetation (bright
green or dark green areas in the color versions of the maps). These areas take on the expected
high-to-low xeric to lower xeric change trajectory that Schweik (1998) identifies as a spec-
tral pattern of natural regrowth in these forests. This evidence also strongly supports
Hypothesis 1.

Hypothesis 2 predicts that the high recreation area will exhibit patterns of development
(permanent soil exposure and nonvegetative land cover) in areas where the property manag-
ers build or maintain recreation infrastructures. Reviewing areas identified by A in Figure 6
reveals a higher level of soil exposure in this area in 1972 than that found in the Deam Wilder-
ness area. Note, however, that in 1992 much of this soil has been replaced with mesic or
mesic/xeric land cover. The history of the area helps to explain this trajectory. The lake adja-
cent to this area was created in the mid-1960s. The soil exposure in 1972 is a result of signifi-
cant recreation infrastructure development that occurred shortly thereafter. However, as time
has passed, property managers have worked to allow the remaining forest surrounding these
recreation areas to continue to age because larger trees are more visually appealing to visi-
tors. Therefore, the change trajectory of the area identified by A depicts a management strat-
egy of development coupled with visual enhancement (natural regeneration) efforts. As trees
have aged, their canopies have grown and covered many of the developed areas (e.g., roads)
that were more visible in the early development in 1972. Areas identified in Figure 6 by B in
these maps exhibit a similar natural regeneration spectral trajectory as managers work to
keep visual enhancement areas pleasing to visitors.

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 59

Figure 6: Change in the Hoosier National Forest Pleasant Run Unit Area Designated 7.1 (inten-
sive recreation area)

NOTE: For a color version, visit the Web site at http://www.indiana.edu/~cipec/publications/
publications.html. MSS = multispectral scanner; SMA = spectral mixture analysis.



DISCUSSION AND CONCLUSION: THE PROMISE OF
SMA FOR STUDYING THE HUMAN DIMENSIONS
OF ENVIRONMENTAL CHANGE

The tremendous advances in desktop computing, even over the past 5 years, has produced
great interest within the social science community on the use of multitemporal, multispectral
satellite images for the study of human-environment relationships. The technology and
know-how that used to only reside in the hands and minds of scholars with graduate degrees
in remote sensing can now be more commonly applied by scholars in various disciplines
such as geography, anthropology, public administration, and political science: all fields with
an interest in understanding the human role in global change.

Let us summarize our main thesis. Multitemporal, multispectral satellite images, such as
Landsat, are very useful in the study of the human dimensions of environmental change stud-
ies because the archive covers a tremendous spatial extent of the Earth’s surface and more
than a quarter century in duration with very frequent temporal sampling. Their radiometric
and spatial resolution continues to improve as new satellites are launched. However, to link
the information provided in these images to human incentives, actions, and outcomes such
that comparisons can be made between studies, we must carefully process the images to
remove nonsurface sources of variability that cannot be attributed to variations in land cover
and land-cover change. To remove this variability, images should be processed from DNs to
physical units: surface reflectance values, through radiometric calibration and atmospheric
correction procedures. In doing so, analyses can be duplicated and replicated by researchers
studying other locations, and accurate comparisons can be made across space and time. Fur-
thermore, by reporting endmember characteristics and locations and applying SMA to gen-
erate endmember fraction images, the researcher can define classifications grounded on a
physical measure of the Earth’s surface: reflectance and based on how much of what material
is present. By using the xeric/mesic distinction, we have classified our forests based on quan-
titative, ecologically significant properties tied to measurable forest properties such as leaf
morphology. Unlike traditional classification, these categories can be readily applied and
replicated by others working in distant but comparable environmental regions.

To understand the role humans have played in changing land cover, one must give careful
attention to both natural as well as human parameters. The analyst must use effective sam-
pling strategies and GIS products (e.g., digital elevation models or other topographic prod-
ucts) to account for natural influences such as rainfall, elevation, and slope aspect. To under-
stand the human dimensions of change and the motivation behind them, the analyst
should identify (a) the suite of alternative actions undertaken by humans in a particular con-
text, (b) how these actions physically disturb the landscape and affect light reflectance, and
importantly, (c) what incentive structures (e.g., rules, social norms, and legislation) have
driven human decision-making and activity choice (Schweik, 1998). By linking incentives
to actions and actions to outcomes measured with physically based yardsticks such as sur-
face reflectance values, more straightforward human-environment relationships can be
understood.

Image processing and SMA have the potential to produce other benefits as well. If physi-
cal relationships can be identified between vegetation reflectance and SMA percentages
with other important global change parameters, such as those associated with forest health
under climate change, researchers may be able to use spectral measurements as proxies for
these vitally important but costly-to-measure parameters and develop image-based inven-
tory estimates appropriate for particular forested landscapes.

We conclude with guidelines on how to link the study of human incentives and actions
with SMA-based landscape measurements. The researcher/analyst should

60 SOCIAL SCIENCE COMPUTER REVIEW



1. investigate and document the incentive structures that drive human actions in the context of the
study location;

2. inventory the suite of potential activity choices that humans in this area undertake;

3. consider the temporal frequency of land-cover disturbances that these activities will produce
and select sufficient images to critically sample the phenomenon of interest;

4. control for season and weather effects through careful selection of satellite images;

5. identify a subset of locations where these actions have occurred and investigate the physical
response produced in the image time series;

6. control for topographic effects through the use of GIS products such as a digital elevation mod-
els; and

7. convert the multispectral images from DNs to surface reflectance values such that nonsurface
sources of variation can be removed and not attributed to human actions. In doing so, findings
can be compared with other studies of similar forest types.

A much more detailed description and example of this process is shown in Schweik
(1998). For more information, contact the authors or visit our World Wide Web site at
http://www.indiana.edu/~cipec.

NOTES

1. Image preprocessing involves the removal of pixel-to-pixel variability not associated with atmospheric or sur-
face phenomena. It involves the conversion of digital images from CD or tape to the appropriate image-processing
software format, removing any artifactual noise from the image, such as line dropouts and line striping, and image
georeferencing.

2. At-satellite reflectance describes the ratio of light received by the sensor from the entire Earth-atmosphere
system over the solar irradiance at Earth-sun distances. These values still contain the effects of the atmosphere, such
as scattering and absorption.

3. In our case, we use model maker, which is part of Imagine® software, to produce this classification.
4. These classifications cannot easily be interpreted in black and white. Therefore, the color classifications for this

article are available at http://www.indiana.edu/~cipec/publications/publications.html.

REFERENCES

Adams, J. B., & Adams, J. D. (1984). Geologic mapping using Landsat MSS and TM images: Removing vegetation
by modeling spectral mixtures (Third thematic conference, remote sensing for exploration geology). ERIM, 2,
615-622.

Adams, J. B., Sabol, D., Kapos, V., Filho, R. A., Roberts, D. A., Smith, M. O., & Gillespie, A. R. (1995). Classifica-
tion of multispectral images based on fractions of endmembers: Application to land-cover change in the Brazil-
ian Amazon. Remote Sensing of Environment, 52, 137-154.

Adams, J. B., Smith, M. O., & Gillespie, A. R. (1993). Imaging spectroscopy: Interpretation based on spectral mix-
ture analysis. In C. M. Pieters & P. Englert (Eds.), Remote geochemical analysis: Elemental and mineralogical
composition (pp. 145-166). New York: Cambridge University Press.

Adams, J. B., Smith, M. O., & Johnson, P. E. (1986). Spectral mixture modeling: A new analysis of rock and soil
types at the Viking Lander 1 site. Journal of Geophysical Research, 91, B8, 8098-8112.

Ballew, G. I. (1975). A method for converting Landsat 1 MSS data to reflectance by means of ground calibration
sites (Stanford remote sensing lab technical report, SRSL 75-5). Stanford, CA: Stanford University.

Bateson, A., & Curtiss, B. (1996). A method for manual endmember selection and spectral unmixing. Remote Sens-
ing and Environment, 55, 229-243.

Brondizio, E. S., Moran, E. F., Mausel, P., & Wu, Y., (1994). Land use change in the Amazon estuary: Patterns of
Caboclo settlement and landscape management. Human Ecology, 22(3), 249-278.

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 61



Bryan, W. B., Finger, L. W., & Chayes, F. (1969). Estimating proportions in petrographic mixture equations by
least-squares approximation. Science, 163, 926-927.

Campbell, J. B. (1996). Introduction to remote sensing (2nd ed.). New York: Guilford.
Chavez, P. S. (1989). Radiometric calibration of Landsat thematic mapper multispectral images. Photogrammetric

Engineering and Remote Sensing, 55, 1285-1294.
Cohen, W. B. (1991). Response of vegetation indices to changes in three measures of leaf water stress. Photogram-

metric Engineering and Remote Sensing, 57, 203-207.
Drury, S. A. (1990). A guide to remote sensing: Interpreting images of the Earth. Oxford, UK: Oxford University

Press.
Eggertsson, T. (1990). Economic behavior and institutions. Cambridge, UK: Cambridge University Press.
ENVI. (1996). ENVI tutorials, ENVI version 2.5. Boulder, CO: Better Solutions Consulting Limited Liability Co.
ENVI. (1998, February 7). ENVI tutorial: Multispectral processing using ENVI’s hyperspectral tools. Available:

http://www.envi-sw.com/e_msihsi.htm.
Green, G. (1988). Physical basis for remotely sensed spectral variation in a semi-arid shrubland and an oak-

hickory forest: Implications for mapping soil types in vegetated terrains. Unpublished doctoral dissertation,
Washington University, Department of Earth and Planetary Sciences, St. Louis, MO.

Green, G. (1996). Forest structure and leaf spectral control on canopy reflectance: Implications for geologic map-
ping using remote sensing in forested terrains (Working paper). Bloomington: Indiana University, Anthropo-
logical Center for Training and Research on Global Environmental Change.

Green, G., & Arvidson, R. E. (1986). Soil types and forest canopy structures in southern Missouri: A first look with
AIS data. In G. Vane & A. Goetz (Eds.), Proceedings of the Second Airborne Imaging Spectrometer Data Analy-
sis Workshop (pp. 153-161). Pasadena, CA: NASA Jet Propulsion Laboratory.

Green, G., Arvidson, R., Sultan, M., & Guinness, E. (1985, April). Geobotanical information contained in Landsat
thematic mapper images covering southern Missouri. Paper presented at the Fourth Thematic Conference,
Remote Sensing for Exploratory Geology, San Francisco.

Green, G., Schweik, C. M., & Hanson, M. (1998). Radiometric calibration of Landsat multispectral scanner and
thematic mapper images: Guidelines for the global change community (Working paper). Bloomington, IN:
Center for the Study of Institutions, Population, and Environmental Change.

Hall, F. G. (1994). Adaptation of NASA remote sensing technology for regional-level analysis of forest ecosystems.
In V. A. Sample (Ed.), Remote sensing and GIS in ecosystem management. Washington, DC: Island Press.

Hall, F. G., Strebel, D. E., Nickeson, J. E., & Goetz, S. J. (1991). Radiometric rectification: Toward a common radio-
metric response among multidate, multisensor images. Remote Sensing of Environment, 35, 11-27.

Hill, J. (1991). A quantitative approach to remote sensing: Sensor calibration and comparison. In A. S. Belward &
C. R. Valenzuela (Eds.), Remote sensing and geographical information systems for resource management in
developing countries (pp. 97-110). Brussels, Belgium: ECSC, EEC, EAEC.

Idrissi. (1997). Idrissi users guide version 2.0. Worcester, MA: Clark University.
IGAP. (1998, March 10). [Indiana GAP homepage]. Available: http://139.102.7.220/h1/bertha/gap.
Jackson, R. D. (1983). Spectral indices in n-space. Remote Sensing of the Environment, 13, 409-421.
Jensen, J. R. (1996). Introductory image processing: A remote sensing perspective. Englewood Cliffs, NJ: Prentice

Hall.
Knipling, E. (1970). Physical and physiological basis for the reflectance of visible and near-infrared radiation from

vegetation. Remote Sensing of the Environment, 1, 155-159.
Koontz, T. (1997). Federalism and national resource policy: Comparing state and national management of public

forests. Unpublished doctoral dissertation, Indiana University, Bloomington.
Kornder, S. C., & Carpenter, J. R. (1984). Application of a linear mixing algorithm to the normal alkane patterns

from recent salt marsh sediments. Organic Geochemistry, 1(7), 61-71.
Kramer, P., & Kozolowski, T. (1960). Physiology of trees. New York: McGraw-Hill.
Lee, C. T., & Marsh, S. E. (1995). The use of archival Landsat MSS and ancillary data in a GIS environment to map

historical change in an urban riparian habitat. Photogrammetric Engineering and Remote Sensing, 61(8),
999-1008.

Lillesand, T. M., & Kiefer, R. W. (1994). Remote sensing and image interpretation. New York: John Wiley.
Lunetta, R. S., Lyon, J. G., Guindon, B., & Elvidge, C. D. (1998). North American landscape characterization data-

set development and data fusion issues. Photogrammetric Engineering & Remote Sensing, 64(8), 821-829.
Markham, B. L., & Barker, J. L. (1986, August). Landsat MSS and TM post-calibration dynamic ranges, exoatmos-

pheric reflectances and at-satellite temperatures. EOSAT Landsat Technical Notes, 1, 3-8.
Moran, E., Brondizio, E., & Mausel, P. (1994). Secondary succession. National Geographic Research and Explora-

tion, 10(4), 458-476.

62 SOCIAL SCIENCE COMPUTER REVIEW



North American Landscape Characterization (NALC). (1997, April 4). North American landscape characteriza-
tion. Available: http://edcwww.cr.usgs.gov/glis/hyper/guide/nalc.

Ostrom, E. (1990). Governing the commons: The evolution of institutions for collective action. Cambridge, UK:
Cambridge University Press.

Ostrom, E. (in press). Self-governance of common-pool resources. In P. Newman (Ed.), The new Palgrave diction-
ary of economics and the law. London: Macmillan.

Price, J. C. (1987). Calibration of satellite radiometers and comparison of vegetation indices. Remote Sensing of
Environment, 18, 35-48.

Richards, J. A. (1986). Remote sensing digital image analysis: An introduction. New York: Springer-Verlag.
Robinove, C. J. (1982). Computation with physical values from Landsat digital data. Photogrammetric Engineering

and Remote Sensing, 48(5), 781-784.
Sader, S. A. (1995). Spatial characteristics of forest clearing and vegetation regrowth as detected by Landsat the-

matic mapper imagery. Photogrammetric engineering and remote sensing, 61(9), 1145-1151.
Sader, S. A., & Wynne, J. C. (1992). RGB-NDVI color composites for visualizing forest change dynamics. Interna-

tional Journal of Remote Sensing, 13, 3055-3067.
Schweik, C. M. (1997). Spatial analysis of natural resources in East Chitwan, Nepal: Conceptual issues and a

multiscale research program. In G. Shivakoti, G. Varughese, E. Ostrom, A. Shukla, & G. Thapa (Eds.), People
and participation in sustainable development: Understanding the dynamics of natural resource systems (Pro-
ceedings of an International Conference, 17-21 March 1996, Institute of Agriculture and Animal Science,
Tribuvan University, Rampur, Chitwan, Nepal). Bloomington: Indiana University, Workshop in Political The-
ory and Policy Analysis.

Schweik, C. M. (1998). The spatial and temporal analysis of forest resources and institutions. Unpublished doc-
toral dissertation, Indiana University, Bloomington.

Scott, J. M., Davis, F., Csuti, B., Noss, R., Butterfield, B., Groves, C., Anderson, H., Caicco, S., D’Erchia, F.,
Edwards, T. C., Ulliman, J., & Wright, R. G. (1993). Gap analysis: A geographic approach to the protection of
biological diversity. Wildlife Monographs. Available: http://www.gap.uidaho.edu/GAP/AboutGAP/
WildlifeMonographs/C.htm.

Smith, M. O., Ustin, S. L., Adams, J. B., & Gillespie, A. R. (1990a). Vegetation in deserts: I. A regional measure of
abundance from multispectral images. Remote Sensing of Environment, 31, 1-26.

Smith, M. O., Ustin, S. L., Adams, J. B., & Gillespie, A. R. (1990b). Vegetation in deserts: II. Environmental influ-
ences on regional abundance. Remote Sensing of Environment, 31, 27-52.

Spurr, S. H., & Barnes, B. V. (1992). Forest ecology. Malabar, FL: Krieger.
Teillet, P. M., & Fedosejevs, G. (1995). On the dark target approach to atmospheric correction of remotely sensed

data. Canadian Journal of Remote Sensing, 21(4), 374-387.
Tucker, C. J., VanPraet, C. L., Sharman, M. J., & Van Ittersum, G. (1985). Satellite remote sensing of total herba-

ceous biomass production in the Senegalese Sahel, 1980-1984. Remote Sensing of the Environment, 17,
233-249.

Ustin, S. L., Smith, M. O., & Adams, J. B. (1993). Remote sensing of ecological processes: A strategy for develop-
ing and testing ecological models using spectral mixture analysis. In J. Ehleringer & C. Field (Eds.), Scaling
physiological processes: Leaf to globe (pp. 339-357). San Diego, CA: Academic Press.

Verbyla, D. L. (1995). Satellite remote sensing of natural resources. Boca Raton, FL: Lewis.
Westman, W. E., Strong, L. L., & Wilcox, B. A. (1989). Tropical deforestation and species endangerment: The role

of remote sensing. Landscape Ecology, 3(2), 97-109.
Wilkie, D. S., & Finn, J. T. (1996). Remote sensing imagery for natural resources monitoring: A guide for first-time

users. New York: Columbia University Press.
Wolter, P. T., Mladenoff, D. J., Host, G. E., & Crow, T. R. (1995). Improved forest classification in the northern lake

states using multi-temporal Landsat imagery. Photogrammetric Engineering and Remote Sensing, 61(9), 1129-
1143.

Charles M. Schweik has a Ph.D. in public policy from Indiana University and is a research associate with
Indiana University’s Center for the Study of Institutions, Population, and Environmental Change (CIPEC).
He currently is a visiting scholar at the Center for Public Policy and Administration, University of Massa-
chusetts, Amherst. His e-mail address is Cschweik@pubpol.umass.edu.

Glen M. Green is a remote-sensing postdoctoral scholar at CIPEC, Indiana University. He received his
Ph.D. from Washington University (Department of Earth and Planetary Sciences) in St. Louis, Missouri. His
e-mail address is Glgreen@indiana.edu.

Schweik, Green / SPECTRAL MIXTURE ANALYSIS 63