Spectral Mixture For Remote Sensing: Yosio Edemir Shimabukuro Flávio Jorge Ponzoni

Download as pdf or txt
Download as pdf or txt
You are on page 1of 90
At a glance
Powered by AI
The document discusses topics related to remote sensing such as spectral mixture analysis, fraction images, mathematical algorithms for spectral mixture modeling, and sensors used in orbital remote sensing.

Some of the remote sensing techniques discussed include spectral mixture analysis, the use of orbital images, and the application of remote sensing to detect selective logging and monitor natural resources.

The fractions discussed in relation to spectral mixture analysis are the shade/water, soil, and vegetation fractions.

Springer Remote Sensing/Photogrammetry

Yosio Edemir Shimabukuro
Flávio Jorge Ponzoni

Spectral
Mixture
for Remote
Sensing
Linear Model and Applications
Springer Remote Sensing/Photogrammetry
More information about this series at http://www.springer.com/series/10182
Yosio Edemir Shimabukuro
Flávio Jorge Ponzoni

Spectral Mixture for Remote


Sensing
Linear Model and Applications
Yosio Edemir Shimabukuro Flávio Jorge Ponzoni
Remote Sensing Division Remote Sensing Division
National Institute for Space Research National Institute for Space Research
São José dos Campos, São Paulo, Brazil São José dos Campos, São Paulo, Brazil

ISSN 2198-0721     ISSN 2198-073X (electronic)


Springer Remote Sensing/Photogrammetry
ISBN 978-3-030-02016-3    ISBN 978-3-030-02017-0 (eBook)
https://doi.org/10.1007/978-3-030-02017-0

Library of Congress Control Number: 2018958512

© Springer Nature Switzerland AG 2019


This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, express or implied, with respect to the material contained herein or for any errors
or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims
in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword

Before introducing the book Spectral Mixture for Remote Sensing: Linear Model
and Applications, I feel impelled to speak about the authors and my relationship
with them. Both authors are of the utmost personal and intellectual integrity, pio-
neers in the implementation of remote sensing in Brazil. Both have solid academic
education from institutions of international renown. I had the privilege of being the
Master degree advisor of the first author, Yosio Shimabukuro, and have hired the
second author, Flávio Ponzoni, for the Remote Sensing Division of the National
Institute for Space Research (INPE). I spent most of my professional life with both,
as co-worker and scientific research collaborator. I followed their professional
growth and had the satisfaction to see that they exceeded the one who could be con-
sidered their master one day. They published several books and book chapters with
prestigious publishers, and they also published several scientific articles of interna-
tional impact, which elevated the first author to the 1A Researcher level of the
National Council for Scientific and Technological Development (CNPq), consid-
ered the highest recognition of a researcher by that institution.
The authors’ experience includes the development of several models for applica-
tion of remote sensing based on physical principles, aiming to the relevant applica-
tions for the knowledge and monitoring of Earth resources. This book describes in
detail a model that allows transforming spectral mixture – into useful information
for users. This book allows readers to navigate through the fundamentals of remote
sensing since its origin. It describes the main satellites, sensors, and spectral indices
used in the interpretation of images. It also presents the operational remote sensing
projects, as well as the physical principles, cartographic bases, features, and image
formats, to focus on linear spectral mixture model in detail.
This book presents the history of remote sensing since the use of aerial photo-
graphs to the modern concepts of information extraction exploring the various reso-
lutions of the images, with emphasis on the concept of spectral mixture and how to
decompose this mixture into its various fractions. It emphasizes, in a didactic way,
the applications of this model in large projects, especially in the Amazon region
where the dynamics of land use transformation is intense. Thus, the PRODES,
DETER, PANAMAZONIA, and AMAZONICA projects are presented for the esti-

v
vi Foreword

mation of deforested and burned areas, based on several spatial and temporal resolu-
tions of images, and presented at different scales.
By focusing on the main subject of the book, which is the decomposition of the
elements of images in biophysical components familiar to every researcher as veg-
etation, soil, and water/shade, the authors present different mathematical approaches
and justify the advantages of the linear model described in detail. The authors have
a large experience in the implementation and application of this model, since the
main author developed the linear spectral mixture model for the MSS and TM sen-
sors of the Landsat satellite, in 1987, in his Ph.D. program at Colorado State
University, USA. They describe the difficulties that users have working with “digi-
tal numbers” or “gray levels” of the images and the effect of the atmosphere in the
geophysical or biophysical objects characterization.
After the conceptualization of the mixture model, the authors describe how it can
be implemented in the major image processing systems such as SPRING, ENVI,
and PCI. The description resembles an operating manual, with actual examples of
fraction images of actual scenes. They emphasize the advantages of fraction images,
such as the facility of interpretation, instead of using gray levels analysis. They also
show how the model can be used to reduce the data dimensionality, for example, by
transforming data from several sensor bands into three fractions (vegetation, soil,
and water/shade).
This book progresses with a generic description of the main sensors/observation
platforms of the Earth and concludes with a detailed description of how the products
derived from the linear spectral mixture model are used in the operational project
analysis steps from INPE, which include image segmentation of fraction images,
classification, and matrix edition for presentation on various scales for dissemina-
tion of the results. The contribution of the fraction images was fundamental for the
automation of these projects. They also proved their use in burned areas mapping
and selective logging monitoring using the Landsat images in Mato Grosso state.
I am very confident that this book is a relevant contribution to the science and
applications of remote sensing.
Getúlio Teixeira Batista
Ph. D. in Remote Sensing and Agriculture, Purdue University, 1981.
Editor of Ambiente & Água – an International Journal of Applied Science
Environmental Science Post-Graduate School, University of Taubaté
Taubaté, SP, Brazil
Preface

The development of remote sensing techniques was marked by three moments. The
first of these was the improvement of photography that made possible the develop-
ment of photogrammetry and photointerpretation techniques. The second came
almost as an extension of these two forms of information extraction, now applied to
images generated by electro-optical sensors placed onboard orbital platforms (satel-
lites). Initially the applications explored approaches similar to those employed in
the extraction of information from photographs and, subsequently, motivated by the
recent discussions on climate change and the consequent needs to provide data on
greenhouse gas emissions and other contributors to global warming. The third
moment was characterized by more applications related to quantification of geo-
physical and biophysical parameters.
One of the most successful space programs dedicated to Earth observation is the
American Landsat program. It consists of medium spatial resolution sensors with
spectral bands strategically positioned in the electromagnetic spectrum to allow
data acquisition ensuring both minimal redundancy and periodicity (temporal
resolution).
The sensors onboard the Landsat satellite collect data in different regions of the
electromagnetic spectrum by using different sensors throughout a program started
in 1972 and which is continuing to this day: Multispectral Scanner System (MSS)
and Return Beam Vidicon (RBV) were onboard the Landsat 1, Landsat 2, and
Landsat 3 satellites; MSS and Thematic Mapper (TM) onboard the Landsat 4 and
Landsat 5 satellites; Enhanced Thematic Mapper Plus (ETM+) onboard the Landsat
7 satellite; and Operational Land Imager (OLI) onboard the Landsat 8 satellite.
These data are relayed to Earth receiving stations and processed into digitally
encoded images stored on computers.
The digital images are originally represented by digital numbers. These numbers
are defined in each element of spatial resolution (often called “pixel,” which is a
term originated from the words picture and element in English) according to the
intensity of radiant electromagnetic energy flow (radiance) that focuses on a detector
inside the sensor, which converts this intensity in an electrical signal that is c­ onverted
to a digital number proportional to this electrical signal. Whereas this flow of elec-

vii
viii Preface

tromagnetic energy comes from a portion of the Earth’s surface with pre-­established
dimensions, the resulting digital number represents in fact a measure proportional to
the “average” radiance of all objects inside this portion. Remember that this process
occurs in specific spectral bands and independently. Depending on the sensor system
and the altitude of the satellite, the spatial resolution of the image varies, since the
dimensions of this portion of observed surface (pixel) vary. For example, the spatial
resolutions of the MSS and TM sensors are approximately 0.45 hectares (57 m × 79
m) and 0.10 hectares (30 m × 30 m), respectively, on the Earth’s surface.
An important phenomenon to take into consideration is the fact that the radiance
that will lead to a digital number is a sum of integrated radiances from all objects or
materials contained within the field of view (IFOV) snapshot of the sensor (which
ultimately will lead to the pixel). Thus, the radiance effectively detected by the sen-
sor will be explained by spectral mixture of various materials “in” pixel added to the
atmospheric contribution. Therefore, the signal recorded by the sensor does not
represent the physical and chemical composition of any object exclusively.
This phenomenon of “spectral mixture” has been considered by some research-
ers (e.g., Horwitz et  al. 1971; Detchmendy and Pace 1972; Shimabukuro 1987).
Usually the problem arises when trying to classify correctly a pixel containing a
mixture of materials on the surface of the Earth such as soil, vegetation, rocks,
water, and others. The nonuniformity of most natural scenes usually results in a
large number of components in the mixture. The problem is further complicated by
the fact that the proportion of specific materials contained “inside” a pixel can vary
from pixel to pixel, generating varying degrees of ambiguity at the time of extract-
ing information.
The spectral mixture becomes more critical in the application of digital image
processing techniques than for the visual interpretation of images performed by
trained interpreters using the so-called elements of interpretation, which are color,
hue, texture, relative size, form, context, etc. The digital image classification is
based predominantly on the radiometric/spectral characteristics of the pixels and
then the classifier has the options: (1) classify the resolution element as a “pure”
pixel when, in fact, it can contain only a small percentage of the “pure” material or
(2) do not classify the pixel. The problem of spectral mixture is related to the prob-
lem of extracting “spectral signatures” or the spectral characterization of objects. To
minimize problems caused by spectral mixture, it is necessary to have a better
understanding of the effects of mixtures on pixel level.
This book presents the basic concepts that explain the spectral mixture, as well
as the development of methods that aim to find the solution for different studies
involving the application of remote sensing techniques. We will see that important
aspects of these methods take into account the spectral characterization of different
objects that compose the spectral mixture, and that based on this characterization it
will be possible to quantify the proportions of each component contained inside a
pixel. The solution of these methods is the fraction images, whose digital numbers
represent the proportions (or percentages) corresponding to each objects in the
mixture.
Preface ix

Thinking about environmental remote sensing or natural land resources, in


general, the spectral mixture in a pixel is formed by basic components such as soil,
vegetation, and shade. Then, after the solution of spectral mixture, we have the
fraction images of soil, vegetation, and shade. These fraction images have been used
in several areas of research such as those dedicated to forest resources, agriculture,
urban studies, and assessment of flooded areas.
Fraction images have been used in several studies in different areas of applica-
tion. In addition, these images have been used in projects to estimate deforested
areas in the Brazilian Legal Amazon (PRODES), the detection of deforested areas
in a near real time (DETER) also in Legal Amazon, and other projects such as the
PANAMAZONIA II and AMAZONICA in estimating burned areas in the Amazon
region. As we know, these projects analyze large areas on the ground through mul-
tispectral images with high temporal frequency. In that way, the fraction images that
reduce the volume of data highlighting the information required for these projects
are so important.
The purpose of this book is to offer to remote sensing users the opportunity to
know the main aspects of spectral mixture, and the potential of linear spectral
mixture model as a powerful tool for extracting information from remote sensing
products.

São José dos Campos, SP, Brazil Yosio Edemir Shimabukuro


Flávio Jorge Ponzoni
Contents

1 Background������������������������������������������������������������������������������������������������   1
1.1 Mixture Problem in the Pixel����������������������������������������������������������������  6
2 The Origin of the Digital Numbers (DNs)������������������������������������������������   9
3 Orbital Sensors������������������������������������������������������������������������������������������  17
3.1 MODIS������������������������������������������������������������������������������������������������  18
3.2 SPOT Vegetation ��������������������������������������������������������������������������������  18
3.2.1 Landsat MSS, TM, ETM+, and OLI��������������������������������������  21
3.3 Hyperion ��������������������������������������������������������������������������������������������  21
4 The Linear Spectral Mixture Model��������������������������������������������������������  23
4.1 Mathematical Algorithms��������������������������������������������������������������������  27
4.1.1 Constrained Least Squares (CLS)������������������������������������������  27
4.1.2 Four Spectral Bands and Three Components��������������������������  28
4.1.3 Six Spectral Bands and Four Components������������������������������  31
4.1.4 Weighted Least Squares (WLS)����������������������������������������������  38
4.1.5 Principal Components������������������������������������������������������������  39
4.2 Endmembers Selection������������������������������������������������������������������������  40
5 Fraction Images������������������������������������������������������������������������������������������  43
5.1 Error Images����������������������������������������������������������������������������������������  48
6 Fraction Images Applications ������������������������������������������������������������������  51
6.1 Monitoring of Deforestation ��������������������������������������������������������������  51
6.1.1 The Brazilian Legal Amazon Monitoring Program����������������  52
6.1.2 The Digital PRODES Project��������������������������������������������������  54
6.1.3 DETER Project ����������������������������������������������������������������������  58
6.2 Burned Areas Mapping ����������������������������������������������������������������������  62
6.3 Selective Logging Detection��������������������������������������������������������������  65
6.4 Land Use and Land Cover Mapping ��������������������������������������������������  65

xi
xii Contents

7 Final Considerations����������������������������������������������������������������������������������  69

References ����������������������������������������������������������������������������������������������������������  71

Index��������������������������������������������������������������������������������������������������������������������  77
About the Authors

Yosio Edemir Shimabukuro  is a Forest Engineer with a degree from the Federal


Rural University of Rio de Janeiro, a Master’s Degree in Remote Sensing from the
National Institute for Space Research (INPE), and a PhD in Forest Sciences and
Remote Sensing from Colorado State University (USA).
He works as a researcher in the Remote Sensing Division of INPE where, since
1973, he has been developing studies focused on the application of remote sensing
techniques in the study of vegetation. He has been responsible for the design and
improvement of methodologies for the identification and quantification of defores-
tation in the Amazon region, which have been applied in the PRODES and DETER
projects developed by INPE. He is a permanent teacher of the postgraduate course
in remote sensing of INPE.

Flávio  Jorge  Ponzoni  is a Forest Engineer, with a Master’s Degree in Forest


Sciences from the Federal University of Viçosa and a PhD in Forest Sciences from
the Federal University of Paraná. He completed a postdoctorate program at the
Center for Meteorological and Climatic Research Applied to Agriculture at the State
University of Campinas (CEPAGRI/UNICAMP), where he developed a work
focused on the absolute calibration of orbital sensors.
He works as a researcher of the Remote Sensing Division of the National Institute
for Space Research (INPE), where he is dedicated to studies of the spectral charac-
terization of vegetation and the development of methodologies focused on the abso-
lute calibration of orbital sensors. He also acts as a permanent teacher of the
postgraduate course in remote sensing of INPE.

xiii
Chapter 1
Background

Abstract  The main aspects related to the so-called spectral mixture under the per-
spective of orbital imagery carried out by Earth observation sensors are presented
and contextualized.

Keywords  Spectral mixture · Spatial resolution · Orbital imagery

Before further rationale for spectral mixture, it is necessary to understand the ori-
gins of all the concerns that have guided its concept in favor of the solution of sev-
eral problems in the application of remote sensing techniques. It is necessary, for
example, to know details about how effectively the intensity of electromagnetic
radiation reflected by a particular portion of the Earth’s surface is recorded by a sen-
sor, but, before that, it will be addressed the origins of the so-called spectral mixture
models.
Since the beginning of the application of remote sensing techniques, especially
when the orbital images were available, according to some researchers, the use of
multispectral data was limited, in part, by what they termed mixture problem, which
occurs in the radiant flux reflected by a particular portion of the Earth’s surface and
that is instantly viewed by a sensor. Later the technical terms will be adequately
defined and described that are applied to both the intensity of this flux and that por-
tion instantly viewed by a sensor. However, for now, consider the following: a sen-
sor is able to measure the intensity of radiant flux (of electromagnetic energy) of
portions of the Earth’s surface with defined dimensions. Typically these dimensions
are considered “square” and represented by metric quantities, as 20 m × 20 m, 80 m
× 80 m, 250 m × 250 m, and so on. Therefore, commonplace that portion instantly
observed by one sensor is called “resolution element.” When that resolution element
is represented in an image, it is called “pixel,” which is a word originated from the
merging of the two words in the English language: picture and element.
Here we will assume the terms “resolution element” and “pixel” as synonymous,
but the first term is related to the effective Earth’s surface portion from which the

© Springer Nature Switzerland AG 2019 1


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0_1
2 1 Background

intensity of the reflected electromagnetic energy flux is measured while the second
one is the numerical representation of this intensity in an image.
The mixture is related with different materials or objects contained “inside” the
resolution element (and “recorded” in the pixel) at the time of measuring the
­intensity of radiant flux by a sensor. The expression, “into the pixel,” which often
will be used in this book, assumes colloquial form, since, strictly speaking, there are
no objects “inside” of a pixel. Actually, we are talking about a situation in which
different objects are instantly viewed in an imaginary portion of the Earth’s surface
that has well-defined dimensions. The radiation flux originated by reflection of inci-
dent electromagnetic radiation is, in truth, a mixture of different fluxes of radiation
that will result in a single measure of intensity in each spectral region in which the
sensor is able to operate.
The concept of spectral mixture was discussed by Horwitz et  al. (1971),
Detchmendy and Pace (1972), Ranson (1975), and Heimes (1977), among others.
The spectral mixture can occur in two cases:
• When the materials (or objects) are smaller than the resolution element size: In
this case, the radiation flux detected by a sensor is composed by a mixture of dif-
ferent radiation fluxes reflected by different objects within the resolution
element.
• When the materials (or objects) are greater than the resolution element size: In
this case, the pixel overlaps the border between two or more materials or objects
larger than its size.

Fig. 1.1 Mixture
problems: (a) caused by
objects smaller than the
resolution element (pixel)
and (b) within the limits
(boundaries) of the
materials. (Source:
Shimabukuro (1987))
1 Background 3

In both cases, the signals recorded by the sensor are not representative of any of
the objects present inside the resolution element. The idealized representation of
mixture problem for both cases is illustrated in Fig.  1.1 that shows the objects
­scattered on the Earth’s surface delineated by continuous lines and the pixels on the
ground delineated by dashed lines.
Figure 1.2 shows schematically the mixture problem for the images generated by
three sensors with different spatial resolutions and five classes (or objects) on the
ground (a, b, c, d, and e). The sensor 3 presents the spatial resolution equal to a, the
sensor 2 presents the spatial resolution equal to 2a, and the sensor 1 presents spatial
resolution equal to 4a – that is, 10 m, 20 m, and 40 m, respectively. In that way, it
can be seen that the sensor 1 (lowest spatial resolution) does not present any pixel
with unique content (pure), the sensor 2 presents 5 pure pixels (1 of class b, 2 of
class d, and 2 of class e), and the sensor 3 presents 36 pure pixels (8 of class a, 6 of
class b, 12 of class d, 10 of class e, and 0 of class c).
The pixels painted in black color (Fig. 1.2) represent those occupied by only one
(pure) of the classes on the ground, while the others refer to those that present mix-
ture of classes with different proportions. Therefore the lower the sensor spatial

Fig. 1.2  Mixture for three sensors with different spatial resolutions and five classes (objects) on
the ground. (Source: Piromal (2006))
4 1 Background

Fig. 1.3  TM/Landsat 5 (R5 G4 B3) image of the Manaus (AM) region and a grid corresponding
to the size of the AVHRR pixels (1.1 km × 1.1 km)

resolution (i.e., the larger size of a pixel), the lower the chances of finding pure
pixels in it.
Figure 1.3 shows an actual example for the region of Manaus (AM). Pixels of the
AVHRR/NOAA (Advanced Very High Resolution Radiometer/National Oceanic
and Atmospheric Administration) sensor (1.1 km × 1.1 km) represented by the grid
over an image of the TM/Landsat 5 (Thematic Mapper on board of Landsat 5 satel-
lite) sensor (30 m × 30 m) are composed of a mixture of water, soil, and vegetation.
However, despite the coarser spatial resolution of the AVHRR sensor in relation to
the spatial resolution of TM sensor, it is possible to find pure pixels of water due to
the large extent of the Rio Negro.
Thus, the spectral characteristics of the sensor pixels such as the AVHRR/NOAA
(approx. 120  ha), MODIS/Terra (6.25  ha), the MSS/Landsat 4 (approximately
0.45 ha), and TM/Landsat 5 (approximately 0.10 ha) on the Earth’s surface can be
affected by one or by both phenomena described earlier.
For better understanding the spectral mixture on a pixel, one can imagine a scat-
ter plot similar to the one shown in Fig. 1.4.
In Fig. 1.4 the graph is composed, in X, by digital numbers or DN (whose origin
will be described later) referring to an orbital image generated in the red spectral
range and, in Y, the DNs generated in the near-infrared spectral range. For these two
spectral bands and considering the spectral properties of the main natural resources
(water, soil, and vegetation), the expected responses of these resources in this scatter
plot would be the pixels occupied by water and/or shadow positioned closer to the
origin of the graph, the pixels occupied by exposed soil positioned more distant to
1 Background 5

Fig. 1.4  Dispersion of the image pixels in the graph formed by the red and near-infrared spectral
bands

the origin as X as well Y, and, finally, the pixels occupied by green vegetation posi-
tioned next to Y, but far from X.
Therefore, pure elements or pure objects, i.e., pixels fully occupied by only one
of the natural resources considered, occupy pixels placed in the extreme of the tri-
angular figure formed in this scatter plot. Thus, in the far superior vegetation occu-
pies the entire pixels (in its whole area, or in 100%), while those placed on the far
right bare soil occupies the entire pixels.
So, there is the question: of what are composed the pixels placed at the center of
this triangular figure? The answer would be for equal proportions of the three natu-
ral resources, i.e., 33% of water, 33% of vegetation, and 33% of exposed soil.
Considering “a walk” from the border of this triangular figure leaving for example,
the vertex occupied by pure vegetation pixels, toward the origin of the scatter plot,
gradually the pixels will gain proportions of water or shade in your composition
until becoming the pure water or shade. Changing the direction of this path to the
far right, in the place of these pure pixels of water or shade, gradually would be
found pixels with higher proportions of soil until reaching pure pixels occupied by
exposed soil. Similarly, we could consider that various proportions of these pure
objects form the pixels within that triangular figure in the graph.
6 1 Background

The mixture problem was first considered and addressed by Horwitz et al. (1971),
who developed a technique to estimate the proportions of different types of land
cover within a resolution element (or a pixel). However, it was not actually used in
the remote sensing works until the middle of 1980 decade (Smith et al. 1985; Adams
et al. 1986; Shimabukuro 1987). Since then, the interest of spectral mixture (linear
and nonlinear) greatly increased, and several methods and applications have been
developed for several areas of study (Boardman 1989; Roberts et al. 1993; Atkinson
et  al. 1997; Bastin 1997; Foody et  al. 1997; Novo and Shimabukuro 1994;
Shimabukuro et  al. 1998; Rosin 2001; García-Haro et  al. 2005; Alcântara et  al.
2009).
This entire concept can be solved mathematically. For this, two approaches have
been adopted:
• Least-squares approach
• Parameters estimation using a maximum likelihood approach

1.1  Mixture Problem in the Pixel

The first applications of mixture models aimed at solving the problem of area esti-
mation by conventional methods of digital classification. Other way found in the
literature to use spectral mixture models are the applications involving images
derived from the proportions of the materials that compose the scene, which are the
main focus of this book.
The problem of mixtures is very critical for the classification of varied themes
related to land cover, limiting in any way the classification accuracy. Misclassification
may occur when an area seen by a multispectral sensor contains two or more themes
or classes of land cover, which produces a spectral response that does not match to
the characteristics of any one of these classes (Ranson 1975).
Two general approaches have been taken to address the mixture problem:
• The technique of classification
• Attempts to model the relationships between the types and proportions of one
class inside of a resolution element and the spectral response of this class (Heimes
1977)
In the first approach, two situations can occur:
• Classification of a pixel as a single class by using some decision function
• No classification of a pixel that do not have characteristic response of any one of
individual class, i.e., leaving that pixel as unclassified (Heimes 1977)
The second modeling approach is more complex since it tries to explain the
effects of the types and of the proportions of classes within a pixel related to its
spectral characteristic. Pearson (1973) and Ranson (1975) presented the simple
1 Background 7

least squares approach and some considerations for practical applications. Ranson
(1975) simulated the spectral characteristics or the spectral response of specific
mixtures of different objects in order to reduce the effects of the mixture problem in
digital classifications. Heimes (1977) evaluated the applicability of the least squares
approach (Pace and Detchmendy 1973) using a set of well-defined data. In Heimes
(1977) the observations and the proportions were obtained by simultaneous acquisi-
tion of radiometer data and photographic record of the scene.
Adams and Adams (1984) discussed the problem of separating the spectral
responses of vegetation from rock/soil when these materials are present in a pixel.
The goal was to extract information about rock/soils of pixels containing mixtures
of rock/soils and vegetation. The approach used by the authors was based on the
application of the linear model presented by Singer and McCord (1979). The authors
concluded that the use of this model was successful in two Landsat scenes: one MSS
image obtained in an area in Hawaii (USA) and another TM image obtained over the
mountains of Tucson, Arizona (USA).
Adams et al. (1986) discussed the spectral mixture modeling applied to an image
of the Viking Lander 1 sensor, which was one of the first sensors sent to Mars. The
basic assumption was that the main factor of the spectral variation observed in the
Viking 1 Lander image was the result of the linear mixture of materials present on
the surface and in the shadow. If this hypothesis is valid, then a limited number of
mixtures of spectra of the objects present in the scene (those that represent the prin-
cipal constituents in the image) can determine all other spectra of other objects of
the image, regardless of the instrumental calibration or the atmospheric effects.
Ustin et al. (1986), studying the applicability of data from TM/Landsat 5 sensor
for the vegetation of semiarid region, used the spectral mixture model (Adams and
Adams 1984; Adams et  al. 1986), and they identified four spectral signatures of
objects present in the scene. Those signatures were mixed in addictive mode to get
the best fit in order to estimate the response of other objects, in pixel-to-pixel basis.
The objects defined for their study were light soil, dark soil, vegetation, and shade
representing the topographic variations.
When working with spectral mixture models, these objects of interest taken as a
basis to estimate the responses of other objects are often called as “endmembers.”
In the 1980 decade, during his PhD course in the United States, Shimabukuro
(1987) developed and implemented the linear spectral mixture model applied to
orbital data (Landsat MSS and TM).
At the end of 1980 decade and in the early 1990, with the advancement of tech-
nology in computer science, spectral mixture models began to be implemented in
image processing systems, such as the image processing system (SITIM) and, sub-
sequently, the georeferenced information processing system (SPRING), developed
at the National Institute for Space Research (INPE). Later, similar models were
becoming available in commercial image processing systems. With the availability
of these models, large number of researchers and graduate students begun to explore
this area of research.
8 1 Background

In that way, the linear spectral mixture models have been used in various research
works, and they are currently important tools to operationalize the large-scale proj-
ects, such as the estimation and the monitoring of deforested areas and burning
areas in the Legal Amazon in digital approaches.
However, before to start studying details of these models, it is necessary to know
relevant aspects of the origin of DNs present in the multispectral and hyperspectral
images.
Chapter 2
The Origin of the Digital Numbers (DNs)

Abstract  The origin of the digital numbers in orbital images is presented from both
users’ and engineer’s point of views. This chapter aims to clarify important details
of the generation of the digital numbers, which keep the spectral mixture, the central
theme of this book.

Keywords  Digital number · IFOV · EIFOV · Radiometry

The digital numbers (DNs) present in orbital images or even those generated by
airborne sensors are generated according to a very simple principle: the
electromagnetic radiation reflected from the Earth’s surface and from the objects
follows toward  the sensor in a form of a flux with direction and intensity. That
intensity is named radiance and can be measured in different spectral wavelengths.
So, a flux of radiation reflected from the Earth’s surface contains different “types”
of electromagnetic radiation, differentiated themselves by wavelengths, which have
their own strengths, or radiances. After the reflected radiation interact with the
atmosphere during its trajectory toward  the sensor, its intensity is measured in
specific wavelength bands, according to the capacity of each sensor. These intensities
are converted into electrical signals by specific detectors that register the radiation
in specific spectral bands, and these electrical signals are converted into DNs by
specific criteria in each spectral band.
The DNs are numerical values proportional to radiance (intensity) values mea-
sured in different spectral bands. The relationship with radiance is straightforward.
The amplitudes of variation of the DNs are dependent on the number of bits adopted
on generation of DNs. That number of bits in reality is the base 2 exponent, so that,
if bits = 8, then 28 = 256, i.e., the DNs will vary from 0 to 255 (256 levels of inten-
sity or radiance). If bits = 10, so 210 = 1024, i.e., the DNs will vary from 0 to 1023
and so on. The amplitude of the DNs defines the radiometric resolution of the
sensor.
But how is the relationship between the radiance effectively measured by the
sensor and the corresponding DNs? These relationships are specific for each spectral
band in which the sensor is designed to operate and are usually expressed by linear

© Springer Nature Switzerland AG 2019 9


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0_2
10 2  The Origin of the Digital Numbers (DNs)

equations. Equation 2.1 expresses a linear relationship between the radiance


effectively measured by a sensor and DN:

DN λ = L0 λ .Gλ + offset λ (2.1)



where:
DNλ = DN value in spectral band λ
L0λ = radiance value effectively measured by the sensor in the spectral band λ
Gλ = angular coefficient of linear equation, also called gain, in spectral band λ
Offsetλ = linear equation intercept value, also called offset, in spectral band λ
Routinely L0λ receives the designation of apparent radiance or radiance at the top
of the atmosphere. In international literature, usually published in the English
language, this radiance receives the designation anachronistic TOA (top of
atmosphere), so it is common to find the term TOA radiance.
Some authors, such as Chander et al. (2010), treat Eq. 2.1 differently. The treat-
ment adopted by these authors is particularly interesting when those responsible for
the distribution of sensors information perform it through the dissemination of what
they call Lmin and Lmax. Many users are confused when presenting the regression line
coefficients, which are nothing more than the absolute sensor calibration coeffi-
cients in each spectral band. Lmin and Lmax are thus the minimum and maximum
radiances that the sensor is capable of measuring in a given spectral band. To make
it easier to understand, please refer to Fig. 2.1.
In Fig. 2.1 there are two distinct ways to show the relationship between L0λ and
DNλ. In the chart of Fig. 2.1a, it has DNλ as a function of L0λ, which reflects the most
physical point of view, that is, what is actually happening at the time of the acquisi-
tion of the data within the sensor. In the chart in Fig. 2.1b, it shows the point of view
of the remote sensing users when DNs are used to calculate the corresponding L0λ
values.

Fig. 2.1  Different forms of showing the relationship between L0λ and DNλ
2  The Origin of the Digital Numbers (DNs) 11

Note that in the relationship shown in the chart of Fig.  2.1a the offset values
would be represented by amounts of DNs, whereas, in the ratio shown in the chart
Fig. 2.1b the offset values would be represented by radiance units, which in this
particular case would be negative. This explains the negative values of Lmin
presented by Chander et al. (2007) describing the relationship between L0λ and DNλ.
Often, people who are not familiar with the relationships described herein are won-
dering how it would be possible to find negative L0λ values. It is now understood that
this is an algebraic peculiarity of the relationship between the two variables.
Chander et al. (2010), exploring the user’s point of view (relationship between
L0λ and NDλ as shown in the chart in Fig. 2.1b), deduced the calculation of Gλ and
offsetλ as expressed in Eqs. 2.2 and 2.3.

Lmax λ − Lmin λ
Gλ = (2.2)
DN max λ − DN min λ

 Lmax λ − Lmin λ 
offset λ = Lmin λ −   DN min λ (2.3)
 DN max λ − DN min λ 
So

 Lmax λ − Lmin λ 
L0 λ =  ( DN λ − DN min λ ) + Lmin λ (2.4)
 DN max λ − DN min λ 
or

L0 λ = Gλ .DN λ + offset λ (2.5)



where DNλ refers to the DN value in the spectral band λ, which will be converted to
apparent radiance L0λ.
Gilabert et al. (1994) presented a didactic discussion about the various influent
factors in L0λ, which are represented schematically in Fig. 2.2.
In this figure, vectors represented by letter E refer to different intensities of elec-
tromagnetic incident radiation coming from a source. Taking into account that this
intensity is called irradiance and that the source, in this case, is the sun (the main
source of electromagnetic radiation exploited in remote sensing techniques applied
to natural resources studies), we have E0λ representing the solar irradiance in the top
of the atmosphere. This irradiance E0λ then begins its trajectory through the atmo-
sphere toward the Earth’s surface. Considering a specific target on that surface, this
target actually receives both direct and diffuse fluxes of that incident radiation. Each
one of these fluxes has its own intensities, i.e., its own irradiances. Thus, Ebλ repre-
sents the incident irradiance directly on the target without interference of the atmo-
sphere, and Edλ represents the diffuse incident irradiance on the target. This so-called
diffuse radiation interacting with the atmosphere is scattered and hitting the target.
These two fluxes, both direct and diffuse, interact with the target, and part of both is
reflected toward space.
12 2  The Origin of the Digital Numbers (DNs)

Fig. 2.2  Influent factors in L0λ (Source: Adapted from Gilabert et al. (1994))

Other fractions of E0λ interact with the atmosphere in direct and diffuse ways and
do not even reach the target, having their trajectories altered back into space as rep-
resented by Lp and m. Besides that, other fractions of E0λ directly reach neighboring
targets, and its reflection flux, in Fig. 2.2 represented by a, is also directed to space.
The DNs contained in orbital images or even generated by airborne sensors are
then correlated to L0λ values, and therefore they cannot be directly associated with
the spectral characteristics of the targets that we are intending to extract some
information from. The spectral characterization of targets through the use of
airborne or orbital images depends on the conversion of DNs into physical variables
directly associated to the spectral properties of the targets, without the influence of
the atmosphere and neighboring targets.
The conversion of L0λ in DNλ is done in each spectral band in which the sensor was
designed to operate. This means that the translation of L0λ to DNλ is performed in a
particular way and meeting specific criteria in each spectral band. Thus, the same DN
value found in two or more images of different spectral bands may not represent the
same value of L0λ effectively measured by the sensor. In this fictitious case, an object
that should present differences in brightness levels in different spectral bands will
appear with the same DN value, thus misleading its “spectral characterization.”
Therefore, it is concluded that DN values from different spectral bands are not
meant to characterize objects spectrally and neither to perform arithmetic operations
between images of distinct spectral bands with the objective of associating the result
with some geophysical or biophysical parameter of the specific target.
This does not mean that arithmetic operations between DNs of images generated
in different spectral bands cannot be carried out. This depends on the goal the
users want to achieve. When, for example, the users want to only visually enhance
the objects or facilitate some sort of classification process (qualitative approach)
2  The Origin of the Digital Numbers (DNs) 13

that allows  them to identify objects as they  normally do in mapping jobs, such
arithmetic operations are feasible. The problem arises when the interest is to explore
the spectral differences of objects according to their spectral properties, since these
will not be represented by DNs.
The values of Gλ, offsetλ, Lminλ, and Lmaxλ are and should be made available by
those responsible for generating or distributing the sensor data. This information is
usually available in the form of metadata contained in specific files when accessing
the images or when they are acquired in digital format. Some organizations inform
these data on specific webpages. Of course the nomenclature adopted for these
coefficients is quite variable, and it is up to the user to recognize it with caution and
patience.
In the discussion presented by Gilabert et  al. (1994) concerning the influent
factors on the radiance L0λ illustrated in Fig. 2.2, other aspects related to the engi-
neering adopted by the sensor were disregarded, which effectively registers the
values of L0λ. It is necessary to consider that the radiation flux that will originate L0λ
and consequently its respective NDλ value is dependent on the spatial resolution of
the sensor that is pre-established.
Figure 2.3 presents a schematic representation of the geometric components par-
ticipants of the definition of spatial resolution of a sensor.
In Fig. 2.3, D refers to the dimensions of the detector, angular IFOV (Instantaneous
Field of View) is the angular representation that defines the surface of which the
value of L0λ is measured on the detector, F and H are respectively the focal length of
the sensor and the flight altitude, linear IFOV is the linear dimension designed of the
detector dimension D over the terrain, and EIFOV (Effective Instantaneous Field of
View) is the effective dimension of the spatial resolution element from which the
radiance L0λ is measured.
By the scheme presented in Fig. 2.3, it can be observed that the portion of the ter-
rain from which the radiant flux originates whose L0λ intensity will be measured is
often greater than that arising from the projection of D on the ground. It is worth say-
ing that when one is working with data from a sensor whose nominal spatial resolu-

Fig. 2.3 Geometric
components of the spatial
resolution. (Source: Slater
(1980))
14 2  The Origin of the Digital Numbers (DNs)

tion is 30 m (30 m × 30 m in the X and Y directions), effectively it should be less than
that, i.e., each spatial resolution element should have dimensions larger than 30 m.
Slater (1980) draws attention to this fact highlighting that one should not assume
IFOV, EIFOV, and pixel terms as synonyms, since the latter is only an existing
element in a numerical array in which a specific DN is placed.
The greater the differences between IFOV and EIFOV, it is understood that the
system will be generating greater spatial distortions, which can be quantified by
applying the modulation transfer function (MTF). More information about this
application can be found at Slater (1980).
There are other definitions or appropriations of the term IFOV that vary some-
what from that recommended by Slater (1980). Other designations for the term and
yet another term, FOV (Field of View), often associated with the angle of the optical
aperture of a sensor or the cross track angle of imaging are presented in Fig. 2.4.

Fig. 2.4  Some approaches involving the definition of the term IFOV
2  The Origin of the Digital Numbers (DNs) 15

It is worth noting that for the case of airborne or orbital sensors, the linear IFOV
has dimensions in both X and Y axes since it is also often assumed the square form
for the IFOV (or the EIFOV).
Whatever the size of the IFOV (or the EIFOV) and the way to define it (linearly
or angularly), it must be taking into account that within it there are objects of
different materials. The inclusion chance of objects with different spectral proper-
ties will be higher as large as the dimensions of IFOV (or EIFOV).
It should be noted that not only the proportion in the area of different objects
within a resolution element on the ground is important, but its spectral properties is
also important. It is important to remember that objects reflect, transmit, and absorb
the incident electromagnetic radiation in a particular way, so it is easy to understand
that each object contained within a resolution element on the ground will reflect
more or less than another within a certain spectral band. Thus, considering a hypo-
thetical situation in which within a resolution element there are four different objects
occupying the same proportions, the value of radiance effectively measured in a
given spectral band will be strongly influenced by the object that reflects more inten-
sively in that same spectral band. This explains why the users can observe, in orbital
images for example, objects whose dimensions are smaller than the IFOV of a sen-
sor such as a road crossing a forest, for example. Normally such roads are very nar-
row and in theory should not be detectable by sensors whose IFOV has dimensions
larger than the size of these roads. In this particular case, there is a very large radio-
metric contrast between the road reflectances (bare soil) and the vegetation cover in
its border. The same could happen with a bridge and the water in its surroundings.
So, the radiance effectively measured by the sensor ends up being “contami-
nated” by the higher radiance generated by the road or by the bridge in relation to
those generated by vegetation or water, respectively. This phenomenon is known as
“spectral mixture,” in which the spectral characteristic of a specific target prevails
over the spatial resolution of a sensor. In these cases the reflected radiances of the
road or of the bridge are greater than that of the forest and water, respectively, even
occupying a smaller portion of the resolution element. Figure 2.5 presents didactic
examples of such phenomenon.
In Fig. 2.5 tracings of a bridge and roads with dimensions lower than the IFOV
can be observed in orbital images of different spectral regions. This phenomenon is
very important when studying the mixing models because such proportions of
spectral mixture are closely linked to it.
Important aspects of the origin of the DNs have been presented. They are present
in sensor-generated images that feature specific characteristics from spectral,
spatial, and temporal points of view. These features obviously offer opportunities
16 2  The Origin of the Digital Numbers (DNs)

Fig. 2.5  Examples of possible object visualizations that present lower dimensions than those of
the IFOV

and constraints to attend different applications. They also cannot be neglected when
applying the linear spectral mixing model.
The following section presents some technical details of sensors whose data are
often used when applying linear spectral mixing models.
Chapter 3
Orbital Sensors

Abstract  The main technical characteristics of some more familiar sensors by the
remote sensing community are presented in order to allow users to compare the pos-
sibilities of application of the linear spectral mixture models.

Keywords  Earth observation sensors · Landsat · MODIS · HRV

The development of remote sensing techniques dates back to the end of the eigh-
teenth century assuming that it started with the first photographic cameras and it had
great momentum with the advent of the first space missions in the late 1950s and
early 1960s. Throughout this process, numerous orbital sensors were developed for
Earth observation that also initially aimed to generate data from the Earth’s surface.
So, through the data provided by sensors – such as those placed onboard satellites
of the Landsat program (MSS, RBV, TM, ETM+, OLI) and the SPOT (High resolu-
tion Visible, HRV) program – it was attempted to develop thematic maps, which
enabled monitoring the natural resources for decades.
Later, between the end of the 1980s and the early 1990s, the Earth Observing
System (EOS) program arose in the United States, which pushed for the develop-
ment and launching of sensors with the most varied capacities (resolutions). The
objective was to generate data from the Earth’s surface not only for mapping pur-
poses but also for the quantification of geophysical and biophysical parameters,
important in modeling studies and in varied forecasts, such as global climate change
and future availability of natural resources.
Data from the most varied orbital sensors are currently available, some of which
will be described in the following sections.

© Springer Nature Switzerland AG 2019 17


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0_3
18 3  Orbital Sensors

3.1  MODIS

The MODIS (Moderate Resolution Imaging Spectroradiometer) sensor was


launched in the Earth’s orbit by the United States in 1999 onboard the Terra Satellite
(EOS AM) and, in 2002, onboard the Aqua satellite (EOS PM). This sensor acquires
data in 36 spectral bands in the interval from 0.4 to 14.4 μm and in different spatial
resolutions (2 bands with 250 m, 5 bands with 500 m, and 29 bands with 1 km).
Together the instruments generate images from the whole Earth every day or 2 days.
They are designed to provide large-scale measurements of the global dynamics of
natural resources, including cloud cover changes, radiation balance, and processes
occurring in the oceans, on the continents, and in the low atmosphere. MODIS was
succeeded by the VIIRS (Visible Infrared Imaging Meter Suite) instrument onboard
the satellite Suomi NPP (Suomi National Polar-orbiting Partnership), launched in
2011, and it will be replaced by future JPSS satellites (Joint Polar Satellite System).
With its low and moderate spatial resolution but high temporal resolution,
MODIS data are useful for controlling landscape changes over time. Examples of
such applications include the monitoring of vegetation health through time series
analyses performed by vegetation indices (Lu et al. 2015), of the Earth’s long-term
cover changes to monitor deforestation rates (Klein et al. 2012; Leinenkugel et al.
2014; Lu et al. 2014; Gessner et al. 2015), of global snow cover trends (Dietz et al.
2012, 2013), and of flooding by virtue of precipitation, rivers, or floods in coastal
areas due to the rise of sea level (Kuenzer et al. 2015) and the change of water levels
of the Great Lakes (Klein et al. 2015), in addition to the detection and mapping of
the forest fires in the United States. The United States Forest Service Remote
Sensing Application Center (USFS) analyses MODIS images on a continuous basis
in order to provide management and forest fire information.
The technical characteristics of this sensor are shown in Table 3.1.
The images from spectral bands 1 to 7 are commonly used in the Earth’s surface
studies. Spectral bands 3–7 are resampled from 500 to 250 m and can be used in the
linear  spectral mixing models. Spectral bands 21 and 22 are generated from the
same wavelengths but present different saturation points.

3.2  SPOT Vegetation

The vegetation program is jointly developed by France, the European Commission,


Belgium, Italy, and Sweden. The first satellite of the program, vegetation 1, was
launched on 24 March 1998, onboard the SPOT 4 satellite, while the second instru-
ment, vegetation 2, was released on 4 May 2002 onboard the SPOT 5 satellite. They
provide data to monitor the Earth’s surface parameters with daily frequency on a
global basis with an average spatial resolution of 1 km. The ground segment associ-
ated with the program processes the data generated to provide standard products to
3.2  SPOT Vegetation 19

Table 3.1  Technical characteristics of the MODIS sensor


Spectral bands Wavelength (nm) Spatial resolution (m) Usage
1 620–670 250 Earth/cloud/aerosol limits
2 841–876 250
3 459–479 500 Earth/cloud/aerosol properties
4 545–565 500
5 1230–1250 500
6 1628–1652 500
7 2105–2155 500
8 405–420 1000 Ocean color/phytoplankton/
9 438–448 1000 biogeochemistry
10 483–493 1000
11 526–536 1000
12 546–556 1000
13 662–672 1000
14 673–683 1000
15 743–753 1000
16 862–877 1000
17 890–920 1000 Atmospheric water vapor
18 931–941 1000
19 915–965 1000
20 3.660–3.840 1000 Cloud/surface temperature
21 3.929–3.989 1000
22 3.929–3.989 1000
23 4.020–4.080 1000
24 4.433–4.498 1000 Atmospheric temperature
25 4.482–4.549 1000
26 1.360–1.390 1000 Cirrus water vapor
27 6.535–6.895 1000
28 7.175–7.475 1000
29 8.400–8.700 1000 Cloud properties
30 9.580–9.880 1000 Ozoniun
31 10.780–11.280 1000 Cloud/surface temperature
32 11.770–12.270 1000
33 13.185–13.485 1000 Cloud altitude
34 13.485–13.785 1000
35 13.785–14.085 1000
36 14.085–14.385 1000

the user community. The entire system complements the high-resolution spatial
capacity available on the sensors of the SPOT series (Satellite Pour la Observation
de la Terre) providing similar and simultaneous spectral measurements in the
20 3  Orbital Sensors

visible and infrared spectral regions. The original features of the instruments allow
users to have access to:
(a) Robust and simple multitemporal measurements of solar radiant properties,
continuous and global monitoring of continental areas, data generation for
regional or local studies, extensive data set with accurate calibration and local-
ization, and continuity and consistency, which will be made available by future
generations of these sensors
(b) Multiscale approaches
The decision to carry out this program was the result of the development of many
studies and projects during the last 20 years: the use of remote sensing data in opera-
tional programs or projects that should drive operating applications has increased
strongly until the availability and quality of the data became clearly a limitation.
As working groups, user communities and international programs have been
expressing their needs to increase the details in different domains (spectral, radio-
metric, temporal, and spatial). The idea of taking the opportunity to embarking on a
dedicated and definitely operational mission onboard SPOT 4 was supported by the
program’s partners. The needs of the European Commission’s sectoral policies for
the management of production in agriculture, for forestry, for environmental moni-
toring, and for national partners, as well as for major international programs related
to the study of global changes, have been synthesized by an international committee
of users and have become the basis for the technical development of the entire
system.
The strong commitment of the European Commission was also a clear sign that
the mechanisms by which remote sensing systems are designed and used are chang-
ing. Taking into account that the methodologies for using remote sensing data
become more adapted to a regular and operational need, the decision to undertake
such development is now also in the hands of users and is not only under exclusive
responsibility of the space agencies. The development of vegetation sensors, the
structure of the program, and its achievements constitute a test by which the new
mechanisms are exemplified. Their overall goals, however, should remain as a long-­
term commitment to provide useful data for the user community.
Table 3.2 presents the main technical characteristics of the vegetation 1 and veg-
etation 2 sensors.

Table 3.2  Technical characteristics of the vegetation 1 and vegetation 2 sensors


Spectral Spectral Spatial Temporal
Sensor band resolution resolution resolution SWATH
Vegetation 1 and B0 0.43–0.47 μm 1.15 km 24 h 2,250 km
Vegetation 2 B2 0.61–0.68 μm
B3 0.78–0.89 μm
SWIR 1.58–1.75 μm
3.3 Hyperion 21

3.2.1  Landsat MSS, TM, ETM+, and OLI

The Landsat satellite family began with the launch of Landsat 1 in 1972, which car-
ried onboard the one that would be the first successful Earth observation sensor: the
so-called Multispectral Scanner System (MSS). The Thematic Mapper (TM) sensor
began to be used 10 years later when, in 1982, it was placed onboard the Landsat
4 satellite.
The longest satellite of the program Landsat was the Landsat 5, also carrying
onboard the TM sensor, which was launched in 1984 and was discontinued in 2013.
This sensor generated images of excellent quality up to 2011, something unprece-
dented until then in terms of time period in orbit. The Landsat 7 satellite was
launched in 1999, taking the enhanced Thematic Mapper Plus (ETM+) sensor
onboard with the increase of a panchromatic band compared to its predecessor, the
TM Landsat 5.
The Landsat 8 brought innovations with the replacement of the TM sensor by the
Operational Land Imager (OLI) sensor characterized by the narrowing of spectral
bands and the inclusion of the coastal/aerosol, thermal, and cirrus bands.
Table 3.3 presents the characteristics of the TM, ETM+, and OLI sensors of the
Landsat program.
The radiometric resolution of the TM and ETM+ sensors is 8 bits, while the sen-
sor OLI is 12 bits. Their temporal resolution is 16 days.

3.3  Hyperion

The Hyperion sensor operates in 220 spectral bands between 0.4 and 2.5 μm, which
is why it is called hyperspectral. Launched in November 2000 as part of the EOS
program, it had as main objective to initiate the availability of a series of innovative
Earth observation sensors. The innovation was then focused on the generation of

Table 3.3  Technical characteristics of the TM, ETM+, and OLI sensors
Spectral bands TM (μm) ETM+(μm) OLI (μm) Spatial resolution
Coast/aerosol 0.433–0.453 30 m
Blue 0.45–0.52 0.45–0.52 0.450–0.515 30 m
Green 0.53–0.61 0.53–0.61 0.525–0.600 30 m
Red 0.63–0.69 0.63–0.69 0.630–0.680 30 m
Near infrared 0.78–0.90 0.78–0.90 0.845–0.885 30 m
Shortwave infrared 1.55–1.75 1.55–1.75 1.560–1.660 30 m
Thermal 10.4–12.5 10.4–12.5 120 m/60 m
Shortwave infrared 2.09–2.35 2.09–2.35 2.100–2.300 30 m
Panchromatic 0.52–0.90 1.360–1.390 15 m
Cirrus 0.52–0.90 30 m
22 3  Orbital Sensors

Fig. 3.1  Example of spectra that can be generated based on the Hyperion sensor data

spectral data that allowed the quantification of geophysical and biophysical param-
eters through the spectral characterization of the objects.
This sensor was placed onboard the satellite EO-1, which has sun-synchronous
orbit at altitude of 705 km, and it has an imaging push broom sensor with a width of
7.65 km, spatial resolution of 30 m, and a 12-bit radiometric resolution.
Ideally the spectral characterization of objects intended by the analysis of data
generated by this sensor should happen in pure pixels, that is, in those pixels where
the spectral mixture does not occur. However, it is known that even for miniature
dimensions of IFOV, there will always be the spectral mixture as a practically man-
datory phenomenon.
Figure 3.1 illustrates an example of spectra that can be generated based on the
Hyperion sensor data.
From what was previously exposed, it is clear that spectra such as those observed
in Fig. 3.1 do not refer to pure objects but rather to mixtures of objects contained in
pixels of 30 m × 30 m.
The conception of new sensors is a constant. There are countless innovations
planned for the next decades in the different areas explored by remote sensing tech-
niques (spatial, spectral, radiometric, and temporal). Regardless of innovations it is
certain that the remote sensing data-user community will have at its disposal an
immense amount of the Earth’s surface data. In addition, spectral mixing models
also serve as an alternative to reducing the volume of data to be processed aiming at
the information extraction process.
Chapter 4
The Linear Spectral Mixture Model

Abstract  The linear spectral mixture model is presented in its math concept.
During this discussion, examples are presented in order to facilitate the reader’s
understanding of the concepts involved in the design of the model.

Keywords  Linear spectral mixture model · LSMM concept · Spectral mixture


modeling

The time has come to understand a bit more how it is possible to calculate the frac-
tions or proportions of each “object” within a pixel, which can assume various
dimensions depending on the spatial resolution of each sensor.
These fractions are calculated by applying mathematical models. This chapter
will deal with linear models, but it is important to know that the linearity may not be
the only way to describe the contribution of each object within a pixel. The linear
spectral mixture models (LSMM) will be described because they have been widely
used by researchers with consistent results.
Following this approach, the spectral response in each pixel, in any spectral band
of the sensor, can be considered as a linear combination of the spectral responses of
each component present in the mixture. Thus, each pixel in the image, which can
assume any value within the gray level range (2n bits), contains information about
the proportion (amount) and the spectral response of each component within the
resolution unit on the ground. Therefore, for any multispectral image generated by
any sensor system, considering the knowledge of the components proportion, it will
be possible to estimate the spectral response of each one of these components.
Similarly, if this response is known, then the proportion of each component in the
mixture can be estimated. This characteristic will help in the analysis of different
sensors with diverse spatial resolution. For example, it is possible to generate the
fraction images from a sensor with high spatial resolution (pixels with small
dimensions) and, based on these proportions, to estimate the spectral responses of
the objects present in pixels generated by a medium spatial resolution sensor, and
then to generate the fraction images for this sensor images (Shimabukuro and Smith
1995).

© Springer Nature Switzerland AG 2019 23


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0_4
24 4  The Linear Spectral Mixture Model

Imagine a situation as simple as possible in which a panchromatic image is avail-


able, obtained in a relatively wide spectral range, generated from a sensor of 8-bit
radiometric resolution, i.e., with 256 gray levels. In this case, the system of equa-
tions of the linear spectral mixture model can be formulated as follows:

R = bx1 + px2 (4.1)



x1 + x2 = 1( sum of proportions must be equal to 1) (4.2)

where:
R = spectral response of the pixel of the image
b = spectral response of the bright object
p = spectral response of dark object
x1 = proportion of the bright object
x2 = proportion of dark object
In this way, it would be possible to generate two fraction images (bright and
dark) for this panchromatic image. The fraction images would be the solution of the
system of equations for all image pixels, as shown below.
Making:

x2 = 1 − x1 (4.3)

Replacing then Eq. 4.3 in Eq. 4.1,

R = bx1 + p (1 − x1 ) = bx1 + p − px1 = x1 ( b − p ) + p or R − p = x1 ( b − p ) (4.4)



As the image DNs range from 0 to 255, it can be considered that there is a bright
pure pixel (b = 255) and other dark pure pixel (p = 0). In this case, the pixels with
values between 1 and 254 would be a mixture of responses of these pixels. So, it
would be possible to generate two fraction images (bright and dark) for this
panchromatic image. The fraction images would be the result of the solution of
Eq. 4.4 for all pixels of the image.
For example:
For R = 0, substituting the values of the pixels (b = 255) and dark (p = 0) in Eq. 4.4,
it is obtained 0 − 0 = x1 (255 − 0) = 255x1 and x1 = 0 and, by Eq. 4.2, x2 = 1 − 0 = 1
(dark pixel).
For R  =  255, substituting the values of the pixels (b  =  255) and dark (p  =  0) in
Eq. 4.4, it is obtained 255 − 0 = x1 (255 − 0) = 255x1 and x1 = 1 and, by Eq. 4.2,
x2 = 1 − 1 = 0 (bright pixel).
For R  =  127, substituting the values of the pixels (b  =  255) and dark (p  =  0) in
Eq. 4.4, it is obtained 127 − 0 = x1 (255 − 0) = 255x1 and x1 = approximately 0.5
and, by Eq. 4.2, x2 = 1 − 0.5 = 0.5 (mixture pixel).
Making this for all the image pixels, the bright and dark fraction images were
obtained as the example for a real image shown in Fig. 4.1.
4  The Linear Spectral Mixture Model 25

Fig.4.1 (a) HRC CBERS panchromatic image; (b) Fraction image  color composite; (c) bright
fraction image; (d) dark fraction image

The resulting fraction images could be used together or individually, depending


on the objectives to be achieved. Although it has not been used normally, the linear
mixing model for two endmembers serves to show that the solution of this mixture
model is not so complex as will be shown in case of three or more endmembers in
the mixture.
Generally, then, the linear spectral mixture model can be written as:

R1 = a11 x1 + a12 x2 +  + a1n xn +e1


R2 = a21 x1 + a22 x2 +  + a2 n xn +e2
    
Rm = am1 x1 + am 2 x2 +  + amn xn + em

26 4  The Linear Spectral Mixture Model

or

Ri = sum ( aij x j ) + ei (4.5)



where:
Ri = average spectral reflectance for the i-th spectral band
aij = spectral reflectance of the j-th component in the pixel for the i-th spectral band
xj = value of the j-th component in the pixel
ei = error for the i-th spectral band
j = 1, 2, ..., n (n = number of components undertaken for the problem)
i = 1, 2, ..., m (m = number of spectral bands for the sensor system)
As mentioned previously, this model assumes that the spectral response (in
Eq. 4.5, expressed as reflectance) of the pixels are linear combinations of the spectral
response of the components within the pixel. To solve Eq. 4.5, it is necessary to have
the spectral reflectance of the pixels in each band (Ri) and the spectral reflectance of
each component (aij) in each band in order to estimate the proportion values, or vice
versa.
The linear spectral mixture model is a typical example of inversion problem
(indirect measures) in remote sensing. Some concepts of inversion problem and
three mathematical approaches to the solution of the system of linear equations will
be discussed below.
On the inversion problem, the mixing model without the relative term to the error
(ei), which was previously defined, can be rewritten in the matrix form:

R = Ax (4.6)

where:
A = matrix of m rows by columns containing input data, representing the spectral
reflectance of each component
R = vector of m columns, representing the pixel reflectance
x = vector of n columns, representing the proportion values of each component in
the mixture (variables to be estimated)
The procedure to solve a problem of remote sensing such as of Eq. 4.6 is called
the inversion problem or method of indirect measures. In this case, the average
spectral reflectance of the pixel (R) is assumed to be linear dependent of spectral
reflectance of each component (A). Therefore, the proportion value (xj) will be zero
if the respective aij and Ri are not dependent on each other.
Numerical inversions can produce results that are mathematically correct, but
physically unacceptable. It is important to understand that most of the problems of
physical inversion are ambiguous, since they do not have a unique solution and a
discreet reasonable solution is achieved by imposing additional boundary conditions.
4.1  Mathematical Algorithms 27

In remote sensing, the users are typically interested in knowing the state of a
physical quantity, biological or geographic (or several of them), such as the biomass
of a specific agricultural crop, the amount of a pollutant gas into the atmosphere or
the extent and the condition of the global coverage of snow in a given date.
For the solution of system of linear equations that represents the spectral mixture
model, there are several mathematical approaches based on the method of least
squares. Following it will be presented three algorithms that are available in the
current image processing software (SPRING, ENVI, PCI).

4.1  Mathematical Algorithms

As seen previously, the linear spectral mixture model is a system of equations, with
one equation for each considered sensor band. For example, for the MSS, there are
four equations, corresponding to the bands 4, 5, 6, and 7, while for the TM, there are
six equations, corresponding to the bands 1, 2, 3, 4, 5, and 7, taking into account
only the solar optical spectrum. It is important to keep in mind that it is not necessary
to use all the available bands, but it must obey the condition that the number of
reference spectra (or pure pixels) always be less than the number of spectral bands.
In this way, mathematical algorithms are required for the solution of the system of
equations formed by the spectral response of the pixel, which is the function of the
proportion of each reference spectrum (or pure pixels) weighted by the respective
spectral response of the endmember.
These reference spectra or pure pixels are routinely referred as endmembers, a
designation that the users more familiar with applying linear mixture models are
more accustomed. So from this point, it will be adopted that name.
Following three mathematical algorithms  will be presented: constrained least
squares (CLS), weighted least squares (WLS), and principal components (PC).

4.1.1  Constrained Least Squares (CLS)

This method estimates the proportion of each component within the pixel minimiz-
ing the sum of squared errors. The proportion values must be nonnegative (physical
meaning) and add to 1. To solve this problem, it was developed a method of quasi-­
closed solution (e.g., a method which finds the solution making approximations that
satisfy these constraints). In this case, the proposed method will be presented for the
cases of three or four components within the pixel. It is important to keep in mind
that this model can be developed for a greater number of endmembers but the
solution is becoming increasingly complex, as it will be seen between three and four
endmembers models. Thus, the mixture model can be written as:
28 4  The Linear Spectral Mixture Model

ri = a11 x1 + a12 x2 + a13 x3 + e1


    
rm = am1 x1 + am 2 x2 + am 3 x3 + em

It is possible to write as:

ri = ∑ ( aij x j ) + ei (4.7)

or,

ei = ri − ∑ ( aij x j ) (4.8)

The function to be minimized is:

F = ∑ ei2 (4.9)

where m is the number of sensor spectral bands used, for example, m = 4 for the
MSS sensor or m = 6 for the TM sensor.

4.1.2  Four Spectral Bands and Three Components

In this case, the mixture problem can be written as:

r1 = a11 x1 + a12 x2 + a13 x3 + e1



r2 = a21 x1 + a22 x2 + a23 x3 + e2

r3 = a31 x1 + a32 x2 + a33 x3 + e3

r4 = a41 x1 + a42 x2 + a43 x3 + e4

The function to be minimized is:

e12 + e22 + e32 + e42 = E1 x12 + E2 x22 + E3 x32 + E4 x1 x2 + E5 x1 x3


(4.10)
+ E6 x2 x3 + E7 x1 + E8 x2 + E9 x3 + E10
The values of the coefficients E1 to E10 are shown in Table 4.1.
Considering the first restriction, x1 + x2 + x3 = 1 or x3 = 1 − x1 − x2, and replacing
this restriction in Eq. 4.10, the function to be minimized becomes:

e12 + e22 + e32 + e42 = A1 x12 + A2 x22 + A3 x1 x2 + A4 x1 + A5 x2 + A6 (4.11)



4.1  Mathematical Algorithms 29

Table 4.1  values of the coefficients E of Eq. 4.10


E1 = a112 + a212 + a312 + a412
E2 = a122 + a222 + a322 + a422
E3 = a132 + a232 + a332 + a432
E4 = 2 × (a11 a12 + a21 a22 + a31 a32 + a41 a42)
E5 = 2 × (a11 a13 + a21 a23 + a31 a33 + a41 a43)
E6 = 2 × (a12 a13 + a22 a23 + a32 a33 + a42 a43)
E7 = − 2 × (a11 r1 + a21 r2 + a31 r3 + a41 r4)
E8 = − 2 × (a12 r1 + a22 r2 + a32 r3 + a42 r4)
E9 = − 2 × (a13 r1 + a23 r2 + a33 r3 + a43 r4)
E10 = r12 + r22 + r32 + r42

Table 4.2  Coefficients values of A of Eq. 4.11


A1 = a112 + a212 + a312 + a412 + a132 + a232 + a332 + a432 − 2× (a11 a13 + a21 a23 + a31 a33 + a41 a43)
A2 = a122 + a222 + a322 + a422 + a132 + a232 + a332 + a432 − 2× (a12 a13 + a22 a23 + a32 a33 + a42 a43)
A3 = 2 (r12 a132 + a232 + a332 + a432 + a11 a12 + a21 a22 + a31 a32 + a41 a42 − a11 a13 − a21 a23 − a31 a32 − a41
a43 − a12 a13 − a22 a23 − a32 a33 − a42 a43)
A4 = 2 (− a132 − a232 – a332 − a432 + a11 a13 + a21 a23 + a31 a33 + a41 a43 − a11 r1 − a21 r2 − a31 r3 − a41
r4 + a13 r1 + a23 r2 + a33 r3 + a43 r4)
A5 = 2 (− a132 − a232 − a332 − a432 + a12 a13 + a22 a23 + a32 a33 + a42 a43 − a12 r1 − a22 r2 − a32 r3 − a42
r4 + a13 r1 + a23 r2 + a33 r3 + a43 r4)
A6 = a132 + a232 + a332 + a432 + r12 + r22 + r32 + r42 − 2 × (a13 r1 + a23 r2 + a33 r3 + a43 r4)

The values of the coefficients A1 to A6 are shown in Table 4.2.


The function to be minimized is:

F = A1 x12 + A2 x22 + A3 x1 x2 + A4 x1 + A5 x2 + A6 (4.12)



In which the coefficients A1 to A6 are functions of the spectral values, aij (response
values of the endmembers) and ri (response values of the pixel).
To solve this problem, it is necessary to find a minimum value within the area
defined by lines: 0  ≤  x1  ≤  a, b  ≤  0  ≤  x2, and x1/a  +  x2/b  =  1, where a  =  b  =  1
(Fig. 4.2). Considering the function to be minimized, in order to find the minimum
value, the partial derivatives are calculated and equated to zero:

dF
= 2 A1 x1 + A2 x2 + A4 = 0
dx1

30 4  The Linear Spectral Mixture Model

Fig. 4.2  Region that meets


the restrictions for the
number of components
equal to 3

Table 4.3  Possible outcomes to the equation system solution


Outcome X1 X2 Inside the region Values to be recalculated X3
1 Positive Positive Yes 1 − x1 − x2
2 Positive Positive No x1 e x2 0
3 Negative Positive No x2 (x1 = 0) 1 − x2
3 Negative Negative No x1 = x2 = 0 1
3 Positive Negative No x1 (x2 = 0) 1 − x1

dF
= 2 A2 x2 + A3 x1 + A5 = 0
dx2

Solving for x1 and x2:


(
x1 = ( A3 A5 − 2 A2 A4 ) / 4 A1 A2 − A32 )


(
x2 = ( A3 A4 − 2 A1 A5 ) / 4 A1 A2 − A32 )
So, there are five possible outcomes (Table 4.3), which are described below.
Outcome 1  The minimum value is within the region of interest. So, this is the final
solution and x3 = 1 − x1 − x2.

Outcome 2  The minimum value is outside the region, and x1 and x2 are positive. In
this case, the minimum value wanted is searched on the line defined by x1 + x2 = 1
(i.e., x3 = 0). Now, making x2 = 1 − x1, the function to be minimized is:

F = ( A1 + A2 − A3 ) x12 + ( A3 + A4 − A5 − 2 A2 ) x1 + ( A2 + A5 + A6 ) (4.13)

The minimum value will be obtained by:
4.1  Mathematical Algorithms 31

dF
= 2 ( A1 + A2 − A3 ) x1 + ( A3 + A4 − A5 − 2 A2 ) = 0
dx1

So:

( A3 + A4 − A5 − 2 A2 )
x1 = −
2 ( A1 + A2 − A3 )

If x1 > 1, make x1 = 1 > 1, or, if x1 < 0, make x1 = 0 and x2 = 1 − x1.
Outcome 3  The minimum value is outside the region and x1 is negative and x2 is
positive. In this case, making x1 = 0, the function to be minimized becomes:

F = A2 x22 + A5 x2 + A6 (4.14)

Solving to find the minimum, x2 = −A5/2A2. If x2 > 1, then make x2 = 1, or, if x2 < 0,
make x2 = 0 and x3 = 1 − x2.
Outcome 4  The minimum value is outside the region and x1 and x2 are negative. In
this case, x1 and x2 are equated to zero and x3 = 1.

Outcome 5  The minimum value is outside the region and x1 is positive and x2 is
negative. In this case, making x2 = 0, the function to be minimized becomes:

F = A1 x12 + A4 x1 + A6 (4.15)

Solving to find the minimum, x1 = −A4/2A1. If x1 > 1, then x1 = 1, or, if x1 < 0, then
x1 = 0 and x3 = 1 − x1.

4.1.3  Six Spectral Bands and Four Components

In this case, the mixture problem can be written as:

r1 = a11 x1 + a12 x3 + a13 x3 + a14 x4 + e1



r2 = a21 x1 + a22 x3 + a23 x3 + a24 x4 + e2

r3 = a31 x1 + a32 x3 + a33 x3 + a34 x4 + e3

r4 = a41 x1 + a42 x3 + a43 x3 + a44 x4 + e4

r5 = a51 x1 + a52 x3 + a53 x3 + a54 x4 + e5

32 4  The Linear Spectral Mixture Model

r6 = a61 x1 + a62 x3 + a63 x3 + a64 x4 + e6



The function to be minimized is:

e12 + e22 + e32 + e42 + e52 + e62 = E1 x12 + E2 x22 + E3 x32 + E4 x42 + E5 x1 x2
+ E6 x1 x3 + E7 x1 x4 + E8 x2 x3 + E9 x2 x4
(4.16)
+ E10 x3 x4 + E11 x1 + E12 x2
+ E13 x3 + E14 x4 + E15
The values of the coefficients E1 to E15 are shown in Table 4.4.
Consider the first restriction: x1 + x2 + x3 + x4 = 1 or x4 = 1 − x1 x2 x3. Replacing
this restriction in Eq. 4.5, the function to be minimized becomes:

e12 + e22 + e32 + e42 + e52 + e62 = T1 x12 + T2 x22 + T3 x32 + T4 x1 x2 + T5 x1 x3


(4.17)
+T6 x2 x3 + T7 x1 + T8 x2 + T9 x3 + T10
in which the coefficients values of T1 to T10 are shown in Table 4.5.
Now, the approach to solve this problem is to find a minimum within the volume
defined by the plans: (a) 0 ≤ x1 ≤ a, 0 ≤ x2 ≤ b, and x1/a + x2/b = 1; (b) 0 ≤ x1 ≤ a,
0 ≤ x3, and x1/a + x3/c = 1; (c) 0 ≤ x2 ≤ b, 0 ≤ x3 ≤ c, and x2/b + x3/c = 1; and (d)
x1/a + x2/b + x3/c = 1 (Spiegel 1968), in which a = b = c = 1 (Fig. 4.3).

Table 4.4  Coefficients values E1 to E15 for Eq. 4.16


E1 = a112 + a212 + a312 + a412 + a512 + a612
E2 = a122 + a222 + a322 + a422 + a522 + a622
E3 = a132 + a232 + a332 + a432 + a532 + a632
E4 = a142 + a242 + a342 + a442 + a542 + a642
E5 = 2 (a11 a12 + a21 a22 + a31 a32 + a41 a42 + a51 a52 + a61 a62)
E6 = 2 (a11 a13 + a21 a23 + a31 a33 + a41 a43 + a51 a53 + a61 a63)
E7 = 2 (a11 a14 + a21 a24 + a31 a34 + a41 a44 + a51 a54 + a61 a64)
E8 = 2 (a12 a13 + a22 a23 + a32 a33 + a42 a43 + a52 a53 + a62 a63)
E9 = 2 (a12 a14 + a22 a24 + a32 a34 + a42 a44 + a52 a54 + a62 a64)
E10 = 2 (a13 a14 + a23 a24 + a33 a34 + a43 a44 + a53 a54 + a63 a64)
E11 = −2 (a11 r1 + a21 r2 + a31 r3 + a41 r4 + a51 r5 + a61 r6)
E12 = −2 (a12 r1 + a22 r2 + a32 r3 + a42 r4 + a52 r5 + a62 r6)
E13 = −2 (a13 r1 + a23 r2 + a33 r3 + a43 r4 + a53 r5 + a63 r6)
E14 = −2 (a14 r1 + a24 r2 + a34 r3 + a44 r4 + a54 r5 + a64 r6)
E15 = r12 + r22 + r32 + r42 + r52 + r62
4.1  Mathematical Algorithms 33

Table 4.5  Coefficients values of T1 to T10 for Eq. 4.17


T1 = a112 + a212 + a312 + a412 + a512 + a612 + a142 + a242 + a342 + a442 + a542 + a642 − 2 (a11 a14 + a21
a24 + a31 a34 + a41 a44 + a51 a54 + a61 a64)
T2 = a122 + a222 + a322 + a422 + a522 + a622 + a142 + a242 + a342 + a442 + a542 + a642 − 2 (a12 a14 + a22
a24 + a32 a34 + a42 a44 + a52 a54 + a62 a64)
T3 = a132 + a232 + a332 + a432 + a532 + a632 + a142 + a242 + a342 + a442 + a542 + a642 − 2 (a13 a14 + a23
a24 + a33 a34 + a43 a44 + a53 a54 + a63 a64)
T4 = 2 [(a11 a12 + a21 a22 + a31 a32 + a41 a42 + a51 a52 + a61 a62) + (a14
2
 + a242 + a342 + a442 + a542 + a642) − (a11 a14 + a21 a24 + a31 a34 + a41 a44 + a51 a54 + a61 a64) − (a12
a14 + a22 a24 + a32 a34 + a42 a44 + a52 a54 + a62 a64)]
T5 = 2 [(a11 a13 + a21 a23 + a31 a33 + a41 a43 + a51 a53 + a61 a63) + (a14
2
 + a242 + a342 + a442 + a542 + a642) − (a11 a14 + a21 a24 + a31 a34 + a41 a44 + a51 a54 + a61 a64) − (a13
a14 + a23 a24 + a33 a34 + a43 a44 + a53 a54 + a63 a64)]
T6 = 2 [(a12 a13 + a22 a23 + a32 a33 + a42 a43 + a52 a53 + a62
a63) + (a142 + a242 + a342 + a442 + a542 + a642) − (a12 a14 + a22 a24 + a32 a34 + a42 a44 + a52 a54 + a62
a64) − (a13 a14 + a23 a24 + a33 a34 + a43 a44 + a53 a54 + a63 a64)]
T7 = −2 [(a11 r1 + a21 r2 + a31 r3 + a41 r4 + a51 r5 + a61
r6) + (a142 + a242 + a342 + a442 + a542 + a642) − (a11 a14 + a21 a24 + a31 a34 + a41 a44 + a51 a54 + a61
a64) − (a14 r1 + a24 r2 + a34 r3 + a44 r4 + a54 r5 + a64 r6)]
T8 = −2 [(a12 r1 + a22 r2 + a32 r3 + a42 r4 + a52 r5 + a62
r6) + (a142 + a242 + a342 + a442 + a542 + a642) − (a12 a14 + a22 a24 + a32 a34 + a42 a44 + a52 a54 + a62
a64) − (a14 r1 + a24 r2 + a34 r3 + a44 r4 + a54 r5 + a64 r6)]
T9 = −2 [(a13 r1 + a23 r2 + a33 r3 + a43 r4 + a53 r5 + a63
r6) + (a142 + a242 + a342 + a442 + a542 + a642) − (a13 a14 + a23 a24 + a33 a34 + a43 a44 + a53 a54 + a63
a64) − (a14 r1 + a24 r2 + a34 r3 + a44 r4 + a54 r5 + a64 r6)]
T10 = r12 + r22 + r32 + r42 + r52 + r62 + a142 + a242 + a342 + a442 + a542 + a642 − 2 (a14 r1 + a24 r2 + a34
r3 + a44 r4 + a54 r5 + a64 r6)

Fig. 4.3  Region that meets


the restrictions for the
number of components
equal to 4

The function to be minimized is:

F = T1 x12 + T2 x22 + T3 x32 + T4 x1 x2 + T5 x1 x3


(4.18)
+T6 x2 x3 + T7 x1 + T8 x2 + T9 x3 + T100
34 4  The Linear Spectral Mixture Model

To find the minimum, the partial derivatives are calculated and equated to zero:

dF
= 2T1 x1 + T4 x2 + T5 x3 + T7 = 0
dx1

dF
= 2T2 x2 + T4 x1 + T6 x3 + T8 = 0
dx2

dF
= 2T3 x3 + T5 x1 + T6 x2 + T9 = 0
dx3

Solving for x1, x2, and x3, the system of linear equations can be formulated as:

2T1 T4 T5 x1 −T7
T4 2T2 T6 x2 = −T8
T5 T6 2T3 x3 −T9

Using a numerical method to solve the system of linear equations (Burden et al.
1981; Conte and De Boor 1980), the minimum not restricted can be obtained. Thus,
there are nine possible outcomes (Table 4.6).
The procedure to calculate the values of x1, x2, x3, and x4 is therefore as follows:
Outcome 1  0 ≤ x1 ≤ 1, 0 ≤ x2 ≤ 1, 0 ≤ x3 ≤ 1, and x1 + x2 + x3 ≤ 1, that is, the mini-
mum is within the region of interest. So, this is the final solution and x4  =  1
− (x1 + x2 + x3).

Table 4.6  Possible outcomes or the solution of system of equations


Inside the Values to be
Outcome x1 x2 x3 region recalculated x4
1 Positive Positive Positive Yes – 1 − x1 − x2
− x3
2 Positive Positive Positive No x1, x2, x3 0
3 Negative Positive Positive No x2, x3 (x1 = 0) 1 − x2 − x3
4 Positive Negative Positive No x1, x3 (x2 = 0) 1 − x1 − x3
5 Negative Negative Positive No x3 (x1 = x2 = 0) 1 − x3
6 Positive Positive Negative No x1, x2 (x3 = 0) 1 − x1 − x2
7 Negative Positive Negative No x2 (x1 = x3 = 0) 1 − x2
8 Negative Negative Negative No (x1 = x2 = x3 = 0) 1
9 Positive Negative Negative No x1 (x2 = x3 = 0) 1 − x1
4.1  Mathematical Algorithms 35

Outcome 2  The minimum is outside the region of interest, and x1, x2, and x3 are
positive. In this case, the problem is to find the minimum in the plane defined by
x1 + x2 + x3 = 1, that is, x4 = 0. So, x3 can be replaced by (1 − x1 − x2) and the function
to be minimized becomes:

F = U1 x12 + U 2 x22 + U3 x1 x2 + U 4 x1 + U 5 x2 + U6 (4.19)



where:

U1 = T1 + T3 − T5

U 2 = T2 + T3 − T6

U3 = 2T3 + T4 − T5 − T6

U 4 = T5 + T7 − T9 − 2T3

U 5 = T6 + T8 − T9 − 2T3

U6 = T3 + T9 + T10

The function to be minimized is similar to the case of three components presented
earlier. Then, x1 and x2 are calculated according to the procedure described and
x3 = 1 − x1 − x2 and x4 = 0.
Outcome 3  The minimum is outside the region of interest, x2 and x3 are positive,
and x1 is negative. In this case, making x1 = 0, the function to be minimized becomes:

F = U1 x22 + U 2 x32 + U3 x2 x3 + U 4 x2 + U 5 x3 + U6 (4.20)



where:

U1 = T2

U 2 = T3

U3 = T6

U 4 = T8

U 5 = T9

U6 = T10

The function to be minimized is similar to the case of three components presented
earlier. So, x2 and x3 are calculated according to the procedure described and x4 = 1
− x2 − x3 and x1 = 0.
36 4  The Linear Spectral Mixture Model

Outcome 4  The minimum is outside the region of interest, x1 and x3 are positive,
and x2 is negative. In this case, making x2 = 0, the function to be minimized becomes:

F = U1 x12 + U 2 x32 + U3 x1 x3 + U 4 x1 + U 5 x3 + U6

where:

U1 = T1

U 2 = T3

U3 = T5

U 4 = T7

U 5 = T9

U6 = T10

The function to be minimized is similar to the case of three components presented
earlier. Then, x1 and x3 are calculated according to the procedure described and
x4 = 1 − x1 − x3 and x2 = 0.
Outcome 5  The minimum is outside the region of interest, x1 and x2 are negative,
and x3 is positive. In this case, making x1 = x2 = 0, the function to be minimized
becomes:

F = T3 x32 + T9 x3 + T10 (4.22)



And, to find the minimum:

dF
= 2T3 x3 + T9 = 0
dx3

Then:

x3 = −T9 / 2T3

If x3 is in the range between 0 and 1, so this is the final solution. If x3 is greater than
1, make x3 = 1, or, if x3 is less than 0, make x3 = 0 and x4 = 1 − x3.
Outcome 6  The minimum is outside the region of interest, x1 and x2 are positive,
and x3 is negative. In this case, making x3 = 0, the function to be minimized becomes:

F = U1 x12 + U 2 x22 + U3 x1 x2 + U 4 x1 + U 5 x2 + U6 (4.23)


4.1  Mathematical Algorithms 37

where:

U1 = T1

U 2 = T2

U3 = T4

U 4 = T7

U 5 = T8

U6 = T10

The function to be minimized is similar to the case of three components presented
earlier. Then, x1 and x2 are calculated according to the procedure described and
x4 = 1 − x1 − x2 and x3 = 0.
Outcome 7  The minimum is outside the region of interest, x2 is positive, and x1
and x3 are negative. In this case, making x1 = x3 = 0, the function to be minimized
becomes:

F = T2 x22 + T8 x2 + T10 (4.24)



And, to find the minimum:

dF
= 2T2 x2 + T8 = 0
dx2

Then:

x1 = −T7 / 2T1

If x2 is in the range between 0 and 1, then this is the final solution. If x2 is greater
than 1, make x2 = 1, or, if x2 is less than 0, make x2 = 0 and x4 = 1 − x2.
Outcome 8  The minimum is outside the region of interest, and x1, x2, and x3 are
negative. In this case, making x1 = x2 = x3 = 0, then x4 = 1.

Outcome 9  The minimum is outside the region of interest, x1 is positive, and x2 and
x3 are negative. In this case, making x2  =  x3  =  0, the function to be minimized
becomes:

F = T1 x12 + T7 x1 + T10 (4.25)



38 4  The Linear Spectral Mixture Model

And, to find the minimum:

dF
= 2T1 x1 + T7 = 0
dx1

Then:

x1 = −T7 / 2T1

If x1 is in the range between 0 and 1, so this is the final solution. If x1 is greater than
1, make x1 = 1, or, if x1 is less than 0, make x1 = 0 and x4 = 1 − x1.

4.1.4  Weighted Least Squares (WLS)

Consider the curve fit the data with a curve having the form:

R = f ( A,,,, x1 ,,,, x2 ,,,,…,,,, xn ) = x1 f ( A ) + x2 f ( A ) +…+ xn f ( A ) (4.26)



where the dependent variable R is linear with respect to x1, x2, ..., xn.
Although there are many ramifications and approaches to the curves fitting, the
least squares method can be applied to a wide variety of curves fitting problems
involving linear form with undetermined values. The values are determined by
minimizing the sum of squared errors (residuals). The obtained solution by this
method is mathematically possible, but an example of what is mentioned being
physically unacceptable (some restrictions are involved: the values should not be
negative and should add to 1). Then, it becomes a problem of constrained least
squares, and the equations of constraints should be added. To solve this problem, it
is necessary to apply the concepts of weighted least squares.
Sometimes, the information obtained in an experiment can be more accurate than
those arising from other sources of information related to the same experiment. In
other cases, it is convenient to use some additional information (previous knowledge)
to make the solution physically relevant. In such cases, it may be desirable to give a
higher “weight” for the information that are considered more accurate or more
important to the problem. To weigh certain information (e.g., additional informa-
tion) is desirable to bring the solution near the physical meaning, then getting an
acceptable solution.
In this case, x1 + x2 + ... + xn = 1 and 0 ≤ x1, x2, ..., xn ≤ 1 are the conditions that
must be met to obtain an acceptable solution. So, n + 1 equations are added to the
system of Eq. 4.8: one corresponding to the conditional sum of proportions equal to
1 (x1 + x2 + ... + xn = 1) and other n corresponding to the condition that the proportions
should not be negative (xj ≤ 1, j = 1, 2, ..., n). To resolve this issue, when restrictions
4.1  Mathematical Algorithms 39

are not met, it is applied a diagonal matrix W containing values of weights associated
with the system of equations to be solved. Initially the first assigned values equal to
1, along the diagonal matrix W, mean that the equations are equally important for
the solution of the problem. The very high value assigned to the diagonal
corresponding to the first constraint (sum of xj = 1) indicates that this equation must
be strictly satisfied. So, if the values of xj’s are satisfied, that is, if they are in the
range between 0 and 1, then the final solution was found. Otherwise, an iterative
process must be used in order to bring all xj’s within the range between 0 and 1. This
is accomplished by a gradual increase of the weights, which are initially zero,
corresponding to n latest equations relating to the restriction that the proportions
should not be negative. The solution to this problem is found by minimizing the
amount: W1 e12 + W2 e22 + ... + W(m + n + 1) e(m + n + 1)2, where W1, W2, etc. are the weight
factors and e1, e2, etc. are the error values for each equation.
The implementation of this method is based on Gaussian elimination and replace-
ment algorithm (forward and backward), described in textbooks of numerical analy-
sis, such as Burden et al. (1981).

4.1.5  Principal Components

Given an image consisting of a number of pixels with measures in a number of


spectral bands, it is possible to model each spectral response of each pixel as a linear
combination of a finite number of components.

dn1 = f1e1,1 +  + fn e1, n band 1


dn2 = f1e2,1 +  + fn e2,n band 2
   
dn p = f1e p,1 +  + fn e p, n band p

where:
dnj = DN for the band i of the pixel
ei, j = pure component dn of the pure component j, band i
fj = unknown fraction of the pure component j
n = number of pure components
p = number of bands
That makes the matrix equation:

dn = ef (4.27)

40 4  The Linear Spectral Mixture Model

A linear constraint is added because the sum of the fractions of any pixel must be
equal to 1; therefore, it needs to increase the vector dn with an additional 1 and the
matrix e with a row of values 1. This becomes a set of p + 1 equations in n unknowns.
Since the number of pure components is generally smaller than the number of
spectral bands, the equations are possibly undetermined and can be solved by any
other techniques. The solution described uses principal component analysis (PCA)
to reduce the dimensionality of the dataset. The matrix of pure component is
transformed into a PCA space using the appropriate number of eigenvectors, the
pixel data are transformed into PCA space, the solutions are found, and the resulting
fractions are stored.
The weighted least squares method and the principal component method are rec-
ommended for the cases where the number of spectral components in the mixture is
greater than three.

4.2  Endmembers Selection

To generate the fraction images, it is necessary to choose the pure components


(endmembers) to apply any available mathematical algorithm. Explicitly defined,
the endmember is just a component that is part of the spectral mixture. So, it is
necessary to choose the endmembers that make sense for the interpretation of the
image considered and also that meet the criteria of the fraction according to the
equations of the mixture already presented earlier. Sometimes, it is easy to choose
these endmembers because targets are already known for the area being studied.
This is true for areas that have been studied for some time, for example, when the
changes in the environment are being monitored. On the other hand, it is neces-
sary to perform experiments to find the appropriate endmembers whenever the
scenes are unknown or when it is needed to extract specified information from the
images.
There are two ways to select the endmembers: directly from the images and
through collections of data obtained in the laboratory and/or in the field. The end-
members derived from the images are called image endmembers, while those
selected from laboratory data and/or in the field are called reference endmembers.
The most convenient is to select the endmembers directly from the images being
studied for the simple reason that the spectrum of the endmember extracted from
the image can be used without calibration. In this case, spectra obtained in the labo-
ratory or in the field present values of reflectance factors. As seen earlier, the orbital
images are, in principle, available in the DN format. In that way, the reference
endmembers (reflectance factors) would be in scale or measure unit different from
that adopted by the images (DNs), which would interfere in the application of the
mixture model. The right procedure, then, would be to convert the DNs of images
in surface reflectance factors, with atmospheric correction, in order to have com-
patibility on the unity of the two data sets.
Image endmembers, although convenient for not needing the calibration, do not
always work in mixture models. To work well, it is necessary to have a good relationship
4.2  Endmembers Selection 41

Fig. 4.4  Example of dispersion of pixels of an image in the graph formed by the red and near-­
infrared bands, showing the potential endmembers of vegetation (green), soil (red), and shadow/
water (blue)

between the pixel scale and the scale in which the materials occur relatively pure on the
ground. In the best case, it can be used image endmember if an image present at least
some pixels fully occupied by a pure material on the ground.
It can be concluded in a way more or less intuitive that the definition of the num-
ber and the selection of endmembers to be considered are essential to the success of
the application of mixture model. Nevertheless, it is known that, in the real world,
the terrain can be spectrally complex. One reason why the images of scenes on the
ground can be well modeled by few endmembers spectra is due to the fact that some
of the potential endmembers is in small proportions in comparison with the end-
members considered in the mixture. In this way, the determination of the number of
endmembers is made by defining the vertices of a geometric figure that encompass
the spectral responses of the pixels of the scene. For example, in the case of three
endmembers, the geometric figure is a triangle on the two-dimensional plane formed
by two spectral bands (in general, for the land cover, red and near-­infrared spectral
bands of the electromagnetic spectrum) (Fig. 4.4). The endmembers will be those in
which the spectral responses are closest to the vertices of the geometric figure
formed.
Now, in the event that it is decided to use image endmembers, there will be frac-
tion images with the proportion of 100% (pure pixel), while, if it is decided to use
reference endmembers, it will be hardly to have any pure pixel in the fraction
images.
Chapter 5
Fraction Images

Abstract  This chapter is dedicated to the presentation of some linear spectral mix-
ture model applications that is based in the so-called fraction images. Details are
also presented on the interpretation of those images and their meanings.

Keywords  Fraction images · Vegetation fraction · Soil fraction · Shade/water


fraction

The fraction images are products generated from the application of the mathemati-
cal algorithms described previously. They represent the proportions of the compo-
nents in the spectral mixture. In general, all algorithms produce the same result, i.e.,
generate the same fraction images when the constraint equations are not used or the
proportions are in the range from 0 to 1. Normally the vegetation, soil, and shade/
water fraction images are generated, which are the components usually present in
any scene on the ground. The fraction images can be considered a form of dimen-
sionality reduction of the data and also a form of enhancement of the information.
In addition, the spectral mixture model transforms the spectral information into
physical information (proportion values of the components in the pixel; not to be
confused with another type of transformation that converts spectral or radiometric
data into physical unities as radiance or reflectance).
The vegetation fraction image highlights the areas of vegetation cover, the soil
fraction image highlights the areas of exposed soil, and the shade/water fraction
image highlights the occupied areas with water bodies such as rivers and lakes and
also the burned areas, the wetlands, etc. The shade and the water are considered
together because these two targets exhibit similar responses in the spectral bands
typically used by Earth observation sensors. In this way, it is important to keep on
mind that the spectral mixture model is not a classifier, but rather an image transfor-
mation technique to facilitate the extraction of information.
For the generation of fraction images, the spectral responses of pure components
(endmembers) are considered known, i.e., can be obtained directly from the images
(image endmember) or from the available spectral libraries (reference endmember).
Figure 5.1 shows an example of the spectral responses of vegetation, soil, and s­ hade/

© Springer Nature Switzerland AG 2019 43


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0_5
44 5  Fraction Images

Fig. 5.1  Spectral response of vegetation (green color), soil (red color), and shade/water (blue
color) components

water components used to generate fraction images in a scene of the OLI/Landsat 8


image Path 231/row 062, corresponding to the Manaus (AM) region. In this case it
was used the images of all bands, 1 (0.43 μm–0.45 μm), 2 (0.45 μm–0.51 μm), 3
(0.59 μm–0.53 μm), 4 (0.64–0.67 μm), 5 (0.85 μm–0.88 μm), 6 (1.65 μm–1.57 μm),
and 7 (2.29 μm–2.11 μm), that were previously converted to apparent reflectance
values. It is worth to remember that this analysis could be performed using these
values of apparent or surface reflectance or even DNs.
Obviously, the users must take into account the data type which they are work-
ing, particularly when selecting the endmembers. Users with more familiarity with
the analysis of reflectance curves of different natural resources probably will feel
more comfortable to select endmembers when working with images converted into
surface reflectance values, since the shape of the curves will inform about the nature
of the pixels selected as pure. It does not mean that they are not able to make good
selection working with images composed by DNs. In this case the shape of the
curves would not be very useful, but it would not interfere in the performance of the
mixture model.
After applying the mixture model, new images are generated, composed then by
numbers that represent the proportion values of a given component within each
pixel. As an example, Fig.  5.2a shows a color composite (R6 G5 B4) based on
images from the OLI/Landsat 8 sensor, while Fig. 5.2b–g presents the corresponding
images of the bands 2–7.
Figure 5.3a shows a color composite of the fraction images Soil Fig. 5.3b,
Vegetation Fig. 5.3c and Shade/water Fig. 5.3d derived from the sensor OLI/Landsat
8 for the Manaus region (AM) soil, vegetation, and shade/water fraction
5  Fraction Images 45

Fig. 5.2 (a) Color composite (R6 G5 B4) of OLI/Landsat 8 for 231/062 image; (b) band 2; (c)
band 3; (d) band 4; (e) band 5; (f) band 6; (g) band 7

Fig. 5.3 (a) Color composite of the fraction images for the region of Manaus (AM) and the frac-
tion images (b) soil, (c) vegetation, and (d) shade/water

In the vegetation fraction image (Fig. 5.3c), the brighter pixels are those which,
at least in theory, have greater amount of vegetation, while the water bodies appear
dark since they do not have any percentage of vegetation cover. A similar analysis
can be done with the other components of fraction images. For example, in the soil
fraction image (Fig. 5.3b), the brighter pixels are those that have the lowest levels of
vegetation cover or are less shaded.
Figure 5.4 shows a color composite (R6 G2 B1) of the bands 1 (red), 2 (near
infrared), and 6 (medium infrared) of MODIS/Terra sensor for the west region of
the state of São Paulo in Brazil. The corresponding vegetation, soil, and shade/water
fraction images are presented in Fig. 5.4b–d.
Figure 5.5 displays a color composite (R6 G5 B4) of the bands 4 (red), 5 (near
infrared), and 6 (medium infrared) of the OLI Landsat 8 sensor for the part of the
image Path 226/row 068 in the state of Mato Grosso, while Fig. 5.5b–d presents the
corresponding vegetation, soil, and shade/water fraction images.
It is observed that the fraction images are monochromatic (grayscale) and the
DNs are directly associated with the proportions (abundance) of each of the
46 5  Fraction Images

Fig. 5.4 (a) Color composite (R6 G2 B1) of MODIS/Terra for the west region of the state of São
Paulo and the fraction images (b) soil, (c) shade/water, and (d) vegetation

r­ espective components of the scene selected for the spectral mixture model. Thus,
the higher the DN value in the vegetation fraction image (Fig. 5.5b), the greater the
proportion of vegetation in the corresponding pixel (bright green in Fig. 5.5). The
same interpretation applies to the other components fraction images: the higher the
DN value in the soil fraction image (Fig. 5.5c), the greater the proportion of soil in
the corresponding pixel (magenta in Fig. 5.5), and the higher the DN value in the
shade/water fraction image (Fig. 5.5d), the greater the proportion of water or burned
in the corresponding pixel (dark magenta or black in Fig. 5.5a).
The literature presents a large amount of works on the use of the linear spectral
mixture model in several regions around the world, showing that this technique is
consistent. In addition, the fraction images generated by this model are being used in
different application areas, such as forest, agriculture, land use, water, and urban areas.
So, the proportion of each endmember can be shown for each pixel, creating an
image useful for photointerpretation. Figures 5.3, 5.4, and 5.5 show the proportions
of each endmember represented in gray levels. The fraction images are derived on
the basis of the information from all multispectral bands used. For each type of
application, a limited number of spectral bands can be employed, for example, the
deforested area analysis in the Amazon region can be based on only three bands:
red, near infrared, and short wave infrared.
Therefore, the conversion of the spectral data in fraction images by means of
linear spectral mixture model can result in a significant reduction in dimensionality
of the data to be analyzed. For example, it is possible to use several bands of ­different
sensors – 6 bands of TM, 7 bands of MODIS, and 242 bands of Hyperion – to gener-
ate a small number of fraction images (usually, three or four endmembers).
5  Fraction Images 47

Fig. 5.5 (a) Color composite (R6 G5 B4) of OLI/Landsat 8 of the part of the image Path 226/row
068 in the state of Mato Grosso and fraction images (b) vegetation, (c) soil, and (d) shade/water

Now, once known the proportions of the fraction images and the spectral
responses of the endmembers, it is possible to retrieve the spectral responses of the
pixels in each of the spectral bands used. This procedure allows evaluating the per-
formance of the models by generating the error images, which will be presented in
the next section.
The grayscale of fraction images are brighter with the higher proportion of end-
member in the pixel. The same convention is adopted for the error image associated
with the model. The choice of the convention is arbitrary, so that sometimes, the
tones of the image can be reversed for helping the visualization of the patterns, for
example, in the case of the shade/water fraction image, it is usually more intuitive
when the tones are reversed, so that dark tones indicate less lighting. The fraction
images can also be presented in color composite (RGB) by selecting three images
corresponding to the endmembers. In this case, the image contrast stretching can be
applied for the visualization of the patterns but distorts the proportion values for the
quantitative information.
48 5  Fraction Images

To analyze a fraction image individually, it is more effective to present the higher


proportions of the endmembers in brighter grayscales, to highlight the target of interest.
In case of considering three endmembers such as vegetation, soil, and shade/
water, the vegetation image enhances the vegetation cover so proportional to the
vegetation greenness, the soil fraction image highlights the areas without vegetation
cover, and the shade/water fraction image highlights the areas occupied by water
bodies and the burned areas.

5.1  Error Images

As mentioned earlier, based on the spectral mixture model, it is possible to calculate


the error for each of the spectral bands and to generate the corresponding error
images, since the responses of the components and their proportions are known.
This is one way to evaluate the performance of the model, that is, when the model
is appropriate, the error images show a nonstandard aspect. If there is a component
that has not been considered in the mixture, it will be highlighted in the error images
of the spectral bands employed.
It is possible to write the mixture model as described earlier:

r1 = a11 x1 + a12 x2 +  + a1n xn +e1


r2 = a21 x1 + a22 x2 +  + a2 n xn +e2
    
rm = am1 x1 + am 2 x2 +  + amn xn + em

or,

ri = ∑ ( aij x j ) + ei (5.1)

So the errors for each band can be obtained by:

e1 = r1 − (a11 x1 + a12 x2 +  + a1n xn )


e2 = r2 − (a21 x1 + a22 x2 +  + a2 n xn )
    
em = rm − (am1 x1 + am 2 x2 +  + amn xn )

or,

ei = ri − ∑ ( aij x j ) (5.2)

where:
ri = average spectral reflectance for the i-th spectral band
aij = spectral reflectance of the j-th component in the pixel for the i-th spectral band
5  Fraction Images 49

xj = proportion value of the j-th component in the pixel


ei = error for the i-th spectral band
j = 1, 2,..., n (n = number of components assumed for the problem)
i = 1, 2,..., m (m = number of spectral bands for the sensor system)
The error images usually present low values according to the accuracy of the
employed models. In this way, these images are used to evaluate the quality of the
defined models, that is, if the number of components is suitable for the analyzed
scene. If there is any component not represented in the mixture, it will be high-
lighted in these error images.
Figures 5.6 and 5.7 present an example of the evaluation of the spectral mixture
model using an OLI/Landsat 8 image obtained over an area of the state of Mato
Grosso (path 226/row 068).

Fig. 5.6 (a) Color composite of the fraction images for the image OLI/Landsat 8 (path 228/row
068), in the state of Mato Grosso, and the fraction images: (b) vegetation, (c) soil, and (d) shade/
water
50 5  Fraction Images

Fig. 5.7  Error images: (a) band 2; (b) band 3; (c) band 4; (d) band 5; (e) band 6; (f) band 7

In this example, considering all bands, the average error was 6.109, while the
error per band was as follows: 9.420 in band 2, 11.601 in band 3, 7.152 in band 4,
1.404 in band 5, 4.338 in band 6, and 2.739 in band 7.
Chapter 6
Fraction Images Applications

Abstract  Examples of fraction image application are given for mapping projects of
large areas of the Earth’s surface such as the operational PRODES project by INPE,
which aims to calculate rates of deforestation in the Brazilian Amazon.

Keywords  PRODES · Vegetation mapping · DETER · Deforestation estimation

This chapter aims to present some examples of the use of fraction images derived
from the linear spectral mixture model in the environmental monitoring projects.
The fraction images, due to their characteristics described above (data reduction and
enhancement), have contributed to the development of large-scale projects that use
a large amount of images.

6.1  Monitoring of Deforestation

The deforestation in the Legal Amazon (AML) has been a concern of various gov-
ernmental and nongovernmental organizations, especially during the last three
decades (Moran 1981; Skole and Tucker 1993). Although there is not a long history
of human occupation in AML, almost 90% of deforestation for pasture and agricul-
ture occurred between 1970 and 1988, as indicated by the estimates based on satel-
lite images (Skole et al. 1994).
Historically, the Brazilian territory was occupied along the coastline, with the
majority of its population concentrated in this region. In an attempt to change this
pattern of occupancy by increasing settlement in the interior of the country, the
federal capitol was transferred from the coast, in Rio de Janeiro, to the central
region, in Brasilia, in mid-1950 (Mahar 1988). This policy of occupation required
infrastructure investments to connect Brasília with the other regions of the country.
The construction of the Belém-Brasília highway (BR-010), in 1958, was the domi-
nant factor that triggered the main activities of deforestation in AML (Moran et al.
1994; Nepstad et  al. 1997). Subsequent events, such as the construction of the

© Springer Nature Switzerland AG 2019 51


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0_6
52 6  Fraction Images Applications

BR-364, crossing the states of Mato Grosso, Rondônia, and Acre, and the PA-150,
in the state of Pará, encouraged further deforestation activities, converting forests
into pasture and agriculture areas (Moran 1993).
To introduce the governance in AML, the “Superintendência do Desenvolvimento
da Amazônia” (SUDAM) and the “Banco da Amazônia” (BASA) were created in
1966. Small producers have been financed to encourage investment in projects of
agriculture (Moran et  al. 1994). Large producers were also financed through tax
incentives in trade for converting forest into pasture areas (Moran 1993). The incen-
tives granted to the large producers were the main causes of deforestation; the small
producers have had a smaller impact on deforestation due to the practice of com-
paratively smaller dimension of the subsistence agriculture (Fearnside 1993).
Other activities with high economic value, such as mining and selective exploita-
tion of timber, also contributed to the deforestation in AML (Cochrane et al. 1999).
The areas of deforestation in Brazilian Legal Amazon have been concentrated in the
so-called arc of deforestation, located in the southern and eastern parts of AML,
from Acre to the Maranhão states (Cochrane et al. 1999; Achard et al. 2002).

6.1.1  The Brazilian Legal Amazon Monitoring Program

Since 1973 Brazil has access to the remote sensing images from Landsat satellite
series that allow quantifying the extent of natural resources and the changing of the
Amazon region. Based on availability of these images, the Brazilian government
started monitoring the Amazon rainforest for quantifying deforestation areas at sev-
eral years’ intervals.
The Brazilian government has conducted annual monitoring of the Amazon for-
est since 1988 using images generated by the Landsat program through the PRODES
(Monitoring of Brazilian Amazon Forest by Satellite) project conducted by the
National Institute for Space Research (INPE). It is the largest remote sensing proj-
ect in the world for monitoring deforestation activity in tropical forests aiming to
assess all deforested areas within the five million square kilometers of the AML,
which is covered by approximately 229 TM/Landsat 5 scenes (Fig. 6.1).
This project estimated approximately 750,000 km2 of deforestation in the AML
until 2010, a total which accounts for approximately 17% of the original forest
extension. These data showed annual rates of deforestation which varied signifi-
cantly in response to political, economic, and financial conditions of the country as
well as foreign market demands.
The PRODES information are based mainly on images of sensors with medium
spatial resolution (30 m), such as those generated by the Landsat program, and with
relatively low temporal resolution (16 days frequency of acquisition), allowing the
annual monitoring of the deforestation. A faster update of forest change is not pos-
sible with these images due to the low frequency of acquisition of cloud-free images,
which is a serious problem for the Amazon region, limiting the number of viable
observations of the ground surface. This fact prevents the Brazilian government and
6.1  Monitoring of Deforestation 53

Fig. 6.1  Brazilian Legal Amazon covered by 229 TM or ETM + Landsat images for the estimation
of annual deforestation. (Source: INPE (2002))

environment control agencies to make quick and appropriate interventions to stop


the activities of illegal deforestation.
Then to monitor the deforestation in near real time is possible using the almost
daily images acquired by MODIS sensor on Terra and Aqua platforms. Thus,
through the project DETER (near real-time deforestation detection), a new method-
ology based on MODIS images was developed for the rapid detection of deforesta-
tion in the Amazon (Shimabukuro et  al. 2006). While the MODIS is a moderate
spatial resolution sensor and does not generate viable images to estimate the exten-
sion of deforested areas, the MODIS data can be valuable as indicators of changes
or as the alarm product for the service of management and control polices of the
ground surface.
The following sections show an overview of the digital PRODES and DETER
projects for monitoring deforestation activities in the Brazilian Legal Amazon
yearly and monthly, respectively. First, the deforestation history and the description
of methodology developed at INPE for monitoring based on geographic informa-
tion system (GIS) and remote sensing image processing are presented. Then, it is
shown the important contribution of the fraction images derived from the linear
spectral mixture model to make the digital PRODES and DETER viable, which
allowed the releasing of the deforestation data in transparent mode to the national
and international community. The results provide invaluable contribution to the
decision-makers in establishing public policies and strengthen environmental gov-
ernance in critical ecosystems of the Brazilian Legal Amazon.
54 6  Fraction Images Applications

6.1.2  The Digital PRODES Project

Since the late 1970, INPE has performed evaluations of deforestation in AML using
remote sensing images. These assessments were made in conjunction with the for-
mer Brazilian Institute for Forest Development (IBDF), which was later incorpo-
rated into the Brazilian Institute of Environment and Renewable Natural Resources
(IBAMA). The first assessment was carried out with the use of images acquired by
the MSS sensor, with four spectral bands and spatial resolution of ~80 m, onboard
of Landsat 1, 2, and 3 satellites, during 1973–1975 and 1975–1978 periods, employ-
ing visual interpretation techniques (Tardin et al. 1980).
From 1988 onward, the annual deforestation assessments were provided to the
entire AML using the TM sensor, six spectral bands with spatial resolution of 30 m,
on board the Landsat 5 satellite, with improved quality of mapping due to its better
spatial and spectral resolutions when compared with the MSS data. The methodol-
ogy applied to map deforested areas was based on visual interpretation of color
composites (R5 G4 B3) of TM images in hard-copy format on a scale of 1:250,000.
The polygons interpreted visually of the deforested areas were added together to
calculate the total deforested areas for each state of the AML and presented in table
format. This method, known as analogic PRODES, was performed until 2003.
At the end of 1990 decade, a semi-automatized methodology using the fraction
images began to be developed and was named the digital PRODES (Shimabukuro
et  al. 1998). The digital PRODES project is an automation of the activities per-
formed in the project PRODES, which was based on analogical data since the 1970
decade.
According to PRODES the deforestation in the Brazilian AML is shown by a
mask of intact forest that is updated annually through the identification of new
deforestation event excluding the non-forest vegetation areas and identifying other
dynamic changes, as clear-cut of secondary regeneration areas. Landsat TM images
or images from similar sensors are selected for July, August, and September period,
which is within the local dry season period on the arc of deforestation and repre-
sents an atmospheric window when cloud-free images are normally available. These
images are geometric corrected using the technique of sampling the nearest neigh-
bor to the UTM projection resulting in a cartographic product with internal error of
50 m. The spectral bands TM 3 (red), TM 4 (NIR), and TM 5 (MIR) are used to
generate the fraction images. The legend for the maps contains the following classes:
forest, non-forest (shrub savanna (cerrado), grassy savanna (campo limpo de cer-
rado), campinarana etc.), deforestation accumulated of previous years, deforesta-
tion of the analyzed year, hydrography, and cloud.
The digital PRODES consists of the following methodological steps: (1) genera-
tion of vegetation, soil, and shade/water fraction images; (2) image segmentation
based on the growing region algorithm; (3) image classification based on unsuper-
vised classifier; (4) mapping classes based on the following legend – forest, non-­
forest (vegetation that is not characterized by a forest structure), deforestation
(clearing accumulated up to the previous year), hydrography, and clouds; and (5)
6.1  Monitoring of Deforestation 55

edition of the classified map based on visual interpretation to minimize the omission
and commission errors of automatic classification in order to produce the final map
of deforestation in digital format. PRODES products are available on the official
INPE website.
A linear spectral mixture model is used to produce vegetation, soil, and shade/
water fraction images applied to the spectral bands of Landsat TM (Shimabukuro
and Smith 1991) (Fig.  6.2). This method reduces the dimensionality of data and
highlights the specific targets of interest. The vegetation fraction image highlights
the areas of vegetation cover, the soil fraction image highlights the uncovered soil,
and the shade/water fraction image highlights the areas of water bodies and the
burned areas. The shade/water fraction image was used to characterize the total area
deforested before 1997  in AML according to the methodology proposed by
Shimabukuro et al. (1998). Subsequently, the deforested areas were accumulated by

Fig. 6.2 (a) Landsat TM image (R5 G4 B3) and fraction images (b) vegetation, (c) soil, and (d)
shade/water
56 6  Fraction Images Applications

the year 2000. From there, the soil fraction image, which highlights the areas with-
out vegetation cover, was used to classify the annual increment of deforestation
based on the high contrast between forested and deforested areas in the following
years. The digital PRODES, recognized nationally and internationally, allowed
INPE to put at the disposal of the general community information of deforested
areas in the AML.
The linear spectral mixture model used was:

ri = a vegei + b soili + c ( water or shade )i + ei (6.1)



where:
ri = spectral response for the pixel in the band i of Landsat TM image
a, b, and c = proportion of vegetation, soil, and shade/water in each pixel
vegei, soili, and (water or shade)i = spectral responses of each component
ei = error term for each band of Landsat TM image
TM 3, TM 4, and TM 5 bands are used to form a system of linear equations that
can be solved by any algorithm developed, for example, weighted least squares
(WLS), described earlier. The resulting fraction images were resampled to 60 m to
minimize computer processing time and disk space, with no loss of information
compatible with the map scale of the final product of 1:250,000.
The next step is the application of a digital image processing technique called
image segmentation, which is based on grouping data into contiguous regions with
similar spectral characteristics. Two thresholds are required to perform the image
segmentation: (a) similarity, which is the minimum value defined by the user to be
considered as similar and to form a region, and (b) area, which is the minimum size,
in number of pixels, for the region to be individualized (Fig. 6.3). The unsupervised
classification method (ISOSEG) is used to classify the segmented fraction images
segmented and employs the statistical attributes (mean and covariance matrix)
derived from polygons generated by image segmentation (Fig. 6.4).
After the unsupervised classification, it is necessary to check the resulting maps,
according to the previously established legend of the PRODES project (Fig. 6.5).
Next, the task of editing the map is executed by interpreters using interactive image-­
editing tools (Fig. 6.6). Omission and commission errors identified by the interpret-
ers are corrected manually in order to improve the classification result.
Then the images individually classified are mosaicked to generate the final maps
for each state and for the entire AML (Fig. 6.7). To the mosaic of the states, the
spatial resolution is maintained at 60 m, and the scale for presentation is 1:500,000,
while for the AML the spatial resolution is resampled to 120 m, and the scale for
presentation is 1:2,500,000, due to the large amount of information.
However, the deforestation information provided by the PRODES was not suffi-
cient for the needs of most frequent surveillance of various agencies of the Brazilian
government. Therefore, the DETER project was developed, based on images from
the MODIS sensor, with high temporal resolution, to provide geospatial information
from deforestation activities in near real time.
6.1  Monitoring of Deforestation 57

Fig. 6.3  Segmented shade/water fraction image

Fig. 6.4  Segmented and classified image using the unsupervised classifier
58 6  Fraction Images Applications

Fig. 6.5  Thematic classes according to the previously established legend of the PRODES project

6.1.3  DETER Project

From 2004, the DETER project was implemented in order to provide real-time
detection of deforestation activities to support the federal government’s action plan
for the prevention and control of deforestation in the Brazilian AML. The procedure
mimics the methodology adopted in the PRODES project but is intended to detect
deforestation activities in real time, exploring the high temporal resolution of
MODIS sensor.
The first step in the method of DETER project is “to mask” the intact forest
based on the evaluation of PRODES in the previous year. The map of intact forest is
used as a reference for identifying new real-time deforestation events throughout
the analyzed year. The monitoring activity with MODIS images starts in January but
6.1  Monitoring of Deforestation 59

Fig. 6.6  Image edition to generate the final map

Fig. 6.7  Digital PRODES thematic map showing the deforested areas up to 2000 (in yellow), non-­
forest areas (in magenta), and annual deforestation of 2001–2010 according to the legend

becomes more effective after March, when a larger number of MODIS images are
available due to less cloud cover in AML. In addition, during the rainy season, from
November to March, there is not much deforestation expected to happen.
Daily MODIS images (surface reflectance-MOD09) used to identify sources of
deforestation are selected based on two criteria: (a) amount of cloud coverage and
60 6  Fraction Images Applications

(b) track inside the Zenithal angle of the sensor smaller than 35° (~1400 km). The
amount of cloud cover is initially evaluated in accordance with the quick-look
images followed by more detailed analysis with actual spatial resolution of MODIS
images. The AML is covered by 12 MODIS tiles (V09 to V11 and H10 to H13).
The images of MOD09 product are in sinusoidal projection (WGS84 datum),
and the bands are projected for the geographic coordinate system with the same
datum and converted from HDF (hierarchical data format) to GeoTIFF in order to
download the images directly to SPRING software for image processing proposes.
High geometric quality of MODIS products ensures the viability of the project,
because it is crucial to detect points of deforestation within the MODIS pixel size.
From the set of seven bands of MOD09 product, band 1 (red), band 2 (NIR), and
band 6 (MIR) are used to generate the vegetation, soil, and shade/water fraction
images applying linear spectral mixture model (see Sect. 6.1.2), as can be seen in
Fig. 6.8 for the period from 22 April to 7 May 2004. In this figure, the soil fraction
image facilitates the mapping of deforested areas.
The soil fraction images are then segmented, classified, mapped, and eventually
edited by interpreter following the same methods used in the digital PRODES proj-
ect. Figure 6.9 shows the steps of the DETER project, i.e., the deforested areas clas-
sified in the soil fraction images are superimposed on the forest mask, highlighting
the areas of new deforestation (red color).
This procedure is performed for each daily MODIS image acquired for the
Brazilian AML. The results of the deforestation activities detected by DETER can
be accumulated for different intervals such as weekly, biweekly, and monthly and
are available in digital format on the project website of INPE. Figure 6.10 displays
the products available for the year 2004 for this project.

Fig. 6.8 (a) MODIS image corresponding to the period of 22 April to 7 May 2004 and fraction
images (b) soil, (c) shade/water, and (d) vegetation
Fig. 6.9 (a) Soil fraction image (mosaic of 22 April to 7 May 2004); (b) result of the classification
of the soil fraction image; (c) classification of the MODIS image (mosaic of 22 April to 7 May
2004) – total extension: August 2003 + change until 7 May 2004

Fig. 6.10  Result of monitoring of DETER project showing deforestation activities monthly
detected during the year 2004
62 6  Fraction Images Applications

6.2  Burned Areas Mapping

Beyond the capacity of the MODIS images of serving for the real-time detection of
the deforested areas (DETER project), through them it is possible also to identify
other types of anthropogenic action on the forest cover, as is the case of burning
activities. This can be seen with the results obtained in the state of Acre using
MOD09 product acquired on 5, 12, and 21 September 2005, composed by the spec-
tral bands of red (centered at 640 nm), near infrared (858 nm), and middle infrared
(1640 nm) (Shimabukuro et al. 2009).
For the initial interpretative phase of burned areas using MODIS images, it was
used the information from the PROARCO project, which presents the daily
­monitoring of hot pixels. The burned areas mapping project makes possible to esti-
mate the spatial distribution, the degree of incidence and the spatial direction of the
fire. That information can be compared to thematic information allowing to deter-
mine the vegetation type that is being affected. For validation of the burned areas
map using MODIS data, it was used the images of better spatial resolution, case of
Landsat TM and the CBERS-2/CCD products, acquired on 13 and 12 October 2005,
respectively, in addition to field information in the considered period. In Fig. 6.11 it
is presented the flowchart of procedures performed on the work.
Figure 6.12 shows the color composite (R6 G2 B1) and also the individual veg-
etation, soil, and shade/water fraction images for the MODIS image acquired on 12

Fig. 6.11  Flowchart of the methodology of burned areas mapping


6.2 Burned Areas Mapping 63

Fig. 6.12  MODIS image of 12 October 2005 of the state of Acre: (a) color composite (R6 G2 B1)
and fraction images (b) vegetation, (c) soil, and (d) shade/water
64 6  Fraction Images Applications

Fig. 6.13  Burned areas in the state of Acre identified in the MODIS images acquired in the year
2005

October 2005. The burned areas can be identified as gray levels higher than the
other targets in the shade/water fraction image, facilitating their discrimination.
The results of the analysis of multitemporal shade/water fraction images derived
from MODIS data indicated the occurrence of 6500 km2 of the area burned in the
state of Acre (Fig. 6.13). Of this total, 3700 km2 correspond to the previously defor-
ested areas, where burning activity serves as a traditional practice of cleaning the
ground for the implementation of agricultural crops or new pastures or even as a
practice of improvement of grazing. The other 2800 km2 correspond to the areas of
forest fires, with the forest cover degraded by fire, in both ground level and canopy
level, in burning out of control, whose degree of incidence is determined by the
wind action, availability of dry material in the interior of the forest, and occurrence
of certain species more susceptible to fire.
Concerned with the level of emissions resulting from burning and their impacts
on air quality breathed by the population of the state of Acre, Brown et al. (2006)
estimated using TM/Landsat 5 and CCD/CBERS-2 images and a detailed field-
work; more than 2670 km2 of primary forests were affected by fires in the southeast
of that state in 2005. Such results show the consistency of assessments with MODIS
images, qualifying them as an important source of information for mapping burned
areas on a regional scale.
6.4 Land Use and Land Cover Mapping 65

6.3  Selective Logging Detection

The selective logging of species of high commercial value is a practice used in areas
of the Amazon rainforest requiring several studies for their detection and measure-
ment and the evaluation of the impacts on intact forests (Asner et al. 2005; Grogan
et al. 2008; Matricardi et al. 2010; Shimabukuro et al. 2014).
Selective logging is characterized by the opening of stocking areas and explora-
tion tracks. In this way, with the use of medium spatial resolution images (e.g., TM/
Landsat 5), it is possible to detect these areas using the soil fraction image generated
by linear spectral mixture model. Figure  6.14a shows the soil fraction image
­highlighting the deforested and selective logging areas, which can be classified as
shown in Fig. 6.14b.

6.4  Land Use and Land Cover Mapping

The vegetation, soil, and shade/water fraction images have been used for mapping
the land use and the land cover. The following example is the mapping of the Mato
Grosso state using multitemporal data of MODIS/Terra. Figure 6.15a, b presents the
color composite of MODIS images acquired in the months of January and August
2002, respectively, showing the landscape change during the rainy and drought peri-
ods in the Mato Grosso state. In the image from August, it is observed the contrast
between areas occupied by bare soils and covered by very sparse vegetation (cer-
rado areas and deforested areas) and  areas occupied by dense vegetation (forest-­
covered areas). On the other hand, in the January image, it is observed the agriculture
areas, mainly soybeans, and flooded areas. Then, the fraction images of the MODIS

Fig. 6.14 (a) Soil fraction image derived from a TM image on an area in the state of Mato Grosso,
highlighting the deforested areas (clear-cut) and the areas of selective logging; (b) classification of
selective logging areas (light blue) and deforested areas (brown)
66 6  Fraction Images Applications

Fig. 6.15  Color composites of MODIS images of the state of Mato Grosso acquired in (a) January
2002 and (b) August 2002

Fig. 6.16  MODIS fraction images of August 2002 of Mato Grosso state: (a) color composite, (b)
vegetation, (c) soil, and (d) shade/water

data acquired during the year can be used to map the land use and land cover areas.
Therefore, the fraction images are very useful to reduce the data volume to be ana-
lyzed, in addition to enhance the land cover classes of interest.
Figure 6.16 displays the fraction images derived from the image acquired in the
month of August 2002, facilitating the discrimination between vegetation and
non-­vegetation areas. It is also possible to differentiate the deforested areas and cer-
6.4  Land Use and Land Cover Mapping 67

Fig. 6.17  Map of land use and land cover of Mato Grosso state obtained based on the MODIS
images. (Source: Anderson (2004))

rado areas that are without vegetation cover during that period of the year, as well
as the water bodies.
Similarly, the fraction images of the month of January 2002, not presented here,
highlight the areas of agriculture, mainly soybeans, and different types of savanna
(cerrado) and wetlands. Anderson (2004), combining the fraction images of MODIS
images acquired during the periods of the year, mapped land use classes and the
land cover of the state of Mato Grosso in 2002 as shown in Fig. 6.17.
Chapter 7
Final Considerations

Abstract  This chapter presents some final thoughts about the book.

Keywords  LSMM objectives · Future applications

The spectral mixture can be linear and nonlinear. The linear model was discussed
because of the facility of implementation with very satisfactory results.
The linear spectral mixture model is a technique of data transformation of remote
sensing data, i.e., converts the spectral information into physical proportion infor-
mation of the components (endmembers) within the pixel. This information of pro-
portion of the components is represented in new images called fraction images. In
that way, the linear spectral mixture model is a data reduction technique, and in
addition it enhances the information of these components within the image pixel. It
is not a thematic classifier, but provides useful information of fraction images for a
variety of applications in several areas.
In general, these endmembers are vegetation, soil, and shade/water elements
present on the ground. The vegetation fraction image presents similar information
of vegetation indices such as NDVI, SAVI, and EVI, highlighting the vegetation
cover areas, while the soil fraction image highlights the areas without vegetation
cover, and the shade/water fraction image highlights the water bodies and the burned
areas.
The soil and shade/water fraction images were important for automating the
PRODES project, which was done through the digital PRODES project, providing
the estimate of deforested areas and the map of spatial distribution of these areas.
Hopefully, at the end of this book, we  have contributed to the provision of
useful information to the deepest reflections for those who intend to use the fraction
images derived from linear spectral mixture model in the development of their
works.

© Springer Nature Switzerland AG 2019 69


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0_7
References

Achard, F., Eva, H. D., Stibig, H. J., Mayaux, P., Gallego, J., Richards, T., & Malingreau, J. P.
(2002). Determination of deforestation rates of the world’s humid tropical forests. Science,
297, 999–1002.
Adams, J.  B., & Adams, J.  D. (1984). Geologic mapping using Landsat MSS and TM images:
Removing vegetation by modeling spectral mixtures – remote sensing for exploration geology.
In Proceedings of the third thematic conference: International symposium on remote sensing
of environment, Colorado (pp. 16–19). Colorado Springs: ERIM.
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.
Alcântara, E. H., Barbosa, C. C. F., Stech, J. L., Novo, E. M. M., & Shimabukuro, Y. E. (2009).
Improving the spectral unmixing algorithm to map water turbidity distributions. Environmental
Modelling & Software, 24, 1051–1061.
Anderson, L.  O. (2004). Classificação e monitoramento da cobertura vegetal de Mato Grosso
utilizando dados multitemporais do sensor MODIS. 247 f. Dissertação (Mestrado em
Sensoriamento Remoto) – Instituto Nacional de Pesquisas Espaciais, São José dos Campos.
INPE-12290-TDI/986.
Asner, G. P., Knapp, D. E., Broadbent, E. N., Oliveira, P. J. C., Keller, M., & Silva, J. N. (2005).
Selective logging in the Brazilian Amazon. Science, 310, 480–482.
Atkinson, P., Cutler, M., & Lewis, H. (1997). Mapping subpixel proportional land cover with
AVHRR imagery. International Journal of Remote Sensing, 18, 917–935.
Bastin, L. (1997). Comparison of fuzzy c-means classification, linear mixture modeling and MLC
probabilities as tools for unmixing coarse pixels. International Journal of Remote Sensing, 18,
3629–3648.
Boardman, J. W.(1989). Inversion of imaging spectrometry data using singular value decomposi-
tion. In: Proceedings of the 12th Canadian symposium on remote sensing. (Vol. 4, pp. 2069–
2072). Vancouver, Canadá: IEEE IGARSS.
Brown, I. F., Schroeder, W., Setzer, A., Maldonado, M., Pantoja, N., Duarte, A., & Marengo, J.
(2006). Monitoring fires in southwestern Amazonian rain forests. EOS Transactions, 87(26),
253–259.
Burden, R.  L., Faires, J.  D., & Reynolds, A.  C. (1981). Numerical analysis (2nd ed.). Boston:
Prindle, Weber and Schmidt.
Chander, G., Markham, B. L., & Barsi, J. A. (2007). Revised Landsat 5 thematic mapper radiomet-
ric calibration. IEEE Transactions on Geoscience and Remote Sensing Letters, 4(3), 490–495.

© Springer Nature Switzerland AG 2019 71


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0
72 References

Chander, G., Haque, O., Micijevic, E., & Barsi, J. A. (2010). A procedure for radiometric recali-
bration of Landsat 5 TM reflective-band data. IEEE Transactions on Geoscience and Remote
Sensing, 8(1), 556–574.
Cochrane, M. A., Alencar, A., Schulze, M. D., Souza, C. M., Jr., Nepstad, D. C., Lefebvre, P., &
Davidson, E. A. (1999). Positive feedbacks on the fire dynamic of closed canopy tropical for-
ests. Science, 284, 1832–1835.
Conte, S.  D., & De Boor, C. (1980). Elementary numerical analysis: An algorithmic approach
(p. 445). New York, USA: McGraw-Hill.
Detchmendy, D. M., & Pace, W. H. (1972). A model spectral signature variability for mixtures.
In F. Shahrokhi (Ed.), Remote sensing of Earth resources (Vol. 1, pp. 596–620). Tullahoma:
The University of Tennessee.
Dietz, A., Wohner, C., & Kuenzer, C. (2012). European snow cover characteristics between 2000
and 2011 derived from improved MODIS daily snow cover products. International Journal of
Remote Sensing, 4, 2432–2454.
Dietz, A., Kuenzer, C., & Conrad, C. (2013). Snow cover variability in Central Asia between 2000
and 2011 derived from improved MODIS daily snow cover products. International Journal of
Remote Sensing, 34(11), 3879–3902.
Fearnside, P. M. (1993). Deforestation in Brazilian Amazonia: The effect of population and land
tenure. Ambio, 22(8), 537–545.
Foody, G.  M., Lucas, R.  M., Curran, P.  J., & Honzak, M. (1997). Non-linear mixture model-
ing without end-members using an artificial neural network. International Journal of Remote
Sensing, 18, 937–953.
García-Haro, F., Sommner, S., & Kemper, T. (2005). A new tool for variable multiple endmem-
ber spectral mixture analysis (VMESMA). International Journal of Remote Sensing, 26,
2135–2162.
Gessner, U., Machwitz, M., Esch, T., Tillack, A., Naeimi, V., Kuenzer, C., & Dech, S. (2015).
Multi-sensor mapping of West African land cover using MODIS, ASAR and TanDEM-X/
TerraSAR-X data. Journal of Remote Sensing Environment, 164, 282–297.
Gilabert, M.  A., Conese, C., & Maselli, F. (1994). An atmospheric correction method for the
automatic retrieval of surface reflectances from TM images. International Journal of Remote
Sensing, 15(10), 2065–2086.
Grogan, J., Jennings, S. B., Landis, R. M., Schulze, M., Baima, A. M. V., Lopes, J. C. A., Norghauer,
J. M., Oliveira, L. R., Pantoja, F., Pinto, D., Silva, J. N. M., Vidal, E., & Zimmerman, B. L.
(2008). What loggers leave behind: Impacts on big-leaf mahogany (Swietenia macrophylla)
commercial populations and potential for post-logging recovery in the Brazilian Amazon.
Forest Ecology and Management, 255, 269–281.
Heimes, F.  J. (1977). Effects of scene proportions on spectral reflectance in Lodgepole pine.
Dissertation (Master of Science) – Colorado State University, Fort Collins.
Horwitz, H. M., Nalepka, R. F., Ryde, P. D., & Morgenstern, J. P. (1971). Estimating the propor-
tions of objects within a single resolution element of a multispectral scanner. In Proceedings of
the 7th international symposium on remote sensing of environment, May 7–21, 1971 (pp. 1307–
1320). Ann Arbor: Willow Run Laboratories.
INPE – Instituto Nacional de Pesquisas Espaciais. (2002). Monitoring of the Brazilian Amazonian
forest by satellite, 2000-2001 (p. 21). São José dos Campos, SP, Brasil: Instituto Nacional de
Pesquisas Espaciais.
Klein, I., Gessner, U., & Kuenzer, C. (2012). Regional land cover mapping in Central Asia using
MODIS time series. Applied Geography, 35, 1–16.
Klein, I., Dietz, A., Gessner, U., Dech, S., & Kuenzer, C. (2015). Results of the global waterpack:
A novel product to assess inland water body dynamics on a daily basis. Remote Sensing Letters,
6(1), 78–87.
Kuenzer, C., Klein, I., Ullmann, T., Foufoula-Georgiou, E., Baumhauer, R., & Dech, S. (2015).
Remote sensing of river delta inundation: Exploiting the potential of coarse spatial resolution,
temporally-dense MODIS time series. Remote Sensing, 7, 8516–8542.
References 73

Leinenkugel, P., Wolters, M., Oppelt, N., & Kuenzer, C. (2014). Tree cover and forest cover dynam-
ics in the Mekong Basin from 2001 to 2011. Remote Sensing of Environment, 158, 376–392.
Lu, L., Kuenzer, C., Guo, H., Li, Q., Long, T., & Li, X. (2014). A novel land cover classification
map based on MODIS time-series in Nanjing, China. Remote Sensing, 6, 3387–3408.
Lu, L., Kuenzer, C., Wang, C., Guo, H., & Li, Q. (2015). Evaluation of three MODIS-derived
vegetation index time series for dry land vegetation dynamics monitoring. Remote Sensing, 7,
7597–7614.
Mahar, D. (1988). Government policies and deforestation in Brazil’s Amazon Region. Washington,
D.C.: World Bank.
Matricardi, E.  A. T., Skole, D.  L., Pedlowski, M.  A., Chomentowski, W., & Fernandes, L.  C.
(2010). Assessment of tropical forest degradation by selective logging and fire using Landsat
imagery. Remote Sensing of Environment, 114, 1117–1129.
Moran, E. F. (1981). Developing the Amazon. Bloomington: Indiana University Press.
Moran, E. F. (1993). Deforestation and land use in the Brazilian Amazon. Human Ecology, 21(1),
1–21.
Moran, E. F., Brondizio, E., Mausel, P., & Wu, Y. (1994). Integrating Amazonian vegetation, land-­
use, and satellite data. Bioscience, 44(5), 329–338.
Nepstad, D.  C., Klink, C.  A., Uhl, C., Vieira, I.  C., Lefebvre, P., Pedlowski, M., Matricardi,
E., Negreiros, G., Brown, I.  F., Amaral, E., Homma, A., & Walker, R. (1997). Land-use in
Amazonia and the Cerrado of Brazil. Ciência e Cultura Journal of the Brazilian Association
for the Advancement of Science, 49(1/2), 73–86.
Novo, E.  M. L.  M., & Shimabukuro, Y.  E. (1994). Spectral mixture analysis of inland tropical
waters. International Journal of Remote Sensing, 15(6), 1354–1356.
Pace, W. H., & Detchmendy, D. M. (1973). A fast algorithm for the decomposing of multispectral
data into mixtures. In F. Shahrokhi (Ed.), Remote sensing of Earth resources (Vol. 2, pp. 831–
847). Tullahoma: The University of Tennessee.
Pearson, R. (1973). Remote multispectral sensing of biomass. Dissertation (Ph.D.)  – Colorado
State University, Fort Collins.
Piromal, R. A. S. (2006). Avaliação do modelo 5-scale para simular valores de reflectância de
unidades de paisagem da Floresta Nacional do Tapajós. 151 f. Dissertação (Mestrado em
Sensoriamento Remoto) – São José dos Campos, Inpe. INPE-14645-TDI/205.
Ranson, K. J. (1975). Computer assisted classification of mixtures with simulated spectral signa-
tures. Dissertation (Master of Science) – Colorado State University, Fort Collins.
Roberts, D. A., Smith, M., & Adams, J. (1993). Green vegetation, nonphotosynthetic vegetation,
and soils in AVIRIS data. Remote Sensing of Environment, 44, 255–269.
Rosin, P. (2001). Robust pixel unmixing. IEEE Transactions on Geoscience and Remote Sensing,
39, 1978–1983.
Shimabukuro, Y.  E. (1987). Shade images derived from linear mixing models of multi­spectral
measurements of forested areas. (274 p.) Thesis (Doctor of Phylosophy)  – Colorado State
University, Fort Collins.
Shimabukuro, Y. E., & 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, 16–20.
Shimabukuro, Y. E., & Smith, J. A. (1995). Fraction images derived from Landsat TM and MSS
data for monitoring reforested areas. Canadian Journal of Remote Sensing, 21(1), 67–74.
Shimabukuro, Y. E., Batista, G. T., Mello, E. M. K., Moreira, J. C., & Duarte, V. (1998). Using
shade fraction image segmentation to evaluate deforestation in Landsat Thematic Mapper
images of the Amazon region. International Journal of Remote Sensing, 19(3), 535–541.
Shimabukuro, Y.  E., Duarte, V., Anderson, L.  O., Valeriano, D.  M., Arai, E., & Freitas, R.  M.
(2006). Near real time detection of deforestation in the Brazilian Amazon using MODIS imag-
ery. Revista Ambi‑Agua, 1, 37–47.
Shimabukuro, Y. E., Duarte, V., Arai, E., Freitas, R. M., Lima, A., Valeriano, D. M., Brown, I. F.,
& Maldonado, M. L. R. (2009). Fraction images derived from Terra MODIS data for mapping
burnt areas in Brazilian Amazonia. International Journal of Remote Sensing, 30, 1537–1546.
74 References

Shimabukuro, Y. E., Beuchle, R., Grecchi, R. C., & Achard, F. (2014). Assessment of forest deg-
radation in Brazilian Amazon due to selective logging and fires using time series of fraction
images derived from Landsat ETM+ images. Remote Sensing Letters, 5, 773–782.
Singer, R.  B., & Mccord, T.  B. (1979). Mars: Large scale mixing of bright and dark materials
and implications for analysis of spectral reflectance. In Proceedings of the 10th lunar and
planetary science conference, Houston, Texas (pp. 1825–1848). Houston: Lunar and Planetary
Institute.
Skole, D., & Tucker, C. (1993). Tropical deforestation and habitat fragmentation in the Amazon:
Satellite data from 1978 to 1988. Science, 260, 1905–1910.
Skole, D. L., Chomentowski, W. H., Salas, W. A., & Nobre, A. D. (1994). Physical and human
dimensions of deforestation in Amazonia. Bioscience, 44(5), 14–322.
Slater, P.  N. (1980). Remote sensing, optics and optical systems. Reading: Addison-Wesley
Publishing Company, Advanced Book Program.
Smith, M. O., Johnson, P. E., & Adams, J. B. (1985). Quantitative determination of mineral types
and abundances from reflectance spectra using principal component analysis. Journal of
Geophysical Research, 90(S02), 797–804.
Spiegel, M. R. (1968). Mathematical handbook of formulas and tables. New York: McGraw-Hill.
Tardin, A.  T., Lee, D.  C. L., Santos, R.  J. R., Assis, O.  R., Barbosa, M.  P. S., Moreira, M.  L.,
Pereira, M. T., Silva, D., & Filho, C. P. (1980). Subprojeto desmatamento, Convênio IBDF/
CNPq/INPE, Rel. In Técnico INPE-1649-RPE/103.
Ustin, S. L., Adams, J. B., Elvidge, C. D., Rejmanek, M., Rock, B. N., Smith, M. O., Thomas,
R. W., & Woodward, R. A. (1986). Thematic mapper studies of semiarid shrub communities.
Bioscience, 36(7), 446–452.

Recommended References

Adams, J. B., & Gillespie, A. R. (2006). Remote sensing of landscapes with spectral images: A
physical modelling approach (Vol. 378). Cambridge: Cambridge University Press.
Adams, J. B., Smith, M. O., & Gillespie, A. R. (1993). Imaging spectroscopy: Interpretation based
on spectral mixture analysis. In C. M. Pieters & P. A. Englert (Eds.), Remote geochemical anal-
ysis: Elemental and mineralogical composition (Cap. 7, pp. 145–166). New York: Cambridge
University Press.
Adams, J.  B., Sabol, D.  E., Kapos, V., Almeida-Filho, R., Roberts, D.  A., Smith, M.  O., &
Guillespie, A.  R. (1995). Classification of multispectral images based on fractions of end-
members: Application to land-cover change in the Brazilian Amazon. Remote Sensing of
Environment, 52, 137–152.
Aguiar, A. (1991). Utilização de atributos derivados de proporções de classes dentro de um ele-
mento de resolução de imagem (‘pixel’) na classificação multiespectral de imagens de senso-
riamento remoto. Dissertação (Mestrado) – Instituto Nacional de Pesquisas Espaciais – INPE,
São José dos Campos.
Aguiar, A. P. D., Shimabukuro, Y. E., & Mascarenhas, N. D. A. (1999). Use of synthetic bands
derived from mixing models in the multispectral classification of remote sensing images.
International Journal of Remote Sensing, 20(4), 647–657.
Alcântara, E. H., Stech, J. L., Novo, E. M. M., Shimabukuro, Y. E., & Barbosa, C. C. F. (2008).
Turbidity in the Amazon floodplain assessed through a spatial regression model applied to
fraction images derived from MODIS/Terra. IEEE Transactions on Geoscience and Remote
Sensing, 46, 2895–2905.
Anderson, L. O., Shimabukuro, Y. E., & Arai, E. (2005a). Cover: Multitemporal fraction images
derived from Terra MODIS data for analysing land cover change over the Amazon region.
International Journal of Remote Sensing, 26(11), 2251–2257.
References 75

Anderson, L. O., Shimabukuro, Y. E., Defries, R. S., & Morton, D. (2005b). Assessment of defor-
estation in near real time over the Brazilian Amazon using temporal fraction images derived
from Terra MODIS. IEEE Geoscience and Remote Sensing Letters, 2(3), 315–318.
Anderson, L.  O., Aragão, L.  E. O.  C., Lima, A., & Shimabukuro, Y.  E. (2005c). Detecção de
cicatrizes de áreas queimadas baseada no modelo linear de mistura espectral e imagens índice
de vegetação utilizando dados multitemporais do sensor MODIS/Terra no Estado do Mato
Grosso, Amazônia Brasileira. Acta Amazonica, Manaus, AM, 35(4), 445–456.
Carreiras, J. M. B., Shimabukuro, Y. E., & Pereira, J. M. C. (2002). Fraction images derived from
SPOT-4 VEGETATION data to assess land-cover change over the State of Mato Grosso, Brazil.
International Journal of Remote Sensing, 23(23), 4979–4983.
Cochrane, M. A., & Souza, C. M. (1998). Linear mixture model classification of burned forest in
the Eastern Amazon. International Journal of Remote Sensing, 19(17), 3433–3440.
Cross, A., Settle, J. J., Drake, N. A., & Paivinen, R. T. M. (1991). Subpixel measurement of tropical
forest cover using AVHRR data. International Journal of Remote Sensing, 12(5), 1119–1129.
Ferreira, M. E. (2003). Análise do modelo linear de mistura espectral na discriminação de fitofi-
sionomias do Parque Nacional de Brasília (Bioma Cerrado). 111 f. Dissertação – Instituto de
Geociências, Departamento de Geologia Geral e Aplicada da Universidade de Brasília, UnB,
Brasília.
Ferreira, M. E., Ferreira, L. G., Sano, E., & Shimabukuro, Y. E. (2007). Spectral linear mixture
modelling approaches for land cover mapping of tropical savanna areas in Brazil. International
Journal of Remote Sensing, Grã-Bretanha, 28(2), 413–429.
García-Haro, F., Gilabert, M., & Meliá, J. (1996). Linear spectral mixture modelling to estimate
vegetation amount from optical spectral data. International Journal of Remote Sensing, 17(17),
3373–3400.
Haertel, V. F., & Shimabukuro, Y. E. (2005). Spectral linear mixing model in low spatial resolution
image data. IEEE Transactions on Geoscience and Remote Sensing, 43(11), 2555–2562.
Haertel, V., Shimabukuro, Y. E., & Almeida-Filho, R. (2004). Fraction images in multitemporal
change detection. International Journal of Remote Sensing, 25(23), 5473–5489.
Hall, F. G., Shimabukuro, Y. E., & Huemmrich, K. F. (1995). Remote sensing of forest biophysi-
cal structure in boreal stands of Picea mariana using mixture decomposition and geometric
reflectance models. Ecological Applications, 5, 993–1013.
Hlavka, C., & Spanner, M. (1995). Unmixing AVHRR imagery to assess clearcuts and forest
regrowth in Oregon. IEEE Transactions on Geoscience and Remote Sensing, 33(3), 788–795.
Holben, B., & Shimabukuro, Y. E. (1993). Linear mixing model applied to coarse spatial resolu-
tion data from multispectral satellite sensors. International Journal of Remote Sensing, 14(11),
2231–2240.
Jasinski, M., & Eagleson, P. (1990). Estimation of subpixel vegetation cover using red-infrared
scattergrams. IEEE Transactions on Geoscience and Remote Sensing, 28(2), 253–267.
Kawakubo, F.  S. (2010). Metodologia de classificação de imagens multiespectrais aplicada ao
mapeamento do uso da terra e cobertura vegetal na Amazônia: Exemplo de caso na região de
São Félix do Xingu, sul do Pará. 129 f. Tese (Doutorado em Geografia Física) – Faculdade de
Filosofia, Letras e Ciências Humanas (FFLCH), Universidade de São Paulo, São Paulo.
Keshava, N. (2003). A survey of spectral unmixing algorithms. Lincoln Laboratory Journal, 14,
55–78.
Keshava, N., & Mustard, J.  F. (2002). Spectral unmixing. IEEE Processing Magazine, 19(2),
44–57.
Kruse, F., Lefkoff, A. B., Boardman, J. W., Heidebrecht, K. B., Shapiro, A. T., Barloon, P. J., &
Goetz, A. F. (1993). The spectral image processing system (SIPS) – interactive visualization
and analysis of imaging spectrometer data. Remote Sensing of Environment, 44, 145–163.
Lobell, D. B., & Asner, G. P. (2004). Cropland distributions from temporal unmixing of MODIS
data. Remote Sensing of Environment, 93(3), 412–422.
Lu, D., Moran, E., & Batistella, M. (2003). Linear mixture model applied to Amazonian vegetation
classification. Remote Sensing of Environment, New York, 87, 456–469.
76 References

Mcgwire, K., Minor, T., & Fenstermaker, L. (2000). Hyperspectral mixture modeling for quan-
tifying sparse vegetation cover in arid environments. Remote Sensing of Environment, 72(3),
360–374.
Mertes, L., Smith, M., & Adams, J.  (1993). Estimating suspended sediment concentrations
in surface waters of the Amazon river wetlands from Landsat images. Remote Sensing of
Environment, 43(3), 281–301.
Novo, E. M., & Shimabukuro, Y. E. (1997). Identification and mapping of the Amazon floodplain
habitats using a mixing model. International Journal of Remote Sensing, 18(3), 663–670.
Peddle, D., Hall, F., & Ledrew, E. (1999). Spectral mixture analysis and geometric-optical reflec-
tance modeling of boreal forest biophysical structure. Remote Sensing of Environment, 67(3),
288–297.
Poulet, F., & Erad, S. (2004). Nonlinear spectral mixing: Quantitative analysis of laboratory min-
eral mixtures. Journal of Geophysical Research, 109(E2), 1–12.
Quarmby, N., Townshend, J. R., Settle, J. J., & White, K. H. (1992). Linear mixture modelling
applied to AVHRR data for crop area estimation. International Journal of Remote Sensing,
13(3), 415–425.
Quintano, C., Fernández-Manso, A., Shimabukuro, Y. E., & Pereira, G. (2012). Spectral unmixing.
International Journal of Remote Sensing, 33, 5307–5340.
Quintano, C., Fernández-Manso, A., Fernández-Manso, O., & Shimabukuro, Y. E. (2006). Mapping
burned areas in Mediterranean counties using spectral mixture analysis from a uni-temporal
perspective. International Journal of Remote Sensing, 27(4), 645–662.
Quintano, C., Shimabukuro, Y.  E., Fernández, A., & Delgado, J.  A. (2005). A spectral unmix-
ing approach for mapping burned areas in Mediterranean countries. International Journal of
Remote Sensing (Print), 26(7), 1493–1498.
Richardson, A. J., & Wiegand, C. L. (1977). Distinguishing vegetation from soil background infor-
mation. Photogrammetric Engineering and Remote Sensing, New York, 44, 1541–1552.
Roberts, D. A., Numata, I., Holmes, K., Chadwick, O., Batista, G., & Krug, T. (2002). Large area
mapping of land-cover change in Rondônia using multitemporal spectral mixture analysis and
decision tree classifiers. Journal of Geophysical Research, 107(D20), 40.001–40.017.
Rudorff, C., Novo, E., & Galvao, L. (2006). Spectral mixture analysis of inland tropical Amazon
floodplain waters using EO-1 Hyperion. In IEEE international symposium on geoscience and
remote sensing. (pp. 128–133). Denver, CO, USA: IEEE IGARSS.
Settle, J., & Campbell, N. (1998). On the errors of two estimators of subpixel fractional cover
when mixing is linear. IEEE Transactions on Geoscience and Remote Sensing, 36(1), 163–170.
Settle, J., & Drake, N. (1993). Linear mixing and the estimation of ground cover proportions.
International Journal of Remote Sensing, 14(6), 1159–1177.
Shimabukuro, Y. E., Holben, B. N., & Tucker, C. J. (1994). Fraction images derived from NOAA
AVHRR data for studying the deforestation in the Brazilian Amazon. International Journal of
Remote Sensing, 15(3), 517–520.
Shimabukuro, Y. E., Novo, E. M. L. M., & Ponzoni, F. J. (1998). Índice de vegetação e modelo
de mistura espectral no monitoramento do Pantanal. Pesquisa Agropecuária Brasileira (PAB),
33, 1729–1737.
Shimabukuro, Y. E., Almeida-Filho, R., Kuplich, T. M., & Freitas, R. M. (2008). Use of Landsat
TM fraction images to quantify the optical and SAR data relationships for land cover discrimi-
nation in the Brazilian Amazonia. International Journal of Geoinformatics, 4(1), 57–63.
Van Der Meer, F. (1995). Spectral unmixing of Landsat Thematic Mapper data.
International Journal of Remote Sensing, 16(16), 3189–3194.
Index

A incentives, 52
Airborne sensors, 9, 12 infrastructure investments, 51
Applications, fraction images DETER project, 53, 58, 60, 61
burned areas mapping, 62–64 Diffuse radiation, 11
deforestation monitoring (see Digital approaches, 8
Deforestation monitoring) Digital classifications, 6, 7
land use and land cover mapping, 65–67 Digital numbers (DNs)
selective logging detection, 65 airborne sensors, 12
Arithmetic operations, 12 amounts, 11
AVHRR/NOAA sensor, 4 arithmetic operations, 12
classification process, 12
conversion, 12
B diffuse radiation, 11
BASA, 52 EIFOV (see Effective instantaneous field
Brazilian Legal Amazon monitoring program, of view (IFOV))
52, 53 electromagnetic radiation, 9
Brazilian territory, 51 fraction images (see Fraction images)
Burned areas mapping, 62–64 geometric components, 13
IFOV (see Instantaneous field of view (IFOV))
influent factors, 11, 12
C intensities, 9
Constrained least squares (CLS), 27, 28 irradiance, 11
Lmin and Lmax, 10
and L0λ, 10
D MTF, 14
Deforestation monitoring in orbital images, 9, 12
AML, 51, 52 radiance (intensity) values, 9
Brazilian Legal Amazon monitoring radiometric resolution, 9
program, 52, 53 radiometry, 15
Brazilian territory, 51 sensor-generated images, 15
DETER project, 58, 60, 61 and sensor relationship, 9, 10
digital PRODES project (see Digital spectral characterization, 12
PRODES project) spectral mixture, 15
high economic value, 52 TOA radiance, 10

© Springer Nature Switzerland AG 2019 77


Y. E. Shimabukuro, F. J. Ponzoni, Spectral Mixture for Remote Sensing,
Springer Remote Sensing/Photogrammetry,
https://doi.org/10.1007/978-3-030-02017-0
78 Index

Digital PRODES project, 69 monochromatic (grayscale), 45


AML, 54, 56 OLI/Landsat 8, 44
analogical data, 54 photointerpretation, 46
annual deforestation assessments, 54 proportions, 47
Brazilian AML, 54 quantitative information, 47
deforestation information, 56 shade/water (see Shade/water fraction)
IBDF, 54 soil (see Soil fraction)
image edition, 56, 59 spectral data in, 46
image segmentation, 56, 57 spectral mixture model transforms, 43
Landsat TM images, 54 spectral response, 43, 44
LSMM, 55, 56 TM, 46
methodology, 54 vegetation (see Vegetation fraction)
MSS, 54
omission and commission errors, 56
semi-automatized methodology, 54 G
soil and shade/water fraction, 55, 56 Gaussian elimination and replacement
thematic classes, 56, 58 algorithm, 39
thematic map, 56, 59 Geographic information system (GIS), 53
TM image, 56 Geometric components, 13
unsupervised classification method, 56, 57
UTM, 54
vegetation fraction, 55 H
High resolution visible (HRV) program, 17
Hyperion, 21, 22, 46
E
Earth observation sensors, 21, 43
Earth Observing System (EOS), 17, 21 I
Effective instantaneous field of view (EIFOV), Image endmembers, 40, 41, 43
13–15 Image processing systems, 7
Electromagnetic radiation, 1, 2, 9, 11, 15 Image segmentation, 56, 57
Endmembers, 7, 25, 27, 29, 40, 41, 43, 44, Instantaneous field of view (IFOV), 13–16, 22
46–48, 69 Institute for Forest Development (IBDF), 54
Enhanced Thematic Mapper Plus (ETM+) Inversion problem, 26
sensor, 21 Irradiance, 11
Error images, 47–50

J
F Joint Polar satellite System (JPSS), 18
Fraction images, 69
application areas, 46
applications (see Applications, fraction L
images) Landsat, 21
bands, 44, 45 Land use and land cover mapping, 65–67
color composite, 44–47, 49 Least squares approach, 6, 7
data type, 44 Linear models
DNs, 45, 46 application, 7
error images, 47–50 LSMM (see Linear spectral mixture
generation, 43 models (LSMM))
Hyperion, 46 and nonlinear, 69
LSMM, 46 Linear spectral mixture models (LSMM), 8
mathematical algorithms, 43 components proportion, 23
mixture model, 44 endmembers, 25, 40, 41
MODIS, 46 equations, 24
Index 79

fraction images, 24, 25 (see also Fraction Orbital imagery


images) DNs, 9, 12
inversion problem, 26 IFOV, 15
mathematical algorithms multispectral data, 1
CLS, 27, 28 red spectral range, 4
four spectral bands and three Orbital sensors
components, 28–31 Earth’s surface, 17
PC, 39, 40 EOS program, 17
reference spectra/pure pixels, 27 Hyperion, 21, 22
six spectral bands and four Landsat, 21
components, 31–38 MODIS, 18, 19
solar optical spectrum, 27 SPOT vegetation, 18, 20
WLS, 38, 39
numerical inversions, 26
objectives, 69 P
panchromatic image, 24 Panchromatic image, 24, 25
pixel, 23 Pixel
AVHRR, 4
definition, 1
M dispersion, 5
Manaus (AM) region, 4 mixture problem, 6–8
Mathematical algorithms, 43 NOAA/AVHRR sensor, 4
Maximum likelihood approach, 6 overlaps, 2
Mixture problem, 3 and resolution element, 1
approaches, 6 sensor, 4
area estimation, 6 size, 4
classification, 6 water due, 4
digital classifications, 7 Principal components (PC), 39, 40
Earth’s surface, 3 Principal components analysis
endmembers, 7 (PCA), 40
image processing systems, 7
least squares approach, 7
linear spectral mixture models, 8 Q
MSS image, 7 Qualitative approach, 12
no classification of pixel, 6 Quasi-closed solution, 27
radiant flux, 1
sensors, 3
spectral responses of vegetation, 7 R
TM image, 7 Radiant flux, 1
TM/Landsat 5 sensor, 7 Radiometry
Viking Lander 1 sensor, 7 12-bit resolution, 22
MODIS, 18, 19, 45, 46, 53, 56, 58–67 8-bit sensor, 24
Modulation transfer function (MTF), 14 DNs defines, 9
Multispectral Scanner System (MSS), 4, 7, 17, physical unities, 43
21, 27, 28, 54 road reflectances, 15
TM and ETM+ sensors, 21
vegetation cover, 15
N Reference endmembers, 40, 41, 43
Numerical inversions, 26 Remote sensing techniques
application, 1
development, 20
O images, 52–54
Operational Land Imager (OLI) sensor, 17, 21, users, 10
44, 45, 47, 49 Resolution element, 1–3
80 Index

S natural resources, 5
Selective logging detection, 65 orbital images (see Orbital imagery)
Shade/water fraction, 43–49, 54, 55, 57, 60, pixel, 1
62–66, 69 radiant flux, 1
Soil fraction, 43–49, 54–56, 60–63, 65, 66, 69 resolution element, 1–3
Solar optical spectrum, 27 sensors, 3
Spatial resolution TM/Landsat 5 sensor, 4
average, 18 types of land cover, 6
AVHRR sensor, 4 SPOT vegetation, 18, 20
and classes, 3 SUDAM, 52
definition, 13 Suomi National Polar-orbiting Partnership
dimensions, 23 (Suomi NPP), 18
elements, 13, 14
geometric components, 13
low and moderate, 18 T
medium, 23, 52 Thematic Mapper (TM) sensor, 4, 21
MODIS images, 60 TM image, 7
sensor, 3 TM/Landsat 5 sensor, 4, 7
and spectral bands, 54 TOA radiance, 10
TM sensor, 4
Spectral characterization, 12
Spectral mixture U
application, remote sensing techniques, 1 Unsupervised classification method, 56, 57
AVHRR/NOAA sensor, 4
characteristics, 4
concept of, 2 V
dispersion, image pixels, 4, 5 Vegetation fraction, 43–49, 54, 55, 60, 62, 63,
DNs, 15 65, 66, 69
elements/objects, 5 Viking Lander 1 sensor, 7
least-squares approach, 6 Visible Infrared Imaging Meter Suite (VIIRS),
linear model, 69 18
LSMM, 69
Manaus (AM) region, 4
maximum likelihood approach, 6 W
mixture problems, 1–3, 6–8 Weighted least squares (WLS), 38, 39, 56

You might also like