Linear Mixture Modelling Solution Methods for

Satellite Remote Sensing

College of Aeronautics Report 9603
ISBN 1 871564 83 2

Stephen Hobbs
College of Aeronautics, Cranfield University
Cranfield, Bedford MK43 0AL, UK
[email protected]
27 March 1996, revised 4 Nov 2003

Abstract
This report documents the algorithms used in the program MIXMOD to analyse mixed pixel data (assuming linear mixing). The report describes the mathematical algorithms rather than acting as a users manual for MIXMOD. The
algorithms described obtain the desired solutions, quantify the quality of the solution, and estimate error bounds, using a variety of methods. A novel feature
is the ability to handle uncertainty in the assumed end-member spectra, which
in practice may be the dominant source of error. The report includes a brief
literature review to place the work in its broader context (with references to the
mathematics of linear systems and other applications of mixture modelling in
remote sensing).

Contents
1 Introduction
1.1 Linear mixture modelling . . . . . . . . . . . . . . . . . . . . . .
1.2 Related analysis techniques and applications . . . . . . . . . . . .
1.3 Structure of the report . . . . . . . . . . . . . . . . . . . . . . . .

1
1
2
2

Literature Overview
2.1 Mathematical Methods . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Comparative Studies . . . . . . . . . . . . . . . . . . . . . . . . .

5
5
5
6

Solution Methods Overview

3.1 The assumption of linearity . . . . . . . . . . . .
3.2 Constraining the solution . . . . . . . . . . . . .
3.2.1 Constraints applied to pixels individually
3.2.2 Constraints applied to groups of pixels . .
3.3 Testing solution quality . . . . . . . . . . . . . .
3.4 Solution error bounds . . . . . . . . . . . . . . .
3.5 The MIXMOD program . . . . . . . . . . . . . .

6
7

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

7
7
7
8
8
8
8
9

Solution Methods Ignoring End-Member Noise

4.1 No constraint on the sum of components . . . . . . .
4.2 Sum of components equal to 1 . . . . . . . .P. . . . .
nc1
4.2.1 Forced exact compliance: xnc = 1 j=1 xj
4.2.2 Lagrange Multiplier Method . . . . . . . . . .
4.3 All components positive . . . . . . . . . . . . . . . .
4.4 Multiple constraints . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

11
12
12
13
13
13
14

Solution Methods Allowing for End-Member Noise

5.1 Approximate Methods . . . . . . . . . . . . . . . . . .
5.2 Exact Methods . . . . . . . . . . . . . . . . . . . . . .
5.2.1 No constraint on sum of components . . . . .
5.2.2 Sum of components equal to 1 . . . . . . . . .
5.2.3 All components positive . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

15
15
15
17
17
18

Estimating Error Bounds

.
.
.
.
.
.
.

19

Developments of the Standard Methods

21
7.1 Reducing uncertainty on estimated proportions . . . . . . . . . . 21
7.2 Estimating spectral signatures using known proportions . . . . . 21
i

8 Conclusions

23

Bibliography

25

ii

Chapter 1

Introduction
The principal aim of this report is to document the algorithms used in the
program MIXMOD for the analysis of mixed pixel data. Subsidiary aims are
to summarise the principles on which linear mixture modelling are based and
to provide a brief overview of the literature relevant to the subject.
Linear mixture modelling refers to cases in remote sensing in which the
signal measured is, for practical purposes, a linear mixture of several different
fundamental signals (usually called end-members in the literature). The
goal is often to estimate the relative contributions of the different end-member
components given only measurements of the composite (or mixed) signals for
a scene.
The report lists references to applications and discussions of the methods
specifically in the area of satellite remote sensing. References are also given to
the much larger body of literature concerned with general linear systems and to
some related applications of the techniques.
Program MIXMOD which is used at Cranfield University for the analysis
of mixed pixel data uses several different solution algorithms; it also evaluates
parameters to quantify the goodness-of-fit of the solution, and calculates appropriate error bounds. All the methods currently used for these functions are
documented in this report, which therefore serves as both a reference for MIXMOD and also as a (partial) review of the mathematical methods appropriate
to linear mixture modelling.

1.1

Linear mixture modelling

The general case being considered is that in which the outputs of some linear
combination process are measurable, but in which the parameters of interest are
actually the inputs. In satellite remote sensing, this occurs when the elemental
area of the Earths surface being imaged (a pixel) contains several different
fundamental cover components (e.g. water, bare soil, grass, forest) and so the
measured spectrum is a composite of the spectra from the different individual
components. If the sensors resolution limit is greater than the pixel size then
neighbouring pixels are not independent measurements of surface properties.
In many cases of interest, the composite spectrum is a linear combination of
the spectra of the individual components, and it is the areas covered by the

individual components which are required.

The relationship between the measured signal in band i (yi ), the value of the
reference signal in band i for a pixel wholly filled by component j ([A]ij = aij ),
and the proportion of the pixel actually occupied by cover component j (xj ), is
given by:
y

= Ax

(1.1)

In this case, the inputs are the class proportions and the outputs are the
signals in each band. In general, the number of bands does not equal the number
of cover classes, and the matrix equation cannot be inverted directly since A
is not square. There are usually more bands than classes, and a conventional
least squares matrix solution method is used (which may or may not account for
uncertainty in the measurements). An important feature of the various leastsquares methods is their ability to cope with a degree of noise in the system.
Conventionally, this noise is assumed to lie solely in the signal measurements
made in each spectral band. Methods reported in the literature do not yet
include the ability to deal with noise in the end-member reference spectra which
is likely to be the dominant source of noise in most applications. Appropriate
solution methods are presented in sections 5 and 6 of this report, and are also
discussed in [7].
Practical methods of estimating the pixel proportions take many forms, and
often make use of additional information to try to improve the solution estimate. The sophistication of solution methods also varies widely, from relatively
straightforward inversion methods for a fixed number of bands (e.g. [25]) to
much more general algorithms able to incorporate a wide range of additional
information (e.g. [24]).

1.2

Related analysis techniques and applications

Many fields of study use linear models to relate inputs and outputs, and the
mathematics of such systems is a highly developed subject. [15] discuss the
mathematics of solution methods for linear systems in good detail, and refer
to related topics such as fitting models to data, and regularised solutions (see
below).
In remote sensing, a mature field in terms of the algorithms used is atmospheric profiling. Rodgers [19, 9, 20] authoritatively reviews the inversion
methods used in this field, covering linear methods as well as more sophisticated techniques.

1.3

Structure of the report

Following this Introduction, Section 2 gives an overview of literature relevant to

linear mixture modelling. This is not intended to be a comprehensive review, but
does provide pointers to a wide range of work, covering not only applications
of conventional mixture modelling, but also a range of related fields, and
some relevant mathematical texts discussing the algorithms used. This overview
places mixture modelling in its broader context.
2

Section 3 provides a brief introduction to the different solution methods

available for linear mixture modelling and shows how they relate to each other.
The core of the material presented in the report describes the algorithms implemented in the program MIXMOD (Sections 4 to 6). This gives a sound
background understanding for users of MIXMOD. The descriptions are however quite general and allow the reader to develop his or her own applications
as necessary as well as illustrating the methods discussed earlier in the report.
Section 7 discusses briefly some extensions of the standard methods.

Chapter 2

Literature Overview
The references listed with this report are a sample of the literature relevant to
the subject of linear mixture modelling. The articles generally deal with either
the mathematical principles and relevant algorithms, specific applications, or
with comparative studies for several different approaches to the same problem.
The following notes are included to guide the reader to appropriate articles for
specific topics.

2.1

Mathematical Methods

[15] give an excellent guide to the use of mathematical methods for data analysis. Topics covered include linear systems (Chapter 2), fitting models to data
(Chapter 15), quantifying solution quality and errors (sections 15.1 and 15.6 of
Chapter 15), and regularising solutions (sections 18.4ff of Chapter 18). Papers
covering the application of these methods to remote sensing include [24] and the
work of Rodgers [19, 9, 20].

2.2

Applications

One of the most active groups using mixture modelling techniques is that based
at the University of Washington (Seattle, USA) [26, 1, 18, 27, 21, 25, 22, 28].
Their work, principally for land surfaces includes hyperspectral geological applications, and has spawned projects at related institutions [8, 14]. Monte Carlo
methods have been used to try to analyse the ability of measurements to discriminate surface types and to evaluate error bounds (e.g. [22]), but the approach
is largely empirical, without always justifying underlying assumptions (e.g. the
use of shade as an end-member, or the application of the method to water colour
due to sediment loading).
Perhaps the most comprehensive work, mathematically, (although it does
not deal with end-member uncertainty) is that of [24]. Topics such as the
regularisation of solutions are discussed. Non-mathematical readers may find
the work daunting.
Various case studies have been reported [3, 5, 16, 17, 2, 12, 30, 29]. The work
of [31] makes an interesting link between mixture modelling and data fusion.

Outside the area of surface remote sensing, Rodgers [19, 9, 20] gives authoritative reviews of atmospheric profiling, including the use of linear mixing
models.

2.3

Comparative Studies

Several studies have been carried out comparing mixture modelling with alternative techniques such as artificial neural networks and fuzzy logic [4, 6, 2].
Neural networks, when properly trained, perform well, but do not provide clear
measures of solution quality or likely error bounds.

Chapter 3

Solution Methods Overview

3.1

The assumption of linearity

The linear mixing model is assumed in most work related to the problem of
estimating the proportions of different cover components occupying a single
pixel from the spectral signature of that pixel. This model is equivalent to
assuming that radiation scattered from the pixel is scattered by one component
only, i.e. photons only interact with one cover type. In general this is often a
good assumption and leads to a number of applications.
Note however, that several developments of this method need careful thought:
1. The signal must be linearly related to the component proportions. This
usually requires it to be linearly related to the power received by the
detector. Quantities such as radiance, DN (usually) and reflectance are
suitable parameters, but in general, quantities such as the radar backscatter intensity ( 0 ) in dB and the brightness temperature (TB ) are not. For
small dynamic ranges quantities which are strictly non-linear may be close
enough to linear to be dealt with by linear mixture modelling.
2. Signals produced by coherent scattering (e.g. radar, lidar) do not in general
satisfy the required linear dependence on component proportion.
3. Shade: shade is inherently a product of multiple scattering and is therefore
unlikely in general to be measurable by this technique.
4. Variable strength: the linear method can solve for either variable areas of
fixed spectra (the usual land-cover application), or variable strengths of
given spectra for fixed areas. The two effects cannot however be separated.

3.2

Constraining the solution

Information in addition to the measured signals in each band is often used to

improve the estimated solution. This information may concern only individual
pixels or it may be applied to groups of pixels so that neighbouring pixels help
determine the values estimated for any individual pixel. Constraints are usually
applied to force the solution to be more realistic in some sense.

It should be noted that constraints can be applied in different ways. Constraints can be applied so that they are satisfied exactly. For example, the
equations can be rewritten to ensure that the sum of components is exactly
equal to 1 (Sections 4, 5; the number of classes to solve for is then reduced by
1). Alternatively, the constraint can be applied on the same basis as the agreement between the data and modelled signals is required, i.e. small departures
from exactly satisifying the constraint are allowed (an example of this is referred
to below as the Lagrange multiplier method).

3.2.1

Constraints applied to pixels individually

There are two principal constraints which are applied to individual pixels:
1. No proportion value should be less than 0.
2. The sum of proportions should be equal to 1.
The first of these is usually applied iteratively (as is an option in MIXMOD)
while the second can be applied either exactly (by modifying the basic linear
mixing model equations) or approximately (treating the constraint on the same
basis as the data). A third constraint, that no proportion value should be greater
than 1, is implicit if both the above constraints are applied, and is rarely applied
explicitly.

3.2.2

Constraints applied to groups of pixels

Constraints which apply to groups of pixels, usually groups of contiguous pixels, are described by the general term regularisation, and are discussed by [24]
and [15]. Regularisation methods allow a priori information concerning spatial
properties of the solution (e.g. the size of inter-pixel variations, or the (presumably low) likelihood of large departures from an area mean) to be incorporated
into the solution.

3.3

Testing solution quality

Any serious attempt to solve for the pixel proportions must address the issue of
checking the quality of the solution obtained. This is usually done by comparing
the discrepancies between the modelled and measured signals in each band with
the level of discrepancy which can be explained using known uncertainties in the
measurement system. Several methods of making the comparisons are available.
Residuals (the actual differences between the measured and modelled signals)
are used by some authors (e.g. [22]), while others use quasi or fully quantitative
statistical measures based on the conventional 2 goodness-of-fit parameter (see
Section 6 of this report). However this task is done, it is a crucial step.

3.4

Solution error bounds

If a solution is of acceptable quality, the next step to make the solution estimate
usable is to estimate the level of uncertainty of the solution. If there is no
knowledge of the uncertainty of the solution then the solution is useless in
8

Constraint(s)
applied
None
P
xj = 1

No end-member
uncertainty
Pnc1

(a) xnc = 1 j=1 xj

(b) Lagrange multipliers

x
Pj 0 (a)
xj = 1 & xj 0

OK
OK
ok
OK
OK

End-member uncertainty
treatment method:
Approximate
Exact
OK
OK
OK
OK
ok
ok
OK
OK
OK
OK

Table 3.1: Different solution methods for linear mixture modelling. (OK indicates that this case has been implemented fully in program MIXMOD (current
version = 0.12), and ok that a partial implementation is available, i.e. excluding dedicated error analysis.) The various solution methods referred to are
explained in the text.

practice. The main quantitative approaches to estimating error bounds are (1)
a Monte Carlo method, and (2) an analytical method. The Monte Carlo method
simulates several different sets of measurements with different sets of random
noise added, and then looks at the spread of solutions obtained. The analytical
method uses the curvature of the cost function minimum at the solution to
estimate error bounds and covariances. Both methods assume that the noise
distributions are known. The Monte Carlo method can deal more easily with
general noise models while the analytical method is more effective at revealing
the factors which make the measurements susceptible to noise.

3.5

The MIXMOD program

Table 3.1 desribes the solution methods implemented in program MIXMOD

(nc = number of classes). The current version of the program runs under the
MS-DOS operating system, and is written in C to be relatively portable.
MIXMOD has been applied to a range of linear mixing problems ranging
from the standard mixed pixel analysis [29] to atmospheric composition profiling and the analysis of fluorescence spectra, illustrating the common ground
between these different application areas.

10

Chapter 4

Solution Methods Ignoring

End-Member Noise
Most of the methods reported so far in the literature completely ignore the contribution of end-member noise. Section 5 and 6 of this report include methods
of handling end-member uncertainty. This section describes a variety of methods for dealing with the simpler case which ignores end-member uncertainty. A
range of standard methods is available to solve for the pixel proportions.
All the methods reported here are based on a goodness-of-fit measure G. G
is a sum of the squared differences between the measured (yi ) and modelled (
yi )
signals in each band (label i), normalised by a quantity i .

G =
yi

nb
2
X
(yi yi )
i=1
nc
X

i2
aij xj

(4.1)
(4.2)

j=1

where aij is the array of end-member reference spectra, and xj is the proportion of component j contributing to the (modelled) signal.
Depending on the definition of i , different types of solution (including those
used in most published work) are obtained.
1. i is constant, independent of signal band (i.e. all bands are weighted
equally): the method repeats the standard least squares solution.
2. i varies with signal band, usually according to the level of noise present
with the signal, but only the relative weights are controlled: gives a
weighted least squares solution but the goodness-of-fit parameter is not
quantitative.
3. i is set equal to the expected noise standard deviation in each band,
and the noise is close to Gaussian: a weighted least-squares solution is
obtained together with a quantitative measure of the goodness-of-fit.
Note that in certain cases, e.g. noise level set correctly and the same in all
bands, the different cases may be equivalent.
11

4.1

No constraint on the sum of components

It is not necessary to apply any constraint on the sum of components, although

this is not usually the approach taken in published work.
A solution is obtained by requiring the derivative of the goodness-of-fit parameter G to be zero with respect to each of the components xj .

nb
X
(yi yi ) yi

i2

i=1

xj

(4.3)

Pnc
nb
X
(yi k=1 aik xk )
aij
i2
i=1

(4.4)

nb
X
1
2 yi aij

i=1 i

nc X
nb
X
1
2 aij aik xk

i=1 i

(4.5)

Bj

k=1
nc
X

Cjk xk

(4.6)

k=1

where
Cjk

nb
X
1
aij aik
2
i=1 i

(4.7)

Bj

nb
X
1
yi aij
2
i=1 i

(4.8)

The solution is obtained by matrix inversion:

x

C 1 B

(4.9)

Equations written in matrix form

Using conventional matrix notation, the equations may be written as
)T W (y y
)
G = (y y
= Ax
y

(4.10)
(4.11)

W is the matrix of weightings (a diagonal matrix if all the bands are independent, with wii = 12 , wij = 0 (i 6= j)), and A is the matrix of end-member
i
reference spectra. The solution is then
x =

4.2

(AT W A)1 AT W y

(4.12)

Sum of components equal to 1

This is a natural constraint to apply since in practice the sum of components

should be exactly one (if the end-members are correctly chosen) and the problem
can be reduced in scope by one dimension using this constraint.
12

4.2.1

Forced exact compliance: xnc = 1

Pnc1
j=1

xj

The simplest way to apply this constraint is to express one component, typically
the last component, in terms of all the other components. A disadvantage is
that the method loses the explicit symmetry between all the components.
A solution may be obtained by requiring the first derivative of the goodnessof-fit parameter to be zero. This gives a matrix equation which can be solved by
matrix inversion methods (e.g. Singular Value Decomposition (SVD) or GaussJordan [15] - SVD methods are generally preferred because they can be made
more robust and also make the sensitivity of the solution to noise more explicit).
nc1
nb
XX
k=1 i=1

1
(aij ainc )(aik ainc )xk
i2
nc1
X

Cjk xk

nb
X
1
(4.13)
)
2 (aij ainc )(yi ainc

i=1 i

Bj

(4.14)

k=1

where

Cjk

nb
X
1
2 (aij ainc )(aik ainc )

i=1 i

Bj

nb
X
1
2 (aij ainc )(yi ainc )

i=1 i

(4.15)
(4.16)

These equations may be solved for the nc 1 component fractions xj (j =

1, 2, . . . nc 1). Note that the value of the final component then has to be
Pnc1
obtained from xnc = 1 j=1 xj .
x

4.2.2

C 1 B

(4.17)

Lagrange Multiplier Method

Lagrange multiplier methods can incorporate the contraint on the sum of the
components on the same basis as requiring that the signal in a particular band
is well-modelled. An extra band row is provided for the end-member data
to generate the sum of the component proportions. The corresponding data
value is 1, added to the signal input. The band measurement noise chosen acts
as the Lagrange multiplier and sets the scale of permitted deviation of the sum
of the components from 1.
Lagrange multiplier methods appear to be appropriate to be based on methods with no other constraints applied to the sum of components (or else double
constraints are being applied for the same feature of the problem).

4.3

All components positive

The current method is to check the first solution for any components which are
negative. Any that are found are set to zero, and the problem is re-run using
13

a restricted set of components. The process continues until no components are

negative.
A development of this constraint for the case without the sum of components
required to equal 1 is to test the solution for component values set above 1.

4.4

Multiple constraints

The constraints on the sum of components and their permitted range can be
applied independently.
The constraint on components not exceeding 1 could be applied independently, although once one component is set to 1 all the others must be zero (if
the constraint on the sum of components is also active).

14

Chapter 5

Solution Methods Allowing

for End-Member Noise
Uncertainty in the end-member measurements is probably the most important
source of error in the solution estimates. Such uncertainty is not easily handled
by the conventional methods reported in Section 4, and so two different methods
of handling end-member uncertainty have been developed and are presented
below. Both methods gave identical results in numerical simulations carried out
to test them. This more general problem is not yet dealt with in the literature.

5.1

Approximate Methods

An iterative method based on the case with no end-member noise is possible, and
is identical except for the evaluation of the expected noise standard deviation
in each signal band. Iterations repeat until the solution values for successive
iterations differ by less than some preset threshold. See section 2.5 of [15] for
more discussion of iterative methods for improving solutions to sets of equations.
The sections above describe how the different constraints may be applied.

5.2

Exact Methods

The general problem requiring solution is to minimise G (the goodness of fit

parameter) which is the sum of the squared normalised errors over the number
of bands (nb). This is a non-linear minimisation problem, discussed in Chapter
15 of [15].

G =

nb
2
X
(yi yi )
i=1

i2

(5.1)

where the modelled signal (

yi ) and variance (i2 ) in band i are given by
yi

nc
X

aij xj

(5.2)

j=1

15

i2

nc X
nc
X

xk xl aik ail + 2

k=1 l=1

nc
X

xk ni aik + ni ni

(5.3)

k=1

In general the cross-product ni aik may be ignored.

If the constraint on the sum of the components being equal to 1 is applied by
writing the last class component as the complement of the other components,
then the required solution (xj ) is a vector in nc1 dimensions, and the modelled
signal and variance may be written

yi

ainc +

nc1
X

xj (aij ainc )

(5.4)

j=1

i2

n2i + 2

nc1
X

xk [eik,inc einc,inc ] + einc,inc

k=1

nc1
X nc1
X

xk xl [eik,il einc,il eik,inc + einc,inc ]

(5.5)

k=1 l=1

where the end-member array statistics are written as

eik,jl

= aik ajl

(5.6)

Note that eik,jl = ejl,ik .

The standard method of solution in multiple dimensions requires the ability
to evaluate the function and its n-dimensional derivative.
G
xj

nb
nb
2
X
(yi yi ) yi X (yi yi ) i2

i2
xj
i4
xj
i=1
i=1

(5.7)

To evaluate the effect of measurement errors the second derivative may also
be required:

nb
nb
2
X
2 i2
1 yi yi X (yi yi )
2 i2 i2

2 xk xj
i4
xk xj
i2 xj xk
i=1 i
i=1

2
nb
X
(yi yi )
1 yi i2
1 i2 yi
yi
2
2
2
(5.8)
i2
xk xj
i xj xk
i xj xk
i=1
2

nb
nb
2
X
G
1 yi yi X (yi yi )
2 i2 i2
2 i2
E
' 2

2
(5.9)
xk xj
2 xk xj
i4
xk xj
i xj xk
i=1 i
i=1
2G
xk xj

= 2

This expression is approximately correct in the sense that terms in odd powers of the difference between the measured and modelled signals have been
ignored. (If the errors are random, which they should be for a good model,
then this is a good approximation for the purposes of evaluating typical error
bounds.)
16

The even powers of the difference between the measured and modelled signals
are related to the expected signal variance.

nb
!
X (yi yi )2
i=1

i2

= N umber of degrees of f reedom

(5.10)

= nb + n0 nc

(5.11)

(Writing n0 for the number of constraints applied.) Contributions from

different bands are independent, therefore

(yi yi )2
i2

nb + n0 nc
nb

(5.12)

The expression for the expected value of the second derivative of G is therefore

2G
xk xj

= 2

nb
X
1 yi yi
2 x x

k
j
i=1 i

nb
2(nb + n0 nc) X 1 1 2 i2
1 i2 i2

(5.13)
nb
2 2 xk xj
i2 xj xk
i=1 i

These expressions apply whether or not constraints are applied to the sum
of the components. The constraints affect the expressions used to evaluate the
derivatives of yi and i2 .

5.2.1

No constraint on sum of components

In this case, the derivatives are given by:

5.2.2

yi
xj

aij

i2
xj

2 i2
xk xj

nc
X

(5.14)
xk eij,ik

(5.15)

k=1

2eik,ij

(5.16)

Sum of components equal to 1

The required derivatives are more complicated if written out in full:

yi
xj

aij ainc

i2
xj

2 [eij,inc einc,inc ] + 2

2 i2
xk xj

(5.17)
nc1
X

xk [eik,ij einc,ij eik,inc + einc,inc

(5.18)
]

k=1

2 [eik,ij einc,ij eik,inc + einc,inc ]

17

(5.19)

An alternative may be to use the expressions for the derivatives in section

5.2.1 and then use a second stage of the partial derivatives to express the dependence of the final component on the remaining nc 1 components.

5.2.3

All components positive

The constraints on the permitted range of the components is implemented the

same way as for the other solution methods.

18

Chapter 6

Estimating Error Bounds

(This discussion is based closely on Chapter 15 of [15], to which reference should
be made for full details.)
The variation of 2 with the parameter values can be used to estimate error
bounds on the parameters (if assumptions about the model fit etc. are satisfied).
2

x [] x

(6.1)

where, identifying G with 2 ,

=
=

1 2 2
2 xj xk
1 2G
2 xj xk

(6.2)

The matrix [] thus needs to be evaluated at the solution point to estimate the expected error bounds. Different definitions apply according to the
constraints applied.
[] needs to be inverted to obtain the covariance matrix, which is then projected onto the sub-space of interest for the error bounds before being reinverted
to obtain the coefficients describing the error bound ellipsoid.
The chosen confidence level and number of degrees of freedom for the solution
determine the actual value used for 2 (the permitted increase in 2 relative
to its value at the minimum corresponding to the solution point) to evaluate
error bounds in a particular case.

19

20

Chapter 7

Developments of the
Standard Methods
It is possible to obtain greater accuracy or a wider range of parameters if the
signals from several pixels are used. The penalty for using these methods is that
there is some loss of spatial resolution.

7.1

Reducing uncertainty on estimated proportions

If N independent measurements of pixel proportions are made then the error

standard deviation on the mean is 1N of the error on a single pixel. Since the
underlying model is linear, it is possible to average the signals and then carry
out the inversion rather than invert the signals for many pixels and then average
the resulting proportions.
To estimate the uncertainty on the proportions estimated for the collection
of pixels it is necessary to use variances appropriate to the number of pixels
involved, i.e. the measurement uncertainty variance is N1 that of a single pixel.
The end-member uncertainty values must also be scaled appropriately. It may
be more convenient to simply scale the variance for a single pixel by the factor
1
N.

7.2

Estimating spectral signatures using known

proportions

If the proportions are known then it is possible to use m pixels to estimate the
spectral signature in n bands, averaged over the m pixels, as long as m > n.
It is assumed that the m pixels are independent in the sense that they provide
independent items of information, i.e. the signal for any one pixel is not a linear
function of the signals of a (sub-)set of the other m 1 pixels.
The result applies to the area covered by the pixels, i.e. there is a loss of
spatial resolution.

21

22

Chapter 8

Conclusions
This report describes the solution algorithms used by the program MIXMOD to
analyse mixed pixel data, assuming the linear mixing model. A unique feature
of the program is its ability to handle uncertainty in the assumed end-member
spectra. MIXMOD includes routines to perform the three principal analysis
steps: (1) estimate a solution, (2) evaluate the quality of the solution, and (3)
provide a measure of the error for the solution.
The literature reviewed gives an overview of current applications of linear
mixture modelling, as well as some pointers to the wider literature concerned
with linear systems. Several studies comparing mixture modelling with alternative methods of analysis are also identified. The principal strength of mixture
modelling relative to the alternatives is that because it is based on an explicit
model, its limitations can be clearly identified, and within those limitations
it is possible to quantify the goodness-of-fit and to evaluate appropriate error
bounds.
Although a large number of mixture modelling studies have been presented,
relatively few deal with issues of quantifying the solution quality or of estimating
proper error bounds. This report presents methods of performing these tasks,
taking account of all the principal sources of uncertainty. As the field of mixture
modelling matures, these issues must be addressed.

23

24

Bibliography
[1] Adams, J.B., Smith, M.O., and Gillespie, A.R. (1989), Simple models for
complex natural surfaces: a strategy for the hyperspectral era of remote
sensing. Proceedings of the IGARSS 89, Vancouver, Canada: IGARSS 89,
Vol. I, pp. 16-21.
[2] Atkinson, P.M., Cutler, M.E.J., and Lewis, H. (1995), Mapping sub-pixel
variation in land cover in the UK from AVHRR imagery. Proc. 21st Annual Conference of the Remote Sensing Society, 11-14 September 1995,
Southampton, UK. Published by the Remote Sensing Society, pp. 58-65,
ISBN 0 946226 20 0.
[3] Campbell, J.G. (1981) The Use of Landsat MSS Data for Ecological Mapping, Proc. 9th Ann. Conf. Remote Sensing Society, Univ. London, Dec.
1981. pp 143-161.
[4] Campbell, J.G., and A.A. Hashim (1992), Fuzzy Sets, Pattern Recognition,
and Neural Networks ..., Proc. 18th Ann. Conf. Remote Sensing Socy.,
Univ. Dundee, Sept. 1992 (AP. Cracknell and RA. Vaughan, eds.) pp. 508
-517.
[5] Cross, A.M. (1990). AVHRR as a data source for a GIS: deforestation in
Amazonia. Proceedings of the International Geoscience and Remote Sensing
Symposium, 1990. (IGARSS 90), Washington, D.C. (New York: IEEE),
pp.223-226.
[6] Foody, G.M., and Cox, D.P. (1994), Sub-pixel land cover composition estimation using a linear mixture model and fuzzy membership functions. Int.
J. Remote Sensing, 15(3), 619-631.
[7] Hobbs, S.E., and Thomas, G. (1996), A goodness-of-fit method applied to
spectral mixture modelling. Proc. Image Processing: Mathematical methods and applications. Conference of The Institute of Mathematics and its
Applications at Cranfield University, September 1994. Published by Oxford
University Press (in press).
[8] Holben, B.N., and Shimabukuro, Y.E. (1993), Linear mixing model applied
to coarse spatial resolution data from multispectral satellite sensors. Int.
J. Remote Sensing, 14(11), 2231-2240.
[9] Houghton, J.T., Taylor, F.W., and Rodgers, C.D. (1984), Remote sounding
of atmospheres, Cambridge University Press, Cambridge (ISBN 0 521 31065
2).
25

[10] Jolliffe, I.T. (1990), Principal components analysis: a beginners guide - I.

Introduction and application. Weather, 45, 375-382.
[11] Jolliffe, I.T. (1993), Principal components analysis: a beginners guide - II.
Pitfalls, myths and extensions. Weather, 48(8), 246-253.
[12] Kerdiles, H., and Grondona, M.O. (1995), NOAA-AVHRR NDVI decomposition and subpixel classification using linear mixing in the Argentinean
Pampa. International Journal of Remote Sensing, 16(7), 1303-1325.
[13] Lillesand, T.M., and RW. Kiefer, (1994) Remote Sensing and Image Interpretation, 3rd. ed., John Wiley.
[14] Novo, E.M., and Shimabukuro, Y.E. (1994), Spectral mixture analysis of
inland tropical waters. Int. J. Remote Sensing, 15(6), 1351-1356.
[15] Press, W.H., Teukolsky, S.A., Vetterling, W.T., and Flannery, B.P., (1992),
Numerical Recipes in C; Second Edition, Cambridge University Press:
Cambridge (ISBN 0 521 43108 5).
[16] Quarmby, N.A., (1992), Towards continental scale crop area estimation,
International Journal of Remote Sensing, 13: 981-989.
[17] Quarmby, N.A., Townshend, J.R.G., Settle, J.J., White, K.H., Milnes, M.,
Hindle, T.L. and Silleos, N., (1992). Linear mixture modelling applied to
AVHRR data for crop estimation. International Journal of Remote Sensing,
13:415-425.
[18] Roberts, D.A., Adams, J.B., and Smith, M.O. (1990), Transmission and
scattering of light by leaves: effects of spectral mixtures. Proc. IGARSS
90, Remote Sensing for the Nineties, Vol. II, 967-970.
[19] Rodgers, C.D. (1976), Retrieval of atmospheric temperature and composition from remote measurements of thermal radiation. Reviews of Geophysics
and Space Physics, 14 (4), November 1976, pp 609-624.
[20] Rodgers, C.D. (1990), Characterization and error analysis of profiles retrieved from remote sounding measurements. Journal of Geophysical Research, 95(D5), pp 5587-5595, 20 April 1990.
[21] Sabol, D.E., Adams, J.B., and Smith, M.O. (1990), Predicting the spectral detectability of surface materials using spectral mixture analysis. Proc.
IGARSS 90, Remote Sensing for the Nineties, Vol. II, 967-970.
[22] Sabol, D.E., Adams, J.B., and Smith, M.O. (1992), Quantitative subpixel
spectral detection of targets in multispectral images. J. Geophysical Research, 97 (E2), 2659-2672, 25 February 1992.
[23] Schanzer, D.L. (1993), Comments onThe least squares mixing models to
generate fraction images derived for remote sensing multispectral data.
IEEE Transactions on Geoscience and Remote Sensing, 31(3), May 1993.
[24] Settle, J.J., and Drake, N.A. (1993), Linear mixing and estimation of
ground cover proportions. Int. J. Remote Sensing, 14(6), 1159-1177.

26

[25] Shimabukuro, Y.E., and Smith, J.A. (1991), The least-squares mixing models to generate fraction images derived from remote sensing multispectral
data. IEEE Transactions on Geoscience and Remote Sensing, 29(1), 16-20,
January 1991.
[26] Smith, M.O., Johnson, P.E. and Adams, J.B. (1985). Quantitative determination of mineral types and abundances from reflectance spectra using
principal component analysis. Journal of Geophysical Research, 90:792-804.
[27] Smith, M.O., Adams, J.B., and Gillespie, A.R. (1990), Reference endmembers for spectral mixture analysis. Proc. 5th Australian Remote Sensing
Conference, 8-12 October 1990, Perth, Australia, Vol. I, 331-340.
[28] Smith, M.O., Adams, J.B., and Sabol, D.E. (1994), Spectral mixture analysis - new strategies for the analysis of multispectral data. Imaging Spectrometry - a Tool for Environmental Observations, Ed. J. Hill and J. Megier,
Kluwer Academic Publishers, Dordrecht, pp 125-143 (ISBN 0-7923-2965-1).
[29] Thomas, G, Hobbs, S E and Dufour, M J D (1996). Woodland area estimation by spectral mixing: applying a goodness-of-fit solution method.
International Journal of Remote Sensing, 17, pp 291-301.
[30] van Kootwijk, E.J., van der Voet, H., and Berdowski, J.J.M., (1995), Estimation of ground cover composition per pixel after matching image and
ground data with subpixel accuracy. International Journal of Remote Sensing, 16(1), 97-111.
[31] Zhukov, B., Berger, M., Lanzl, F., and Kaufmann, H., (1995), A new technique for merging multispectral and panchromatic images revealing subpixel spectral variation. Proceedings International Geoscience and Remote
Sensing Symposium 1995, IGARSS 95, Firenze, Italy, 10-14 July 1995,
Vol. III, IEEE.

27