The Evolution of Landsat Data Analysis 1

by David Landgrebe
Professor of Electrical & Computer Engineering
Purdue University
West Lafayette IN 47907-1285
landgreb@ecn.purdue.edu

Preface. In this paper a description is presented of how the multispectral data analysis
technology, which has come to be synonymous with Landsat, was begun and how it developed
and spread among the broader research and user community. The paper is concluded with brief
remarks about key factors which moderated the development and what the future may hold for its
further development. To describe a 25 year long activity as varied and complex as the evolution of
the Landsat related data analysis is a daunting task. Many significant events must be omitted and
others only briefly mentioned. Example activities have been chosen which occurred early and led
to perhaps the largest impact in the development of the technology.

Background: Setting the Stage. The people of the U. S. and much of the
world were surprised, one might even say shocked, when the Soviet Union
successfully launched the first Earth orbiting artificial satellite in October 1957.
The impact on the U. S. was immediate and substantial, affecting tangibles such
as the U. S. education system, but also intangibles such as how this society
thought of itself. One of the significant responses to this event was the National
Aeronautics and Space Act in 1958 which created NASA from the National
Advisory Committee on Aeronautics. These events resulted in people in various
organizations beginning to consider how an ability to operate in space could be
used.

In addition to telecommunications, one of the earliest thoughts relative to Earth

observation from space turned to the weather, and this quickly resulted in the
launch by NASA of TIROS I, the first Earth observational satellite designed for
that purpose on April 1, 1960. During the 20 year period of the '60's and '70's,
some 40 satellites for making weather observations were launched, an average
of two per year.

Probably no one can pinpoint when or where the idea for Landsat originated,
because the idea for it surely evolved over some time and from many different
sources. Indeed, there were many anecdotal stories about the idea. For
example, one source was said to be the weather satellite sensor supplier who
said that there were imaging tubes left over from weather satellites and that it
would be easy to re-adjust their dynamic range for the (darker) Earth surface as
compared to the (brighter) clouds; this, when young NASA was eagerly looking
for useful things to be done with space technology. Another was in that same
context of looking for new missions, in beginning to conceive of exploration
missions to Mars. There were said to be discussions about how close to Mars
an observation satellite would need to come to tell if there were intelligent life
on Mars. In seeking an answer, thought was given as to how close one would
need to come to the Earth to answer the same question.

But perhaps the most significant motivation for beginning to research space-
based remote sensing of the Earth's surface was the growing awareness during

1 Prepared for Photogrammetric Engineering and Remote Sensing, Vol. 63, No. 7, July 1997,
pp. 859-867, Special Issue commemorating the 25th anniversary of the launch of Landsat 1,
July 1972. Reproduced with permission, the American Society for Photogrammetry.
Landgrebe: Evolution of Landsat Data Analysis

this time of the finiteness of the Earth and the need to better manage its
resources. In this context interests tended to be spoken of in categories referred
to as non-renewable and renewable resources. Geological surveys for locating
mineral deposits were an example of the former group, while agriculture was an
important example of the latter. Indeed, agriculture was deemed especially in
need of better survey and management techniques because of its dynamics,
e.g. with regard to such factors as weather and disease. Then-current survey
methods for gathering resource management information were simply not able
to respond quickly enough and in an economically suitable fashion to provide
credible quantitative data.

Research Leading to Landsat. This need motivated the National Research

Council in 1961 to form a study committee entitled the Committee on Remote
Sensing for Agricultural Purposes under the chairmanship of J. Ralph Shay of
Purdue University. Two other members of this study group were R. N. Colwell of
the University of California at Berkeley and M. R. Holter of the University of
Michigan. These three individuals, leading research groups at their respective
institutions, organized a cooperative interdisciplinary, inter-institution research
team among the three institutions funded jointly by NASA and the U. S.
Department of Agriculture.

Following several years of preliminary feasibility studies, this effort got formally
under way in early 1966. Professor Colwell, of the University of California,
Berkeley Forestry Department was already a highly respected and
knowledgeable leader in the use of air photo interpretation in forestry and
agriculture. Mr. Holter headed a group2 at the Willow Run Laboratories of the
University of Michigan which was engaged in airborne sensing for the U. S.
Army Electronics Command under Project MICHIGAN. This group operated an
airborne photographic system which, in addition, included some capability to
generate line-scanned imagery in the infrared region. They were also involved
in devising photointerpretive methods for data analysis. The Purdue University
Agricultural Experiment Station supported this effort by making available
personal and a number of diverse experimental stations adjacent to the
University, and also involving personal from their School of Electrical
Engineering to begin devising digital computer means for analyzing such data.

Early thoughts about how to analyze data quite naturally turned to photo
interpretation. However, the added possibility of several bands in the infrared
increased the complexity of this process. For example, an early data collection
tool was the "nine lens camera," which formed three rows of three images on a
sheet of 9 inch film, each lens containing a different filter passing a different part
of the spectrum. Developing a photointerpretive key relative to nine different
images was seen to be a daunting task. To develop a quantitative version of the
data, thought was given to using a film densitometer to quantify the film density
at a given point in each image.

However, in 1966, the University of Michigan group succeeded in extending the

range of the infrared scanner on their aircraft into additional regions of the
spectrum. The result was a scanner that could collect electronic data

2 In 1973 this group separated from the University of Michigan and became the Environmental
Research Institute of Michigan (ERIM).
- 2 -
Landgrebe: Evolution of Landsat Data Analysis

simultaneously in up to 18 spectral bands extending from 0.32 µm in the uv to

14 µm in the thermal infrared. The data were recorded on a multiband analog
tape recorder. This system continued to be modified and improved over the next
several years and eventually became known as the M-7 scanner.3 This system
became the mainstay of this and many later research efforts.

The Purdue group succeeded in developing a capability to simultaneously

electronically sample and digitize all bands of the Michigan scanner analog
tapes, thereby producing individual vectors for each pixel. This made possible,
and indeed convenient, data analysis based upon digital multivariant statistical
methods. Because the early form of the Michigan scanner was actually made up
of two double ended scanners, thus producing four different data sets per
flightline, the Purdue group also created an early capability for the precision
registration of such data sets, concatenating the four sets of measurements into
one vector for each pixel.

The Multispectral Concept. When beginning such an endeavor as this, it is

logical to think first of use of images and image processing methods. The
decision to base analysis methods not on image characteristics but on spectral
characteristics was motivated by several considerations. At the time,
computational capabilities were much more limited than now, as was the state
of image processing technology. The use of spatially based approaches with
their inherently greater complexity, did not seem like it could lead to practical,
usable technology in the near term. Rather, a spectral approach, given the
Michigan system's ability to deliver a fairly complete spectral sampling for each
pixel allowed for a basic simplicity which was an important practical
characteristic.

But perhaps more significantly was the matter of economics and spatial
resolution. There was a strong intention in the research program to produce a
technology that was not any more expensive to devise and use than necessary.
To base the proposed technology on image characteristics would have required
high spatial resolution for such materials as agricultural crops. Spatial
resolution is one of the more expensive parameters of such a system. Data
volume increases as the square of resolution. Pointing precision also increases
rapidly with spatial resolution, as does the required downlink capacity. Thus,
focusing on spectral characteristics while using the lowest spatial resolution
necessary seemed like the optimum approach, if it could be made to work.

In considering approaches to the analysis of such data, one concern at the time
was the great volume of data to be processed. Neural network analysis was
popular then, but it had (and has) the disadvantage of requiring substantial
computing time involved with training the algorithm each time. Since in the case
of remotely sensed data, it was envisioned that the analysis algorithm would
need to be retrained for each new data set, this was seen as a significant
disadvantage. There was also the problem of a paucity of analytical tools
related to it.

3 The Michigan M-7 Airborne Scanner is described in more detail in Swain and Davis, 1978. See
also Hasell, 1972, 1974.
- 3 -
Landgrebe: Evolution of Landsat Data Analysis

To handle the large volume of data, analog computer processing was also
considered briefly, however, the rapid development of digital computation at
that time, together with its inherent flexibility quickly favored the digital
approach. Thus the research effort became focused upon basing the derivation
of information for managing Earth resources using a so-called single aperture
scanner coupled with digital statistical pattern recognition data analysis
methods.

Promising preliminary results came within the first few months of the effort4 and
a broader, more fundamentally based research effort was formulated. This
included not only research on understanding the spectral response of materials
in laboratory and field environments, but broadening to include additional user
disciplines beyond agriculture.

In 1967, the National Research Council conducted a major summer study

entitled "Useful Applications of Earth-Oriented Satellites." This study was
intended to cover the full range of possible space applications through 13
separate panels, including such applications as several forms of
telecommunication, navigation and traffic control, and geodesy and
cartography. One panel was also included on economic analysis. Ones related
to remote sensing included the discipline panels on Forestry-Agriculture-
Geography, Geology, Hydrology, Meteorology, Oceanography, and Sensor and
Data Systems. The study illustrates the breadth with respect to the various
disciplines to which interest had spread by that time. Discussion also took place
with regard to both photographic film and electronic data collection and a wide
variety of optical and microwave sensor configurations. A sense emerged that
first generation systems would need to be image oriented with the scale tipping
toward electronic imaging devices with electronic data transmission to Earth,
but a second generation approach using multispectral line scan sensors was
already seen as appropriate. This study provided a major impetus for devising a
satellite series to be known as Earth Resources Technology Satellites (and later
as Landsat), the design of which began in earnest within the next year.
However, though there were promising preliminary results existing by this time,
clearly much more research was needed to be ready for the launch of a land-
looking satellite series.

The Structure of the Remote Sensing Research Effort. It was early

recognized that to create the new remote sensing technology and begin its
utilization would require a multidisciplinary, or more ideally an interdisciplinary,
effort. Figure 1 illustrates the ideal but difficult relationships that had to be
achieved for the optimal development environment, as well as the relationship
between engineering, science, and applications objectives.

4 See, for example, Landgrebe and Staff, 1967 soon followed by Fu, et al, 1969.
- 4 -
Landgrebe: Evolution of Landsat Data Analysis

Figure 1 5. The multifaceted nature of remote sensing technology research and development.

Such an effort must be led by engineering research to produce a capability to

collect and analyze data, for the other parts cannot proceed without this
capability. This is coupled with an effort to deepen the understanding of the
scene materials and how they exist in their environment, which, in turn, leads to
research to focus the capabilities and knowledge developed to the various
application situations which would become of interest. Operational uses could
then result. All of these research efforts contribute to the knowledge base for the
field. The significance of the double-ended arrows of the diagram is that they
imply a great deal of two-way communication, thus the desirability for the overall
effort to be interdisciplinary rather than simply multidisciplinary.

It is this scheme that was followed during the 1960's and 70's for research
leading to the various Landsat systems. Engineering research led to significant
innovative advancements in airborne and spaceborne sensor systems as
already mentioned. This, in turn, allowed for significant work in areas such as
pattern analysis algorithms, spectral feature selection algorithms, image
registration methods, and the use of spatial relationships as an augmentation to
spectral features.

Substantial work was also needed to develop suitable laboratory and field
instruments and measurement techniques. Spectroscopy, long used in the
chemistry laboratory, had to be adapted to the rigors of the field with its heat,
dust, changing lighting conditions, and the like, and had to be made capable of
observing the target from varying angles. Laboratory and field studies were
deemed important to achieving a more sound understanding of the way in
which Earth surface materials reflect, transmit, and absorb optical energy as a
function of wavelength. Thus, new instruments had to be specified, designed,
and constructed, and once available, suitable techniques for collecting samples
in the field and in the laboratory had to be developed. Means for calibrating
such data also needed to be created.

As a result of these activities, several extensive data bases of laboratory and in

situ spectra were assembled.6 The intention of these data were not that they
would be used in direct conjunction with aircraft and spacecraft data, since the
problem of reconciliation of the circumstances of its collection with those of the

5 Adapted from the introduction to a special issue of the Proceedings of the IEEE entitled
"Perceiving the Earth's Resources from Space," Vol. 73, No. 6., June 1985.
6 See, for example, http://dynamo.ecn.purdue.edu/~frdata/FRData/ on the World Wide Web.
- 5 -
Landgrebe: Evolution of Landsat Data Analysis

aerospace data were felt to be either insoluble or unnecessarily complex.

Instead, the focus was on scene understanding and building the knowledge
base (Figure 1) so that the most knowledgeable choices could be made when
new experiments, sensors, or processing schemes are to be designed.

Early Research Literature. Being cross-disciplinary, this newly emerging

field was significantly different from those then existing. Also, there were no
research journals directly suited to report new findings. Thus much of the early
work was only discussed and documented in various symposia specially
focused for that purpose. The earliest of these, the then-so-called Michigan
Symposia which was the forerunner of the current ERIM International Symposia
on Remote Sensing of Environment. This series had begun somewhat earlier
focused upon infrared imaging technology, a subject under study for the U. S.
military. They were quickly adapted to the broader thrust including multispectral
sensing. Another set of meetings which served the purpose of documenting and
providing for technical interchanges between the now growing community of
researchers were the NASA annual program review meetings. In 1968, a series
of meetings were begun at Purdue University called the Machine Processing
Symposia focused specifically on methods for digital processing of multispectral
data. The proceedings of all of these series of meetings provide not only a
picture of how this technology emerged, but contain many results and ideas still
relevant today.

It was not long before serial journals on this topic began to be published. The
journal Remote Sensing of Environment was first published in 1969 with
Professor David Simonett, then at the University of Kansas, as the organizing
Editor-in-Chief. It was specifically intended as an interdisciplinary journal, and
as such, was not associated with any of the existing technical societies. The
American Society for Photogrammetry renamed its journal Photogrammetric
Engineering and Remote Sensing adding the "and Remote Sensing" in 1975. It
officially added "and Remote Sensing" to its Society name in 1985. The IEEE
Geoscience Electronics Society changed its name to the IEEE Geoscience and
Remote Sensing Society and changed the name of its journal to the
Transactions on Geoscience and Remote Sensing in 1979. In 1981 the first of
the annual International Geoscience and Remote Sensing Symposia was held
in Washington, DC. A number of additional journals and symposia were begun
in this period in the U. S. and elsewhere.

Various types of books soon began to appear. An early handbook, the Manual
of Remote Sensing (1st edition), was published by the American Society of
Photogrammetry in 1974. The research monograph, Remote Sensing of
Environment, was published by Academic Press in 1976. One of the first books
intended specifically as a textbook, Remote Sensing: The Quantitative
Approach, was published by McGraw-Hill International Book Company in 1978.

Early Tests and Demonstrations of the Technology. Many of the uses

of Landsat technology were envisioned to require detailed thematic maps of
small (e. g., 25 mile by 25 mile) areas. However, the new and really unique
capability that Landsat technology could provide is a way to rapidly and
economically acquire information of a broad scale regional nature. During the
1970's a series of the latter type of use were performed. These served a very
important research function in that they were a way to test the overall

- 6 -
Landgrebe: Evolution of Landsat Data Analysis

technology as it was being developed to see where there were weaknesses.

They also served the additional function of demonstrating to the potential user
community the viability of the new technology.

The first of such broad scale tests of the technology came about in a crisis
environment and became known as the 1971 Corn Blight Watch Experiment
(CBWE) (MacDonald, et al, 1972). During the latter stages of the 1970 growing
season, a pathogen to which essentially all of the U. S. corn crop was
susceptible emerged in the southern U. S. and began to move northward. This
pathogen, called Southern Corn Leaf Blight, developed from airborne spores
and first showed up as brown lesions on the lower leaves of the corn canopy.
The lesions would grow in size and spread upward in the canopy until
ultimately, the entire plant was destroyed. The broad susceptibility of the corn
crop stemmed from the fact that most varieties of corn in the U. S. were hybrids
utilizing a single type of cytoplasm and it was this cytoplasm that was the basis
for the susceptibility. By the time the danger was realized, seed corn for the
1971 year had already been produced, and so it was feared that, if Southern
Corn Leaf Blight spores could over-winter through the 1970-71 winter, it could
spread rapidly during the 1971 growing season and devastate corn production.

Because of the potentially devastating nature of threat, even though much of the
technology was quite new and untested, it was decided to see if remote sensing
technology could be used to monitor the spread of the blight over the entire U.
S. Corn Belt throughout the 1971 growing season. Unfortunately Landsat was
still more than a year away, and so the experimental test had to be done with
aircraft. A stratified sampling scheme was laid out in which some 200 segments
were to be overflown every two weeks. NASA made available an RB-57F high
altitude reconnaissance aircraft for photographic data collection over these
sites, there being no multispectral sensor system at that time capable of
covering the Corn Belt with that frequency. A map showing the flightlines
containing these segments is given in Figure 2.

Figure 2. 1971 Corn Blight Watch Experiment sample flightlines.

In order to test the new, multispectral technology, a series of 30 segments was

located in the western third of Indiana, and the (formerly University of Michigan)
now ERIM aircraft system, a C-47 platform carrying its multispectral scanner,

- 7 -
Landgrebe: Evolution of Landsat Data Analysis

was assigned to the task. Figure 3 shows the location of these segments in the
Intensive Study Site region.

Figure 3. The sample segments of the CBWE Intensive Study Site.

Even though the CBWE arose due to an anticipated emergency rather than
being motivated by a perceived research need, it proved to be a propitious
research event. The core idea was, in addition to using photographic methods,
to test the ability for a multispectral sensor in conjunction with supervised
pattern and related algorithms. It provided an effective test of a number of
aspects of the technology, just as Landsat 1 (then known as ERTS-1) was in
preparation for launch. Some of these aspects are as follows.

• A challenging discrimination problem. The problem at hand,

identification of the level of the degree of stress in a specific plant
species was a challenging one. It not only required reliably
discriminating between a specific plant species, corn, and all others in
the scene but subdividing that species, determining the level to which
the disease had affected the plants in view. The primary
discrimination algorithm used was the Gaussian maximum likelihood
classification scheme. Calibration and preprocessing matters were
held to a minimum, thus demonstrating that such preprocessing,
which is expensive and time consuming, can be avoided in large
measure.

• Use of a quantitative feature selection algorithm. The multispectral

data was collected in 17 spectral bands over the range from 0.40 to
11.1 µm. Most classifications were done at much lower dimensionality
by using a feature selection algorithm to select the best 4 or so
features with which to do the classification at hand. This procedure
worked very well. It also provided insight in what bands were the most
useful for a problem of this nature. For example, it is common today to
see people dismiss the thermal band of Thematic Mapper data
- 8 -
Landgrebe: Evolution of Landsat Data Analysis

"because of its poor resolution," and yet the thermal band of the
CBWE data was at lower spatial resolution than the others, but over
the growing season, it was the band most frequently selected by the
feature selection algorithm.

• Stratified sampling methods. Such sampling methods were then and

are now commonly used. However, this was perhaps one of the most
comprehensive uses, certainly in remote sensing applications. The
area involved extended over seven states, and an entire growing
season. The intensive study area involved a third of a state. This
approach proved the significance of such a technique to remote
sensing. So called "wall-to-wall" sensing was not necessary and
indeed would have been wasteful. Further, this problem could not
have been done with a sensor of the spatial resolution of, for
example, the AVHRR (Advanced Very High Resolution Radiometer) if
it had been available, because the pixels would have been too large,
and the effect to be measured too subtle.

• Variability aspects. The CBWE problem provided a unique

opportunity to become familiar with the degree of variability that exists
in the scene over a growing season and over a large geographical
area. The collection of such high spatial resolution data over that
many sites and times probably has rarely if ever been duplicated.

• Comparison with then-conventional remote sensing methods. The

fact that the new multispectral remote sensing was being used along
side of more conventional air photo interpretation on such an
extensive scale was a unique opportunity to judge the viability and
potential of the new technology. The two technologies are clearly
different with different strengths and weaknesses. In terms of final
conclusions for the detectability of corn blight, the general post-
experiment evaluation was that the photointerpretation of
photography was not as sensitive, making possible discrimination to a
yes-or-no level of detail, while the multispectral data allowed for
reliable discrimination into three classes, little or none, moderate, or
substantial to severe blight.

• Technology Robustness. Was the technology sufficiently developed

that a large group could rapidly learn to use the technology? The
short term and emergency nature of the manner in which the problem
arose, where there was not time draw up budgets, reallocate funds,
hire people with specific skills and the like, meant that the work had to
be done by people already in place, if for other purposes. Some 1000
people from 17 federal and state agencies were ultimately directly
involved.

• Many operational aspects. In carrying out such a far flung activity with
the time schedule of data collection and analysis cycles which must
be completed for 230 segments every two weeks using ground
observations gathered each cycle from over a seven state area using
personnel already in place at the various sites resulted in many
operational matters that had to be dealt with. For example, the

- 9 -
Landgrebe: Evolution of Landsat Data Analysis

scheduling of missions over the various sample segments relative to

the expected cloud conditions resulted in a substantial increase in
understanding about cloud relationships to remote sensing.

Though the CBWE and its impact on the preparation for the launch of the
Landsat series can only be briefly outlined here, it is perhaps clear from the
above that this experiment made a very significant contribution to the
technology needed for such a satellite program and gave confidence that this
technology was indeed ready. The data set collected, with its large number of
spectral bands collected over a large area throughout an entire growing season
has probably not been equaled to this day.

One of the remaining problems at the time, then, was that the community of
people who possessed this technology at that point was very much too small
relative to the potential for use which such a system had. This pointed to the
need for substantial efforts in so called technology transfer. This point will be
addressed further shortly.

There were many other activities that contributed to the preparation for the
launch of ERTS-1 (Landsat 1). One of the important ones, especially with regard
to providing a wide opportunity for involvement with it was the call for proposals
to analyze data from it. Such a call was issued as the time for first launch
approached. There were more than 700 proposals submitted to this call, with
more than 400 ultimately being funded.

The Landsat Era. The Landsat era really can be said to have begun with the
launch of ERTS-1 in July 1972. As is well known, the first series contained a set
of Return Beam Videcons (RBV) which were for the purpose of producing
framed images, whose geometry was more familiar to photogrammetrists. ERTS
also contained the Multispectral Scanner (MSS), the first generation
spaceborne multispectral sensor system. MSS must certainly be seen as a
primitive multispectral device, both in terms of today's capabilities and in terms
of the ERIM aircraft system which led to it. It had only four bands that were quite
broad and closely spaced, and the signal-to-noise ratio could only support a 6-
bit dynamic range, i.e., 64 shades of gray per band. It could not have supported
a problem as subtle as the corn blight monitoring. Nevertheless, it broke new
ground. It demonstrated to skeptics that line scanning devices were viable.
Indeed, not only did the MSS become essential as a result of the early failure of
the RBV system on ERTS-1, but the RBV failure resulted in an acceleration in
the rate at which techniques for dealing with the new geometry of line scan
devices presented to cartographers. Further, the demand for data in digital tape
form, as compared to hard copy film, grew well beyond planned levels.

Figure 4 shows a time line for some major events of the Landsat era. It shows
the preparation period described above including the first regional survey, the
1971 Corn Blight Watch Experiment, as well as the times of later launches.

- 10 -
Landgrebe: Evolution of Landsat Data Analysis

s
on
t

ti

s
ep

ch o n
y
da

y
e
nc

ch
ve

un t i
h rv
en

ch

ch

re
La d a
Co

un
Re nc Su

ur
m

un

ilu
un

ch

6 h
n
lS

La
om

c
t me
al

au al

La

Fa
un
La

s a un
na
tr

2n I L ion

4
ec

La
sa m

n d La
3

5
ec

t
nd c o
IR

TS eg

sa
gi

t
1
sp

La 2
IS
sa

sa
La R e

nd
ER t R

T
TS

T
IR
ti

nd

nd

O
La
ul

TM

V
d
s

SP
ER

La

SP
La
1
M

A
65 70 75 80 85 90 95

Year

CIVIL LAND REMOTE SENSING MILESTONES

Figure 4. A timeline for the Landsat era.

A second regional-sized survey, one of the first done using Landsat, or more
specifically, the MSS, was begun in 1973. In the early 1970's there was
growing concern about the Great Lakes, and the anthropogenic effects on their
water quality. The U.S.- Canada International Joint Commission, in an effort to
begin to solve the problem, indicated a need for land use maps of the Great
Lakes drainage basin. As a result, a project was funded during the first year of
Landsat to generate land use maps for each of the 192 counties of the US
portion of the drainage basin. The region involved is shown in Figure 5.

Figure 5. The Great Lakes drainage basin and the boundaries of the US counties involved.

The project resulted in thematic maps being made using MSS data and
supervised maximum likelihood classification for each county in classes of the
Anderson Level 1 land use classification system, and tables for Anderson Level
1 and 2 classes, showing the approximate proportion of each in each county.
The total cost of the project was approximately $2500 per county, a cost much
below that possible by any other means. This demonstrated the viability of the
Landsat sensor and analysis technology, although it was clearly limited by the

- 11 -
Landgrebe: Evolution of Landsat Data Analysis

limited spectral capabilities of the MSS. A more complete description of the

project and the methodology used is given in Swain and Davis, 1978.

In addition to the diverse and numerous smaller demonstrations of what the

Landsat technology could do, there was a series of larger scale programs that
were conducted during the 1970's and 80's. Notably among there were LACIE
(the Large Area Crop Inventory Experiment), and AgRISTARS (Agriculture and
Resources Inventory Surveys Through Aerospace Remote Sensing).

LACIE was begun in November 1974 and was intended to test and demonstrate
a real time capability for inventorying crop production. It culminated in such an
inventory of the USSR wheat production in the 1977 growing season which
gave an estimate of the number of bushels of wheat produced by the USSR
within 6 per cent of the final Soviet figures 6 months before their release. MSS
data were used to identify and determine the area of acres in wheat. This was
followed by the use of available meteorology data from the area applied to crop
yield models to predict the number of bushels of wheat which would result.
Results were well within the USDA accuracy goals. The program was
separately budgeted with a budget of $10M/yr. at its peak. More than 200
support contractor personnel were involved, with the majority of the effort
conducted at the NASA Johnson Space Center. A more complete description of
LACIE and the results achieved is given in MacDonald and Hall, 1980.

There were a number of notable contributions of the program to the technology

of data analysis. One was a means to label training samples for the supervised
classifier in an unbiased fashion without observations from the ground, since
these were not possible in this program. Another was the Tasseled Cap
transformation, (Kauth and Thomas, 1976), a transformation providing a feature
space in which human perception of the interpretation of the contents of a given
pixel is enhanced. This program also stimulated the work to use temporal
information, e.g., in the for of a "crop calendar" for purposes of determining or
predicting the state of maturity of crops in the analysis process.

AgRISTARS was initiated in 1980 as a cooperative program between NASA,

USDA, USDC, USDI, and USAID. Its goals were to extend the crop commodity
forecasting capability which was begun with a wheat emphasis by LACIE to all
major grains. The program accounted for about 75 per cent of the federally
funded research in the remote sensing of renewable resources area in the
1980-83 period. The program was divided into 8 project areas covering crop
condition assessment; inventory technology; yield modeling; soil moisture
assessment; domestic crops and land cover; renewable resources inventory
(forestry); conservation and pollution; and supporting research. It, too,
generated a whole series of developments which moved the technology
forward. A special issue of the IEEE Transactions on Geoscience and Remote
Sensing, January, 1986 was devoted to the results of this program. Bauer
(1985) contains a review of the developments in crop identification and
condition assessment up to that time, including an extensive bibliography.

The parameters for Thematic Mapper, the second generation Landsat

multispectral sensor system were set beginning with a NASA working group
meeting held April 30-May 1 & 2, 1975 at Purdue University. At that meeting a
group of 35 invited participants selected the location of the spectral bands and

- 12 -
Landgrebe: Evolution of Landsat Data Analysis

other parameters for this instrument. Members of the working group felt strongly
that a seventh band should be included, however, the NASA managers had
limited considerations to six bands on projected cost grounds. Thus bands
number 1 to 6 were selected, and detailed planning for Thematic Mapper
began. Later, the NASA managers relented allowing the second middle infrared
band to be added as the band now labeled band 7. This is what led to the non-
sequential numbering of the Thematic Mapper bands in use to this day.

Technology Transfer. Prior to the launch of the first Landsat, because the
technology of multispectral sensing and analysis was new, but the community of
potential users of this technology would be large, there was a significant
technology transfer problem to be overcome. The technology had to be moved
from the small number of people who knew it to the large number who could
use it. What was needed went well beyond the normal vehicles used within the
research community, since most of the user community would not see such
documentation occurring in archival research journals and symposia
proceedings. Thus significant efforts began to be exerted to make available
user-oriented documentation, short courses, and other forms of communication
appropriate to this problem.

One particularly unique scheme resulting from the unusual characteristics of

remote sensing was the Remote Terminal System created and operated by
Purdue's Laboratory for Applications of Remote Sensing (LARS) based on
NASA funding. The LARS group, having been involved in the creation of the
technology, codified the digital computer algorithms emerging from their
research into a computer software system called LARSYS. Beginning in 1970,
this software system, implemented on a mainframe time-share computer, was
made available for use by others via remote terminals connected either by
leased lines or dial-up modem. Various training materials, including audio
tutorial tapes, videos, and pamphlets were created to show how to use the
system for analyzing multispectral data from a user's own laboratory. The cost to
the user was only the cost of supplying a teletype style terminal and the cost of
the communication line. Since the processing and the data were resident at
LARS, only processing instructions and results needed to be sent over the
communication lines, thus the bandwidth required between the host computer
and the remote terminals was modest.

This system was operated in this fashion from 1970 prior to the launch of
Landsat until 1982. Figure 6 shows some of the sites which had such
installations. The intention was for sites to have an opportunity to try out
multispectral analysis methods with very little expense. If it appeared to be
useful, then they could create their own capability either by building their own
hardware/software system or by receiving a copy of the LARSYS7 system,
which was available at no cost either from Purdue or from NASA Office of
Technology Transfer sources. A number of NASA, University, other government,
and private commercial organizations did so.

7 An application software system which is a descendent of LARSYS, but implemented for

personal computers is available for downloading from the world wide web at no cost at the
following URL: http://dynamo.ecn.purdue.edu/~biehl/MultiSpec/

- 13 -
Landgrebe: Evolution of Landsat Data Analysis

Figure 6. Location of sites of terminals on the Purdue/LARS remote terminal system.

On Analysis Methods and their Development. The fundamental premise

of remote sensing is that

• Information is transmitted through space via force fields and, in

particular, via
- spatial
- spectral
- temporal
variations of those force fields.

Then to capture remotely sensed information, one must,

• Measure those variations and relate them to classes of material of

interest.

The sensor system must measure the variations, and then the analysis system
must provide for relating the measurements to the classes of materials of
interest in any particular case and with acceptable accuracy. As has already
been pointed out, the decision was made early on relative to preparing for
Landsat to focus on spectral variations for pragmatic reasons, although spatial
and temporal variations have not been ignored over the years.

The matter of how the variations are represented mathematically and

conceptually is an important first step in defining how the analysis process
should proceed. There have been three principal ways in which multispectral
data is represented quantitatively and visualized.

• In image form, i. e., pixels displayed in geometric relationship to one

another,
• As spectra, i. e., variations within pixels as a function of wavelength,

- 14 -
Landgrebe: Evolution of Landsat Data Analysis

• In feature space, i. e., pixels displayed as points in an N-dimensional

space.

See Figure 7. We will refer to these three as image space, spectral space and
feature space, and next summarize some of the ramifications of these three
perspectives.

Image Space Spectral Space Feature Space

Figure 7. The forms for representing multispectral data.

Image Space. Though the image form is perhaps the first form a new researcher
or user thinks of when first considering remote sensing as a source of
information, its principal value has been somewhat ancillary to the central
question of deriving thematic information from the data. Data in image form
serve as the human/data interface in that image space helps the user to make
the connection between individual pixel areas and the surface cover class they
represent. It also provides for supporting area mensuration activities usually
associated with use of remote sensing techniques. Thus, it becomes very
important as to how accurately the true geometry of the scene is portrayed in the
data. However, it is the latter two of the three means for representing data that
have been the point of departure for most multispectral data analysis
techniques.

Spectral Space. We have already described why spectral variations early on

became the focus of the field. In the 1960's, the term "spectral signature" began
to be used, first to succinctly indicate to non-technical persons the then-new
idea that materials could be identified by their spectral variations alone.
However, some extended this idea to hypothesizing that each material has a
unique spectral response which is identifiably different in an absolute sense
from all others. This assumption set a course for analysis studies based upon
this assumption, even though, after 30 years of study, there is not general
agreement on the efficacy of this hypothesis. Still, many analysis algorithms
which appear in the literature begin with a representation of a response function
as a function of wavelength. Early in the work, the term "spectral matching" was
often used, implying that the approach was to compare an unknown spectrum
with a series of pre-labeled spectra to determine a match, and thereby to
identify the unknown. This line of thinking has, at various times, led to attempts
to construct a "signature bank," a dictionary of candidate spectra whose identity
had been pre-established.

- 15 -
Landgrebe: Evolution of Landsat Data Analysis

A second example of the use of spectral space is the "imaging spectrometer"

concept, whereby identifiable features within a spectral response function, such
as absorption bands due to resonances at the molecular level, can be used to
identify a material associated with a given spectrum. This approach, arising
from the concepts of chemical spectroscopy which has long been used in the
laboratory for molecular identification, is perhaps one of the most
fundamentally, cause/effect based approaches to multispectral analysis.

Feature Space. The third basis for data representation also begins with a
spectral focus, i.e., that energy or reflectance vs. wavelength contains the
desired information, but it is less related to pictures or graphs. It began by noting
that the function of the sensor system inherently samples the continuous
function of emitted and reflected energy vs. wavelength and converts it to a set
of measurements associated with a pixel which constitute a vector, i.e., a point
in an N-dimensional vector space. This conversion of the information from a
continuous function of wavelength to a discrete point in a vector space is not
only inherent in the operation of a multispectral sensor, it is very convenient if
the data are to be analyzed by a machine-implemented algorithm. It, too, is
quite fundamentally based, being one of the most basic concepts of signal
theory. Further, it is a convenient form if a more general form of feature
extraction is to precede the analysis step, itself.

Another key characteristic which is fundamental to the engineering task of

optimally designing a data analysis system is the basis for the mathematical
representation of the data. A number of approaches have been considered for
multispectral data over the years. The following are some examples.

• Deterministic Approaches
• Stochastic Models
• Fuzzy Set Theory
• Dempster-Shafer Theory of Evidence
• Robust Methods, Theory of Capacities, Interval Valued Probabilities
• Chaos Theory and Fractal Geometry
• AI Techniques, Neural Networks

Deterministic approaches have been by far the most common. This is no doubt
because they are the most intuitive. Spectral matching and imaging
spectroscopy are specific examples of these. Stochastic methods are also quite
common. They have the advantage of being powerful, rigorous, and well-
developed in the engineering literature, having many background tools
available.

As the Landsat data analysis technology has formed over the years, another
fundamental characteristic of analysis methods is that of identification vs.
discrimination. Is the intent of the algorithm to identify the contents of a pixel or
area, or is the algorithm to discriminate between a defined set of candidate
classes. The former is spoken of as absolute classification while the latter is
said to be a relative classification scheme. That of identification is clearly the
more conceptually appealing, especially to the non-technical user, while
schemes based on the latter are likely to produce the more accurate results in a
given situation at the expense of usually being the more difficult to implement.

- 16 -
Landgrebe: Evolution of Landsat Data Analysis

A final fundamental characteristic of Landsat data analysis methods that

distinguish between various approaches is the human data interface. How does
the human become involved in specifying what classes are of interest in a given
analysis task, and how is this interest related to the spectral response of the
various materials in the scene? As displayed in Figure 8, the data analysis
process in the remote sensing context is basically a merging of spectral
characteristics present in the current Landsat data set to be analyzed with that
of reference information about the scene.
Landsat
Data

Preprocessing
A
Reconcilliaton Data
of Conditions Analysis
Preprocessing
B
Reference
Information
Figure 8. Data analysis as a merging of Landsat data with reference information.

This Reference information can take a wide variety of forms. It can be in the form
of quantitatively expressed reference spectra gathered for the classes present
in the Landsat data, but at a previous time, and perhaps from laboratory or field
measurements. At the other extreme, the required reference information can
simply be input by the analyst, designating in the Landsat data set to be
analyzed, example areas of each class of interest.

As indicated in Figure 8, there must be a reconciliation of the scene, weather,

and illumination conditions at the time the Landsat data is collected with those
present in the reference data. This can be done by adjusting the Landsat data to
those of the reference data (Preprocessing A) or by adjusting the Reference
data to those of the Landsat data (Preprocessing B). As a third alternative, both
could be adjusted to a third set of conditions, for example, converting both sets
of data to geophysical units. Each approach has its advantages, however, any
such data adjustment is never perfect and to some extent could have an
unknown or unsuspected deleterious effect.

For example, "correcting" for the atmosphere, a part of Preprocessing A, might

seem to be an obviously good thing to do. However, it must usually be done
based on measurements of the atmospheric state which are inexact. Thus there
are circumstances in which it is not helpful and, indeed, could reduce overall
performance. On the other hand, if the reconciliation is achieved by the analyst
manually or otherwise designating example areas for each class of interest in
the data set to be analyzed, this has the advantage of automatically reconciling
the reference information with the current data set, thus preprocessing can be
held to a minimum or omitted altogether.

- 17 -
Landgrebe: Evolution of Landsat Data Analysis

The Future. It is appropriate to speculate upon the future for this technology,
but as a part of doing so, it is appropriate to first take a look at the past. At the
beginning of the space age in the early 1960's multispectral remote sensing
was seen to have great potential both for gathering land cover information over
broad areas and very detailed information over more limited areas. The
Michigan/ERIM airborne system together with multispectral discriminant
analysis methods were able to quickly demonstrate that promise as seen by the
1971 Corn Blight Watch Experiment. However, there have been several factors
present which have limited the achievement of the potential the technology
seemed to have at the outset.

• Spectral Limitations. The early ability to collect multispectral data from

space was a significant limiting factor. The four relatively broad,
relatively low signal-to-noise ratio bands of the MSS were a long way
from delivering the information content of the 12 to 18 bands of the
Michigan/ERIM scanner. Even the seven bands of Thematic Mapper
still present a substantial such limitation. It is questionable whether a
discrimination problem as subtle as measuring degrees of blight
infestation over the variability conditions of a broad area and an entire
growing season could be done successfully with Thematic Mapper.

• Data delivery delay. The delay in delivering data to the user has been
a chronic problem over the Landsat period. This delay has had two
deleterious effects. (a) For many applications of this technology, the
information contained in the data is quite perishable. Month-long or
even week-long delays in delivering satellite data to the user/analyst
eliminates the interest of a significant user community. (b) Such
delays affect the analyst more directly in the analysis process. They
have served to disconnect the analyst from the scene. Many types of
land cover change substantially between data collection and data
analysis during such a delay period, and this reduces the ability of the
analyst to make subtle perceptions about what shows in the data
relative to what was present in the scene. As a result, many analysts
don't even think of the scene, only the data as data. Add to these two
the uncertainty in being able to obtain data over a given site in a
given time interval due to the relationship between cloud conditions
and the infrequent passage of the satellite over a site of interest.

• Data Costing Practices. Given the other limitations, the decision to

begin commercializing the Landsat system during the Carter
administration was probably premature. The technology and its user
community were simply not that far along by that time. Aside from the
effect on the user community, the effect on the research community
was substantial. Most researchers could not afford to design and carry
out experiments in the normal fashion, not being able to afford the
data. Thus they fell into a mode of "making do" with whatever data
they could have access to.

• Policy Problems. U. S. Government policy over the period relative to

remote sensing has been uneven and inconsistent. This has inhibited
the development of both commercial providers of services and users
of the products that could be produced.

- 18 -
Landgrebe: Evolution of Landsat Data Analysis

From the perspective of these past limitations, however, there is reason to see a
bright future. Advances in the field of solid state technology in the intervening
years have made possible space sensors with substantially larger numbers of
spectral bands while at the same time providing substantially greater signal-to-
noise ratios without sacrificing spatial resolution. Government policy relative to
the commercial community and even with regard to data pricing has been
clarified in recent years, and while the cost of data for research purposes has
not improved as far as would be desirable, it has improved. It remains to be
seen what will be the case with regard to data delivery delays, but there is no
technological reason why it cannot be substantially better, even to the point of
delivering data in near real time.

There are many aspects that remain important research questions. Perhaps
chief among these is the matter of rigorous, broadly applicable, and user-
acceptable analysis procedures for the more complex data that future Landsats
will surely provide. It seems clear that sensor systems with large numbers of
spectral bands and greater signal-to-noise ratios will make theoretically
possible substantially more accurate and detailed information from such
sensors. The challenge, then, is to reduce this theory to practices and to do so
in a way that will be found acceptable to the Earth science and application
communities.

Acknowledgments. The author is indebted to the following people for their

comments on ideas put forth while this paper was being formulated.

Marvin Bauer, U. of Minn.

Marion Baumgardner, Purdue University
Roger Hoffer, Colorado State University
Quentin Holmes, Environmental Research Institute of Michigan
Christian Johannsen, Purdue University
Bill Malila, Environmental Research Institute of Michigan
Vincent Salomonson, NASA Goddard Space Flight Center
James A. Smith, NASA Goddard Space Flight Center
William Stoney, Mitretek, Inc.
Gene Thorley, U. S. Geological Survey
M. C. Trichel, Environmental Research Institute of Michigan

All of these people where involved in the development of the Landsat system
over an extended period of years. The perspective they shared with the author
aided materially in the writing of the manuscript. In the final analysis, however,
not all of their suggestions (nor the thoughts of the author) could be included,
due to space limitations, and the responsibility for the contents of the paper
rests solely with the author.

- 19 -
Landgrebe: Evolution of Landsat Data Analysis

References

Bauer, Marvin E., 1985, "Spectral Inputs to Crop Identification and Condition
Assessment," IEEE Proceedings, Vol. 73, No. 6, pp. 1071-1085.

Fu, K. S., D. A. Landgrebe, and T. L. Phillips, 1969, "Information Processing of

Remotely Sensed Agricultural Data," Proceedings of the IEEE, Vol. 57, No. 4,
pp. 639-653, April 1969.

Hasell, P.G. 1972. "Michigan Experimental Multispectral Scanner System" 4th

Annual Earth Resources Program Review, Vol. II, MSC-05937, NASA Manned
Spacecraft Center, Houston, TX, Jan 17-21, 1972, pp. 34-1 thru 34-13.

Hasell, Jr., Philip G., et al, 1974, "Michigan Experimental Multispectral Mapping
System, A Description of the M7 Airborne Sensor and Its Performance" NASA
Contractor Report, ERIM Report No. 190900-10-T, January 1974, NASA
Scientific and Technical Information Facility, P.O., Box 33, College Park, MD
20740

IEEE Transactions on Geoscience and Remote Sensing Special Issue on

AgRISTARS, Volume GE-24, No. 1, January 1986.

Kauth, R. J. and G. S. Thomas, 1976, "The Tasseled Cap - A Graphic

Description of the Spectral-Temporal Development of Agricultural Crops as
Seen by LANDSAT", Proceedings of the LARS 1976 Symposium on Machine
Processing of Remotely Sensed Data, Purdue University

Landgrebe, D. A., and Staff, 1967, "Automatic Identification and Classification of

Wheat by Remote Sensing," Agricultural Experiment Station Research Progress
Report 279, Purdue University, February 1967

MacDonald, R. B., M. E. Bauer, R. D. Allen, J. W. Clifton, J. D. Ericson, and D. A.

Landgrebe, 1972, "Results of the 1971 Corn Blight Watch Experiments,"
Proceedings of the Eighth International Symposium on Remote Sensing of
Environment, Vol. I. Environmental Research Institute of Michigan, Ann Arbor,
Michigan, pp. 157-190.

MacDonald, R. B., and F. G. Hall, 1980, “Global Crop Forecasting,” Science,

Vol. 208, pp. 670-679, 16 May 1980.

Proceedings of the IEEE Special Issue entitled "Perceiving the Earth's

Resources from Space," Vol. 73, No. 6., June 1985.

Swain, P. H. and S. M. Davis, Editors, 1978, Remote Sensing: The Quantitative

Approach, McGraw-Hill., Inc.. McGraw-Hill, 396 pages

- 20 -