i Protection
kvEPA
;r-c Development
apping Vegetation
Complexes with
Digitized Color
Infrared Film
Wisconsin Power
Plant Impact Study
EP 600/3
80-054
LIBRA.': .
U.S.
£Disoai?
,- r.T*"V
-
-------
RESEARCH REPORTING SERIES
Research reports of the Office of Research and Development, U S Environmental
Protection Agency, have been grouped into nine series These nine broad cate-
gories were established to facilitate further development and application of en-
vironmental technology Elimination of traditional grouping was consciously
planned to foster technology transfer and a maximum interface in related fields
The nine series are
1 Environmental Health Effects Research
2 Environmental Protection Technology
3 Ecological Research
4 Environmental Monitoring
5 Socioeconomic Environmental Studies
6 Scientific and Technical Assessment Reports (STAR)
7 Interagency Energy-Environment Research and Development
8 "Special" Reports
9 Miscellaneous Reports
This report has been assigned to the ECOLOGICAL RESEARCH series This series
describes research on the effects of pollution on humans, plant and animal spe-
cies, and materials Problems are assessed for their long- and short-term influ-
ences Investigations include formation, transport, and pathway studies to deter-
mine the fate of pollutants and their effects This work provides the technical basis
for setting standards to minimize undesirable changes in living organisms in the
aquatic, terrestrial, and atmospheric environments
This document is available to the public through the National Technical Informa-
tion Service, Springfield, Virginia 22161.
-------
EPA-600/3
June 1980
MAPPING VEGETATION COMPLEXES WITH
DIGITIZED COLOR INFRARED FILM
Wisconsin Power Plant Impact Study
by
Warren J, Buchanan
and Frank L, Scarpace
Institute for Environmental Studies
University of Wisconsin-Madison
Madison, Wisconsin 53706
Grant No. R803971
Project Officer
Gary E- Glass
Environmental Research Laboratory-Duluth
Duluth, Minnesota
This study was conducted in cooperation with
Wisconsin Power and Light Company,
Madison Gas and Electric Company,
Wisconsin Public Service Corporation,
Wisconsin Public Service Commission,
and Wisconsin Department of Natural Resources
ENVIRONMENTAL RESEARCH LABORATORY-DULUTH
OFFICE OF RESEARCH AND DEVELOPMENT
U,S, ENVIRONMENTAL PROTECTION AGENCY
DULUTH, MINNESOTA 55804
-------
DISCLAIMER
This report has been reviewed by the Environmental Research
Laboratory-Duluth, U.S. Environmental Protection Agency, and approved for
publication. Approval does not signify that the contents necessarily reflect
the views and policies of the U.S. Environmental Protection Agency, nor does
mention of trade names or commercial products constitute endorsement or
recommendation for use .
11
-------
FOREWORD
Tne U.S. Environmental Protection Agency was established as a focus of
scientific, governmental, and public efforts to improve the quality of the
environment. These efforts require expansion of our understanding of the
mechanisms that govern environmental changes and in particular tnose changes
that result from our own manipulations of the environment. One specific
thrust of these efforts must be the continuous development of more effective
and more efficient metnods for analyzing the environment and the changes
occurring in it.
One such project, which the Environmental Protection Agency is supporting
through its Environmental Research Laboratory in Dulutn, Minnesota, is the
study "The Impacts of Coal-Fired Power Plants on the Environment." The
Columbia Generating Station, a coal-fired power plant near Portage, Wis., has
been the focus of all field observations. This interdisciplinary study,
involving investigators and experiments from many academic departments at the
University of Wisconsin, is being carried out by the Environmental Monitoring
and Data Acquisition Group of the Institute for Environmental Studies at the
University of Wisconsin-Madison. Several utilities and state agencies are
cooperating in the study: Wisconsin Power and Light Company, Madison Gas and
Electric Company, Wisconsin Public Service Corporation, Wisconsin Public
Service Commission, and Wisconsin Department of Natural Resources.
Reports from tnis study will be published within the EPA Ecological
Research Series. Tney will deal with topics related to chemical constituents,
chemical transport mechanisms, biological effects, social and economic
effects, and integration and synthesis.
The Remote Sensing Group of tne Columbia Generating Station impact study
has sought to develop new techniques involving aerial photography of land
areastechniques that can be used to measure tne impact of such facilities as
power plants on the surrounding vegetation. This report describes research
undertaken to use computer technology in the analysis of aerial photographs of
land resources.
Donald I. Mount
Director
Environmental Research Laboratory-Duluth
Duluth, Minnesota
iii
-------
ABSTRACT
Environmental inventories of sites typically include some efforts to map
land cover. These maps are used to describe the existing environment or to
monitor and assess environmental changes. Digitized photo-analysis was
investigated in this context for mapping vegetation complexes, The goal was
to assess the utility of this technique for site analysis by testing several
variables.
Seventy-millimeter color infrared transparencies of upland and lowland
surface-resource complexes were digitized with a drum-type scanning
microdensitometer, The effects of seasonal change and resolution on digital
spectral signature behavior were investigated along with spectral signature
extendability, classification accuracy, and time and cost efficiency of
resource classification. Seasonal studies indicated that, for digital
analysis, at least some spectral signatures were confused in all seasons, but
these overlapping signatures shifted through the growing season such that a
combination of seasonal data would successfully separate nearly all resources.
As the resolution cell increased in size from 0,275 m^ to 14,59 m^, the
spectral range of resource signatures decreased in width and became more
discrete; however, linear resources such as roads were unresolved with the
largest resolution cell. Signatures appeared to be extendable from image to
image if images were from the same roll of film, were processed in the same
way, were exposed within a short time span, were corrected and calibrated
identically, and were scanned at the same time. The best classification
accuracy achieved was 85$ correct for two kinds of classifiers: one defined
spectral resource signatures with a box-like configuration, and the other with
an elliptical configuration, A basic parallelepiped (box-like)
classification, including ground verification surveys, of a scene containing
about 100.000 pixels cost approximately $100 and expended 20 h, A more
sophisticated elliptical table look-up classifier, with various data
transformations and enhancements, takes no additional time, but can increase
the cost by $50,
Some images were successfully classified, but other classification
attempts were abandoned; digitized film analysis was not an entirely reliable
mapping technique. Resource signatures often overlapped. Additional research
should determine if multiseasonal data overlayment, data-enhancement
transformations, and machine adjustments can create the necessary reliability
for operational applications.
This report was prepared with the cooperation of faculty and graduate
students in the Department of Civil and Environmental Engineering at the
University of Wisconsin-Madison,
IV
-------
Most of the funding for the research reported here was provided by the
U,S» Environmental Protection Agency, Funds also were granted by the
University of Wisconsin-Madison, Wisconsin Power and Light Company, Madison
Gas and Electric Company, the Wisconsin Public Service Corporation, and the
Wisconsin Public Service Commission, This report is submitted toward
fulfillment of Grant No, R803971 by the Environmental Monitoring and Data
Acquisition Group, Institute for Environmental Studies, University of
Wisconsin-Madison, under the partial sponsorship of the U,S, Environmental
Protection Agency, The report covers the period 1 July 1975 to 1 July 1977,
and work was completed as of 31 July 1977,
-------
CONTENTS
Disclaimer ii
Foreword iii
Abstract iv
Figures viii
Tables x
Acknowledgments xi
1. Introduction 1
2. Conclusions and Recommendations 2
3. Literature Review 5
4. Methods 7
Selection of imagery 7
Densitometry 7
Computer classification 12
Test variables 19
5. Results and Discussion 27
Ground resolution 27
Seasonal variation 33
Generality attributes 36
Cost-time efficiency 40
Accuracy 45
References 54
VI1
-------
FIGURES
Number Page
1a Columbia Generating Station study area near
Portage, Wis., April 1977 8
1b Hook Lake study area near Oregon, Wis., October 1976 8
1c Waubesa Marsh study area near Oregon, Wis., October 1976 8
1d Waunakee Marsh study area near Middleton, Wis., October 1976. ... 8
2 Location of study areas in the vicinity of Madison, Wis 9
3 Characteristic curve (density versus log-jQ exposure) of corrected
dye densities for blue, green and red dye layers of color
infrared film 11
4 Nine-level density slice of the green band from Hook
Lake digital data file using overprint option 13
5 Example of a histogram display of information in a training
set 15
6 Example of a bar diagram display of information from several
training sets 16
7 Example of a scatter diagram display of information from a
training set in two bands 17
8a Bimodal histograms showing two bands of information from a
two-resource training set 18
8b Same resource signatures resolved by using scatter diagram of
two bands 18
9 Illustration of two discrete elliptical resource signatures and
their confusion by a parallelepiped definition 20
10 Example of line printer output of classification from the
Columbia Generating Station study area, April 1977 21
11 Example of film classification output from the
Columbia Generating Station study area, April 1977 22
viii
-------
12 Hook Lake resource map as manually interpreted and traced
on a mylar overlay of 70-mm color infrared image 25
13 Randomly selected and marked pixels on digitally produced
film image of Hook Lake study area for accuracy check 26
14 Relation between mean signature width and ground
resolution cell size 29
15 Relation between standard deviation of resource signatures
and ground resolution cell size 30
16 Resolving power of various ground cell sizes for a portion of
the Columbia Generating Station study area showing
grassland and woods surrounding a sandblow, June 1976 32
17 Classifications of the Hook Lake study area,
October 1976 47
IX
-------
TABLES
Number Page
1 Ground Resolution Cell Sizes for Different
Combinations of Image Scale and Scanner Aperture 22
2 Signature Behavior of Grassland, Woods, Sandblow, and Duckweed
from the Columbia Generating Station Study Area, June 1976,
in Response to Different-Sized Ground Resolution Cells 28
3 Signature Behavior of Grass, Woods, Duckweed, and Sedge Meadow
from the Columbia Generating Station Study Area
in Response to Seasonal Variation 34
4 Signatures of Corn, Woods, Constructions, Alfalfa, and Sedge
Meadow in the Hook Lake and Waubesa Marsh Study Areas 37
5 Mean Signatures from Hook Lake-Waubesa Marsh Study Area and
Signatures from Waunakee Marsh Study Area Before and
After Standardization 39
6 Time and Expenses of a Basic Film Analysis When University
of Wisconsin Software Package and Madison Academic Computing
Center are Used 41
7 Additional Time and Expenses for Data Enhancements,
Transformations, and More Sophisticated Classifications
in Digitized Film Analysis 43
8 Confusion Matrix for the Unsmoothed Parallelepiped
Classification of the Hook Lake Study Area 48
9 Confusion Matrix for the Smoothed Parallelepiped
Classification of the Hook Lake Study Area 50
10 Confusion Matrix for the Unsmoothed Elliptical Table Look-Up
Classification of the Hook Lake Study Area 51
11 Confusion Matrix for the Smoothed Elliptical Table Look-Up
Classification of the Hook Lake Study Area 52
-------
ACKNOWLEDGMENTS
The authors wish to express their gratitude to Dr, Lawrence T, Fisher for
providing his knowledge and expertise in creating, maintaining, and executing
the digitized film analysis software package. We also acknowledge Dr. Grant
Cottam and Dr. Philip Page for critical review of the manuscript and editorial
assistance, Ms, Linda Quirk and Ms, Jane Lewis entered and managed the text
of this report in our computer system.
XI
-------
SECTION 1
INTRODUCTION
Since environmental impact assessment has become a requisite for many
large-scale works, such as power plant construction, the need for efficient
techniques to inventory natural resources of site environs has arisen. Ground
surveys are usually used to obtain data that describe the environment of some
proposed action. Ground-based resource surveys, including vegetation
quantification and description, are typically laborious, time consuming,
expensive, and often inconvenient for remote or hazardous study areas. The
ability to inventory resources such as vegetation from a distance (i.e., to
sense them remotely) avoids problems stemming from ground hazards or
inaccessibility. For some data requirements, remote sensing may provide an
improved means of data collection. Remote sensing with medium-scale (1:10,000
to 1:40,000) aerial photography provides rapid collection of immense amounts
of data, a unique overview, good geometry and resolution, and powerful
versatility. Although remote-sensing techniques have been applied to many
survey problems, surveys of vegetation are among the most common applications.
Qualitative manual photointerpretation is the typical remote-sensing
approach to surveying vegetation. However, quantitative data are often
desired. Digitized film analysis, the subject of this study, automatically
provides a quantitative assessment of ground resources. Digitized film
analysis involves quantifying the spectral signatures of ground resources
(vegetation types, water, roads, etc.) by measuring the dye densities of film
with a scanning microdensitometer. After calibration and correction of the
digitized film data, the resources can be characterized and classified based
on their spectral signatures.
The specific intention of this study was to determine the versatility,
accuracy, and cost efficiency in terms of time and money of digitized film
analysis for mapping vegetation complexes. In addition to studying the
general nature and capabilities of these classification procedures, we
explored the behavior of individual resource signatures in response to
seasonal change and changes in resolution.
Medium-scale (1:11,500 to 1:33,200) color infrared imagery was used for
this study. The images portrayed complex lowland and upland scenes including
various native and cultivated plant communities and structures (roads, dikes,
buildings, etc.). The complexity of intermingled resources was intended to
test severely the discriminatory power of digitized film analysis. The scenes
were considered representative of nearly any site that may require an
environmental inventory. The study areas ranged in size from several hundred
acres.
-1-
-------
SECTION 2
CONCLUSIONS AND RECOMMENDATIONS
Medium-soale color infrared images portraying complex upland and lowland
resource scenes ware digitized with a scanning microdensitometer. These
scenes ware considered representative of nearly any site requiring
environmental inventory, including cover mapping. Computer analysis employed
a software package developed by personnel of the Institute for Environmental
Studies, Environmental Monitoring and Data Acquisition Group, at the
University of Wisconsin-Madison (EMDAG). This study investigated the effects
of different-sized ground resolution cells (pixels) and seasonal variation on
resource spectral signatures and their separability for classification
purposes. Further studies determined the extendability of resource signatures
within a roll of film, the time and money expenses for a typical
classification effort, and the degree of accuracy to be expected from such
efforts.
p
Picture elements (pixels) with ground areas ranging from 0.275 m to 14.59
m" were used to study the same resource scene. As the ground area represented
by one pixel was increased, the spectral resource signature narrowed and the
resource signatures became more discrete. Also, the reduction of the amount
of data increased processing efficiency. One adverse effect of increasing the
size of the pixel was the loss of resolving power; at the largest pixel size,
linear features such as roads and dikes were difficult to discern. Also, the
proportion of "edge" pixels (border pixels whose values represent an average
of multiple resource signatures) increased compared-to pixels representing
single resources. The largest sized pixel (14.59 m ) seemed slightly too,
large because of resolution problems, and the next smaller pixel (3.565 m~)
resulted in suboptimal signature separation; a pixel size near 10 m would
have been best. The choice of pixel size depends on the detail desired.
Seasonal variation of digital signatures was considerable, but over a
growing season nearly all resources could be distinguished from one another.
Howaver, there was no optimal season for digital analysis. As might be
expected, infrared reflectance increased from early spring into the climax of
tne growing season. Although the predominant red tones in mid- and late
summer made manual interpretation more difficult, the degree of confusion
among digital signatures remained fairly constant. An important result of
this part of the study was the realization that at different times of the
growing season, resource signatures shifted such that nearly all resources
were distinct from one another in a temporal series. If multidate data files
were overlaid and registered, a powerful classification could be developed,
probably for little additional expense.
-------
An important factor in the comparison of digital processing and manual
interpretation was the extsndability of resource signatures. For images of
different scenes, at the same scale, from the same roll of film, expossd
within a short time span, processed the same way, and digitized with like
calibrations, cultivated resources (corn and alfalfa) had nearly identical
respective signatures, whereas signatures of natural vegetation associations
differed slightly more. With little modification the resource signatures
could be generalized from one image to the other. Another aspect of this
study was whether signatures could be standardized for data collected under
different conditions. Again, for cultivated crops the relationship seemed
stable when a simple ratio of signature means was used; however, this method
proved very inaccurate when applied to a natural vegetation complex such as a
woodland. The classification accuracy attained by applying these generalized
signatures is worthy of further study.
A basic digital classification (simple parallelepiped classification
without data enhancements) cost nearly $100 and consumed approximately 20 h
for a single 70-mm film image. A more sophisticated classification can cost
up to $150. These figures include all manual interpretive efforts, ground
truthing, and iterations of computer runs, but do not include salaries.
If digital processing is compared with strictly manual interpretation, the
comparative cost advantage depends on the desired product. For a single image
where nonquantitative classes are needed, manual interpretation is probably
superior. If resource signatures are generalized successfully for multiple
images, an economy of scale favors digital processing over manual
interpretation of multiple images. This economy depends on the areal coverage
of the film imagery, that is, the scale. Quantification of the areal extent
of a complex but distinct set of resources also favors digital analysis, even
for a single image. The reliability of a basic digital analysis, without data
enhancements, however, may influence the efficiency of a digitized approach
compared to a manual approach. Some of the scenes analyzed had enough
confused resource signatures to render those classification attempts futile.
Data enhancements might have corrected the situation, but time was
insufficient to experiment further.
Tne overall accuracy of the classifiers was lessened by a few complicated
resource signatures, for example, woodlands. Tne accuracy including the
problems incurred by woodland signatures was about 75? for the parallelepiped
classifier, but, discounting the woods, an accuracy of better than 90$ was
achieved. For an elliptical table look-up classifier, the accuracy was 12%',
nearly 80^ of the misclassified pixels were due to trees.
The accuracy of the classifications was also tested after erratic pixels
had been removed by using a smoothing routine. Tne result was that overall
accuracy was improved to nearly 85'^ for both the parallelepiped classifier and
the elliptical classifier.
Most confusion in all classifications was between upland and lowland
resources. Little trouble was sncountered distinguishing among upland
resources or among lowland resources by themselves. Most of the accuracy
problems were attributed to two facts: (1) the season was favorable for
-3-
-------
classifying some resources while unfavorable for classifying some others, and
(2) the scene was highly complex with intermingled signatures. A high level
of detail was attempted. Analysis of scenes with a less complex resource base
or use of a simple classification scheme, or both, would probably achieve
greater accuracy.
Further studies could be directed toward improving the efficiency,
reliability, and accuracy of digitized film analysis. The efficiency of this
technology would be enhanced by incorporating into the process a better visual
medium than paper printout, such as a color terminal. The extendability of
signatures should also be further investigated. Accuracy and reliability may
be improved by developing the ability to overlay and register multiseason
data. Also, assembling the best data files obtained from various
transformations and enhancements into one multiband data file might provide a
basis for improved classification.
-4-
-------
SECTION 3
LITERATURE REVIEW
Low-altitude medium-scale (1:10,000 to 1:40,000) imagery is an integral
part of -nany land-related studies, e.g., soil mapping (Bus'nnell 1932, Baldwin
et al. 1947, Kuhl 1970), land-cover or land-usa classification (Colwell 1963,
Nunnally and Witmer 1970), forest management (Spurr 1950, Avery 1955), geology
(Lyon 1969, Cole and Owen-Jones 1974), and geography (Simonett 1959). Tne
results of such studies may include quantitative tabular estimates of
land-cover classes or some map product. Interpretation of low-altitude air
photographs largely employs manual techniques with or without visual
enhancement. The basis of interpretation is commonly qualitative ocular
estimation (Anson 1968).
The development of film densitometry, coupled with the digital computer,
allows quantification, storage, and management of enormous film-dye density
data sets (Moore et al. 1964, Colwell 1965, Doverspike et al. 1955, Hoffer et
al. 1972). Suitable algorithms can be applied to the data for purposes of
correction, transformation, and enhancement (Smedes et al. 1971, Scarpace and
Voss 1975, Berry and Smith 1977, Carter and Gardner 1977).
The advantages of digital film analysis include objective comparisons,
signature extrapolations, quantification of resource signatures and areal
coverage, economy of scale, and the aforementioned ease of data management and
retrieval (Carter and Gardner 1977). Disadvantages are encountered with
overlapping resource signatures and resources having complex textural patterns
that are often readily identifiable to the eye (Whitman and Marcellus 1973),
but extremely complicated for computer analysis (Wiersma and Landgrebe 1975).
Registering picture elements (pixels) from the densitometer output to the
original imagery can be difficult (Bernstein 1975). Additional problems
arise, as with many mechanized processes, from machine (densitometer)
maintenance and calibration procedures. These latter problems are largely
surmountable and unrelated to theoretical concepts of densitometry. They
should not detract from the potential of digital film analysis, but at any
time they can hamper, halt, or invalidate an interpretive effort.
Furthermore, some effort must be expended to gain a working knowledge of the
computer hardware and software. The stats of the art is developmental and
dynamic. At present, the technology is not standardized and packaged such
that potential users (resource planners and managers) could simply and
immediately plug into a digital film-analysis system. The potential utility
and efficiency of this approach has yet to be fully explored.
Recent applications of digital film analysis largely relate to studies of
vegetation resources. These include general classifications of land use and
-5-
-------
terrain (Doverspike et al. 1965, Smedes et al. 1971, Owen-Jones 1977) and
discrimination of general plant communities (Driscoll et al. 197H) such as
forests (Akca 1971, LeShack 1971) and wetlands (Egan and Hair 1971, Scarpace
et al. 1975). Additional vegetation studies have been directed toward
agricultural cropland inventory (Williams and Aggarwal 1976, Kamat et al.
1977), plant productivity and biomass (Driscoll et al. 1972, Pearson and
Miller 1972), and crop infestation and disease detection (Ali and Aggarwal
1976). The use of digitized film for land-resource-management considerations
was explored by Reppert et al. (1969) in regard to wildland sites, and by
Benson et al. (1973) for soil limitations. Bryant and Zobrist (1976) and Cox
et al. (1975) discuss a digital image processing system for overlaying
geographic information for resource management.
-6-
-------
SECTION 4
METHODS
SELECTION OF IMAGERY
Testing quantitative film analysis required the selection of suitable
study sites with existing aerial imagery. To test fully the discriminatory
power of digital film analysis, emphasis was given to scenes with very diverse
and complex collections of resources. Multidate and multi-altitude imagery
taken with a Hasselblad 70-mm format camera and 40-mm focal length lens was
commissioned for the study of the Columbia Generating Station near Portage,
Wis. (Fisher et al. 1977). Scales ranged from 1:11,500 to 1:38,200. This
imagery included 70-mm transparencies of a scene containing a wide variety of
lowland and upland resources (Figure 1a). Present in the scene were upland
mixed oak woods, lowland aspen woods, dry prairie, structures (roads, dikes),
deep sedge meadow, shrub carr, cattail (Typha sp.) beds, and floating duckweed
beds.
Single-date imagery was also available for several similar scenes in the
Madison, Wis., vicinity (Figure 2). These scenes included Hook Lake (Figure
1b), an acid sphagnum bog with zones of tamarack (Larix laricina) and
leatherleaf (Chamaedaphne calvculata) over a sphagnum (Spagnum sp.) mat,
wiregrass sedge (Carex lasiocarpa). and cattail surrounded by woods and
agricultural fields; Waubesa Marsh (Figure 1c), a calcareous shallow wetland
with much shallow sedge meadow, cropland, upland mixed oak woods, and
constructed resources; and Waunakee Marsh (Figure 1d), an extensively
disturbed wetland dominated by cattail and some shallow sedge meadow, and
surrounded by cropland and upland mixed oak. Curtis (1959) thoroughly
describes the flora and phytosociology of the various plant associations
depicted on this imagery.
DENSITOMETRY
Transformation of multi-emulsion imagery into digital form can be
accomplished by using a scanning microdensitometer. The photographic scanning
for this study was performed with an Optronics P-1700 scanning
microdensitometer, a revolving drum-type device that measures film-dye density
with square pixels having linear dimensions of 25, 50, or 100y. Narrow-band
filters are placed in the densitometer's light path to extract data for
individual color layers of the multi-emulsion image. The output of the
scanner is a two-dimensional array of numbers ranging from 0 to 255
proportional to the dye density. These numbers are written as 8-bit words on
magnetic tape. For this study blue, green, and red filters having 100-& band
widths (centered at 4,500, 5,500, and 6,500 X, respectively) were used
-7-
-------
Figure la. Columbia Generating
Station study area near Portage.
Wis., April 1977.
Figure Ib. Hook Lake study area near
Oregon, Wis., October 1976.
Figure Ic. Waubesa Marsh study area Figure Id. Waunakee Marsh study area
near Oregon, Wis., October 1976. near Middleton, Wis., October 1976.
-------
Figure 2. Location of study areas in the vicinity of Madison, Wis.
-------
successively to create three raw data files for each image. By convention,
the data files for the blue, green, and red bands were labeled bands 4, 5, and
6, respectively. The mechanical design of the scanner and the nature of
multi-emulsion film resulted in precise registration of the three data files
to within 2 y.
Before the data can be enhanced or manipulated, several correction
calibrations should be performed (details of sensitometric corrections and
theory are discussed in Scarpace 1977). Because film density is not simply
related to light reflected from the ground, reliance on such an assumption may
cause problems (Scarpace 1977). Two main considerations affecting the use of
film density are: (1) the actual relationship between film exposure and the
resulting film density; and (2) the change in effective exposure on the film
plane due to lens geometry. These effects, along with atmospheric
interference, lens flare, and changes in ambient light, should ideally be
compensated for before launching an interpretive effort (Steiner 1972).
The first step in the correction process is to plot the relationship of
density versus exposure (ergs/cm^) to obtain what is called a characteristic
curve (Figure 3). This relationship which depends on film type, emulsion
number, and, most important, the film processing, is derived by exposing a
portion of film to a sequence of known fixed-energy steps. The developed film
gives the investigator an incremental scale of densities, each with a known
exposure. After this film wedge is digitized, the characteristic curve can be
generated. Film wedges were available for all but the September imagery used
in this study.
After the relationship between overall film density and exposure has been
discerned, corrections for the behavior and interaction of the three dye
layers and the film base for color and color infrared films must be applied.
Any density measurement reflects the combination of all film layers. All
layers contribute to the density because they have broad band sensitivities.
If the specific spectral dye densities are known, the actual amounts of dye in
each layer can be deduced after scanning the film through three narrow band
filters (Scarpace 1977). After correction calibrations, a narrow band
representation of the imagery can be derived similar to data from a
multispectral photographic effort. Ideally, after the correction process, the
data will have nearly the radiometric fidelity of an electro-optical scanner
(where reflected energy is electronically measured) and the geometric fidelity
of a photograph.
Additional corrections to compensate for problems such as darkened film
edges due to lens falloff were not operational at the time of this study.
Therefore, the study scenes were selected as much as possible from the center
of their respective images, away from the darkened edges.
Several algorithms were available to transform, enhance, and separate the
resource signatures in the data files. Three transformations were used most
commonly in this study. (1) The ratio of one band to another was completed.
The result was a band in which aberrant values caused by variations in scene
illumination were removed (Goodenough and Shlien 197H). (2) Canonical
discriminant analysis transformed data into new "pseudo-bands" oriented so
-10-
-------
K1 f 1
1 !
I I I I
Figure 3. Characteristic curve (density versus log Q exposure) of corrected
dye densities for blue, green and red dye layers of color infrared
film.
-11-
-------
that the separation of specified resource signatures was maximized. (3)
Texture characteristics were elucidated by computing the standard deviation of
each pixel and its neighboring pixels within a specified distance. Areas with
highly contrasting textural patterns produced higher standard deviations. An
additional file was created for this information. The texture algorithm was
the only transformation used for final classifications.
COMPUTER CLASSIFICATION
After an image is digitized and corrected, representative sets of pixels
are selected so that the surface resources in the scene can be classified by
computer. To achieve these classifications each pixel's spectral signature is
compared to the signatures of known sets of pixels, selected to represent as
nearly as possible all surface resources in the data (Wacker and Landgrebe
1972). These "training sets" are carefully selected to represent single
"pure" resource signatures. Complexes of resources in a single training set
will result in confused classifications. The selection of pure training sets
becomes the most critical part of these classification processes because
training-set signatures are the defining factors. A supervised computer
classification is only as good as the training sets (Wacker and Landgrebe
1972).
The encoding of training sets is a two-step process relying on knowledge
of the separation and identity of the resources in a given scene. In the
first step each of the three data files containing single-band pixel values
can be assigned a string of characters to represent increments of pixel
values. For example, if the blue-band data file contained pixel signature
values from 1 to 255 and the character string "1 2 3^5" was assigned to
represent all values in this file, a computer map could be generated in which
all pixels in the file having values 1 to 51 are printed as the character "1",
those from 52 to 103 are the character "2", and so on until all five
increments are assigned a character. The vernacular expression for such a
process is a density-level slice or contrast stretching; the file was sliced
into five levels in this hypothetical case. The values in a file can be
assigned any number of characters available from a line printer and,
therefore, can be sliced into whatever number of levels is desired. An
optional string of characters is an overprint string, which maximizes the
contrast among levels on the output map. The darkest characters can be
assigned to the highest or lowest signature values by preference.
Between 9 and 12 levels have been found useful in portraying the general
patterns of resources on a digitized scene (Figure 4). The recognition of
resource patterns on a level slice (such as water, corn fields, and trees in
Figure 4), allied with prior knowledge of resources in the original image
(usually attained with manual photointerpretation and ground truthing), allows
delineation of training sets. Although the digitized scene can be presented
by other methods, for example, by using color terminals or by selecting
training-set coordinates (as with ground surveying and photogrammetry), all
methods rely on the transfer of some knowledge of ground resources to the
digitized scene.
-12-
-------
.j;||-*;il«K(^i^|pr :("** ip;||p;|ip^p-'- if |t "^|||E~"
Figure 4. Nine-level density slice of the green band from Hook Lake
digital data file using overprint option.
-13-
-------
The second step is to recognize and encode the row and column coordinates
of the training sets. The available software allows odd-shaped training sets
to be encoded simply by recording all the vertices. This feature of
extracting general polygons can be an advantage over rectilinearly restricted
sets (Heiber 1976). Some resources, especially woods, contain a multitude of
signatures as one looks at crowns, edges, shaded depressions in the canopy,
etc. (Kan et al. 1975). All such signatures must be recognized if a
category "woods" is to be completely classified. The requisite training sets
are often small, containing perhaps only four pixels. Since the statistics of
a training set with only four members are unreliable, a program was developed
to allow merging of training sets representing like classes and like
signatures. Care was necessary to avoid mixing pure training sets with impure
ones. Much positive effort can be negated by one such misfortune. For the
complex scenes used in this study, 70 or 80 training sets were commonly
designated in an initial effort to represent all resources present.
Additional iterations needed to fill blank spots in a classification might
require several more training sets.
Once training sets are identified and encoded, a program is run that
extracts and files the signature information in each band for every training
set. The signatures can then be presented as computer output in various forms
so that judgments can be made as to their accuracy and adequacy. These forms
include histograms (Figure 5), bar diagrams (Figure 6), and scatter diagrams
(Figure 7). Each has inherent advantages and disadvantages.
The bar diagrams give the most rapid visual approximation of how resource
signatures relate to one another and the extent to which they overlap. In
such diagrams the mean of each signature is designated along with three
standard deviations, and the spread or normality often indicates the purity of
the signature. Since bar diagrams are not detailed enough to allow inference
of specific spectral boundaries, they are best applied as an indication of
data quality. If resources are seen to be badly confounded at this point,
either the classification process can be abandoned or appropriate
transformations can be applied to enhance the data.
Histograms are the usual mode of presenting resource signatures in
sufficient detail to permit assignment of spectral bounds. With histograms
the values in each band of every pixel in a training set can be viewed
collectively, the goodness of fit to a normal distribution can be judged if
enough pixels are present, and the number and degree of aberrant values are
commonly evident. If multiple resources are inadvertently included in a
training set, however, their respective signatures cannot be separated
conclusively.
In such cases, when multiple-resource sets are suspected, scatter
diagrams are best applied. Bimodal histograms for single bands are not very
informative, but when pixel values from two bands are plotted, the extra
dimension can successfully separate the resource signatures. The problem and
its solution by using "scatter diagrams" is illustrated in Figures 8a and 8b.
After the quality and character of the signatures are discovered, several
editing capabilities are available. In addition to being merged, training
-14-
-------
CLASS
! 3--«LF«LFA
b PO [ NTS
PFR SYMBOL ,
107C
108C
109C
I I 0*
Hir
1 |2C
i |3C
I 1 NC
I 15 +
1 1 61
I 1 7C
I 1 8C
1 I 9C
1 20*.
12IC».
I22C
123C-
12^C«
1 2S+«»*
I 26C «
127C»«
1 28T « « «
1 29C »» «
130 + ...... -
1 3IT ...... '
132C .......
I33C .......
135 +
1 36C
1 38C » «
1 39C
1 MO + »«
1 H 1 C
1 MfC
115 +
1 M6C
117C
1 H8C
1 49C
1 50 +
i sir
5 9 e POINTS
r N T I R E
H I
1 Q7C
1 3PC
1Q9C
1 in*
1 I 1C
1 1 2C
1 1 3C
1 1 1C
IIS*
1 1 6C
1 1 7C
1 1 BC
I I 9C
1 20 +
1 21C
1 22C
1 23C
1 2*(.
125 +
1 26C
1 27C
I 28C
1 29C
1 30 +
1 31C»«
13?C«
133C«»
1 31C
1 35**»»
1 36C
BA';D r.
I 38C
1 39C
1 MO +
1 43C *
1 1 1C *
n s * »
1 H6C
1 H7C
1 IRC *
1 M9C»
1 50 +
1S1C«
07C
08C
09C
1 0 + ««
1 1 C »
1 2C»»»*
1 3C«»»
I
-------
HANDS 456
Figure 6. Example of a bar diagram display of information from several
training sets.
-16-
-------
I/"1.
I
a-
fV (V fV (V
f\; rv ~
(V rv r*>
rv rv rv.
(V rv «-«
(V rv cr
(V IT-
rv ~* C£
rv < r-
rv.» x
rv --. IT-
fV - 1 ^j
rv.1 Ki
(V - f\i
rV -- "-1
rv >- c.
rv, o o*
rv,
rv: c rv
rv o .r-
rv c c
o-o-
- o- cr
- o- r-
>- cr >c
o~ tr
cr -?
^- i^- rr
- r- oc
r*- r^
" r^ ir
t^ C!
*- !^- K>
- r- rv.
- r« c.
*c o-
-i CC
i >c r~
-* -X3
« >T IT
- >C ^3
t «o r"
i >C rvj
-^ rv, rv
,- f-- r^
cc cc ~c a
-c cc c r- «
rv i/' n or tTi rv
ro cc f. cc K"
K* P«- IT> rv ^J
»- rv.1 cc/ c? ^j -n
«- IO 5J K1 '
« *
»-( * ^^
«-i.
w-t
ru ru in
rv rvi ^r
rvt rvj r^
rv. »v rv
rv rvi «
rv rv o
rv «-< o-
rv *~* cC'
rv r«-
rv >c
rvi *-< ir.
(V ^f
rv «- KI
rv/ «-i rvi
rv ^-* *
rv ^ c-
rv c; CT1
rv o cc
rv c t--
rv- cr -c
rv o tr
rv c «
rv o f
rv c rv
rv o «
rv e c
^H O" O-
«- o~ cr
- O" ^
c >c
»- O~ IT'
CT- cj
. CT" f*
*^ c? r\j
-* CT
CT C
»^« cC O1
«~ cr cc
^- cc r^
>~ X xC
oc in
CC, =3
co m
or rv
^ cc.
1C 0
r>- er
*-, t~~ a.
~ r- i^
^ r^ IT
*~ i-
*-^ f^ r**i
r- rv
r^ c
-co
« >o cc
^~ -C p^-
-c .r
«-* -C If
< «c :ar
« «C- ro
^ -D rvi
CT- cj C
en
T3
o
4-1
C
H
0)
CO
60
C
H
C
H
CC
S-t
4-1
e
o
C!
O
H
a
rl
D.
CO
H
0
CC
>_
OC
CO
CO
o
C/j
Q.
e
cc
X
cj
O
r^
QJ
oc
H
-17-
-------
BAND 4
HIh
HI 1 1 1
H 1
1 2 3 4 5 6 78 9 10 11 12 13 14 15 16 17 18 19 20 22 24
26
BAND 5
50 52
Figure 8a.
54 56 58 60 62 64 66 68
70
72 74
76
Bimodal histograms showing two bands of information from a two-
resource training set; individual resource signatures cannot be
discerned.
If.
11 -
8.
J .
0 6
z 6
*
9 v
1
> g
> '
' ^
|
1
1
I
i
>.
>'
\
>i
4
>
>
»'
»,
|
)
k
54 56 58 60 62
BAND 5
64 66 68 70
Figure 8b. Same resource signatures resolved by using scatter diagram of two
bands.
-18-
-------
sets can be deleted, identities changed, signatures trimmed to within a
specified number of standard deviations, or arbitrary upper or lower bounds
can be imposed on training-set signatures to remove outside pixels. As
mentioned previously, extreme caution is recommended since useful information
can be obscured or lost during these manipulations.
After signatures are satisfactorily quantified and displayed, a
classification can be applied. Several types of classification algorithms
exist. The simplest is a "parallelepiped" or "box" classification. Each
resource class is assigned upper and lower bounds in each band. Thus, in
three dimensions, a rectilinear set of boundaries defines each resource
signature. All pixel values are tested against the boundaries for all bands.
Those inside a given class parallelepiped are so classified. If a pixel does
not fit into any box, it remains unclassified. This algorithm is simple and
fast (Hyde et al. 1977). However, real signatures seldom occupy a box-like
configuration. Instead, the signatures typically assume an elliptical
configuration when plotted in two or three bands. Two signatures can actually
be distinct, but a box classifier will still confuse them (Figure 9).
More sophisticated classifiers rely on the statistics of the training
sets and data points. Computer programs are available that provide standard
statistics such as mean vectors, eigenvalues, eigenvectors, and covariance
matrices for each resource. An ellipse can be defined for each resource, and
an array of these definitions can be loaded into a tabular file based on these
statistics. Classification decisions can then be achieved by "looking up"
these elliptical definitions Ln the table and comparing pixel values to them.
This table look-up or elliptical classifier was used in this study alternately
with the box classification. Both assume a Gaussian distribution of signature
pixel values.
Various other classifiers exist that were not employed in this study.
Most make use of some minimum distance or maximum likelihood rule whereby
individual pixels are classified according to the statistics of resource
spectral signatures or vectors that associate most closely. These are the
"vector-by-vector" classifiers as described by Wacker and Landgrebe (1972).
The expense of these semi-automatic or fully automatic classifiers is greater
than for thoss used in this study because of the comparative calculations
required for svary pixel in the vsctor-by-vector approach (Hyde et al. 1977).
Classifications can be put out as color television displays, film images,
or line-printar maps. The assignment of characters or colors to each class is
arbitrary; the original scene can be altered or enhanced, or both. This study
employed both line-printer maps (Figure 10) and digitally reconstructed film
images (Figure 11) to present finished classifications.
TEST VARIABLES
This study identified five major variables requiring investigation. They
were selected on the basis of their anticipated importance in a
problem-solving situation involving resource inventory or classification.
-19-
-------
Q
Z
CO
Box
Defining
4> ^r-
^
/*
-4
t s/
s&*
££
Box Defining
^
BAND 5
Figure 9. Illustration of two discrete elliptical resource signatures and
their confusion by a parallelepiped definition.
-------
TITLE: PORTAGE 4/26/77 1:19100 2500AMT SCANNED AT 100 MICRONS 8/3/77
TAPE READ RESOLUTION: 1 PRINT RESOLUTION: 1
START ROW: 1 END ROW: 132 START COLUMN: 4 END COLUMN: 123
TITLE: PORTAGE 4/25/77 1:19100 ZSOOAMT SCANNED AT 100 MICRONS 8/3/7?
TAPE READ RESOLUTION: i PRINT RESOLUTION: i
CHARACTER RESOURCE
UAIER-SHACE
WAlEK-SHAOE
HATER-SHADE
IREE1
TREEZ
TREE 3
SHADE
GRASS
GRASS
SAND
TREE-SHRUE
TREE-SHRU3
SHRUB
BAND LOU UP SAND LOU UP BAND LOU
D
40
15
33
78
54
5
28
76
43
126
103
76
47
118 145
100 133
152 228
73 103
50 100
126 151
5
5
5
5
S
5
5
5
5
5
5
S
5
50
as
68
112
96
80
54
me
117
178
123
88
154
76
107
87
148
122
105
80
168
161
236
147
126
176
6
6
6
6
E
6
6
6
6
6
6
6
6
0
0
0
55
68
43
0
58
53
90
43
36
fa
UP
6
60
10
33
105
90
58
36
105
136
67
66
34
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
NUMBER OF
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
PIXELS
FOR UATCR-SHADE
FOR UAT£R-SHAOE
FOR WATER-SHADE
FOR TREE!
FOR TREL2
FOR TREE 3
FOR SHADE
FOR GRASS
FOR GRASS
FOR SAND
FOR TREE-SHRUB
FOR TREE-SHRUB
FOR SHRUB
UNCLASSIFIED
470
1255
84
4738
515
335
163
2879
1105
569
382
1610
455
681
Figure 10. Example of line printer output of classification from the
Columbia Generating Station study area, April 1977.
-21-
-------
Figure 11. Example of film classification output from the Columbia
Generating Station study area, April 1977. Green= grass-
land, orange = woods, yellow = sand and gravel, brown =
water and shade, olive = shrubs.
Ground Resolution
The different altitudes resulted in imagery scaled at 1:11,500, 1:19,100,
and 1:38,200. The scanning microdensitometer apertures were set at either 50
or 100 y. This combination of image scales and scanner apertures provided
essentially four different-sized ground resolution cells (Table 1).
TABLE 1. GROUND RESOLUTION CELL SIZES FOR DIFFERENT COMBINATIONS
OF IMAGE SCALE AND SCANNER APERTURE3
Aperture (y )
100
Scale
1:11,500 1:19,100 1:38,200
OO7R /C\ C^OR _
1.32/1.15 3.65/1.91 14.59/3.82
aCell size = pixel ground area (m^)/linear dimension of pixel (m).
-22-
-------
The optimum ground resolution cell was studied first so that it could be
used in subsequent portions of the study. Considerations in determining what
constituted an "optimum" pixel size were: (1) preserving ground detail and
accuracy, (2) reducing data to manageable proportion, (3) removing "noise"
from the data, and (4) maximizing time and cost efficiency.
These goals obviously conflict. When compromise was required the
decision as to how to optimize all constraints was bound to be subjective.
For example, the desire to remove some of the textural noise from woods
sometimes required a scale that would not resolve linear features such as
roads. Is resolving roads more or less important than facilitating woods
classification? This sort of question occurs constantly and can only be
answered by individual investigators in the context of their needs. This
study was directed toward ascertaining what pixel size resulted in roads being
unresolved or woods attaining less noisy signatures.
Seasonal Variation
The study of seasonal variation used the same scene as the ground
resolution study. Aerial photographs at a scale of 1:19,100 spanned the 1976
growing season from late May to late September on a monthly basis. An early
spring image taken in late April 1977 was also analyzed. Relative resource
spectral signatures were emphasized. Of special interest were how individual
natural resource signatures change over the growing season, which resources
were confounded, and what season, if any, was optimum for separating,
classifying, and quantifying resources. Since vegetation resources are the
most sensitive to seasonal variation, the rich variety of vegetation types in
the lowland and upland areas was carefully scrutinized.
Generality Attributes
An important determination, which heavily influences cost and time
efficiency, is whether the resource signatures on a given date can be
generalized from one image to another, within the same roll of film, or among
rolls. This factor is crucial because it determines whether the effort of
identifying resource signatures must be repeated for every image analyzed-. If
signatures can be generalized, only new resources on additional scenes must be
defined; previously identified resources can be classified by applying known
signatures.
The signatures of selected resources were compared among film frames
taken from a roll of film depicting the same vicinity on a single date. This
test involved October imagery depicting Hook Lake, Waubesa Marsh, and Waunakee
Marsh (Figure 1), The signatures of resources common to the three sites were
compared. These frames were taken from the same roll of film, which had been
exposed during a single flight.
This portion of the study was not limited to discerning whether like
resources have like signatures. The possibility that signatures vary with
some direct relationship was studied with the prospect that the relationship
could be used to standardize signatures within a single-date roll of film.
-23-
-------
Cost-Time Efficiency
Costs, in terms of time and computer-account money, were documented in a
daily log for the entire classification process. The goal was to determine
representative costs of processing an image from the point of digitizing to
final interpretation. Many repetitions were recorded for each step in the
process, including data enhancement and presentation. For excessively
repetitive processes, most but not all iterations were documented.
On a more limited basis, the costs of ground-level interpretation and
manual air-photo interpretation were explored. These methods, the two most
common approaches to resource inventory, provided an important basis for
comparing computer-interpretation expenses.
Accuracy
In addition to providing economic bases of comparison, the ground
observations and manual photointerpretations act as evaluation techniques
(Curtis 1977, Townshend 1977)-. Each resource scene was visited on the ground
and manually interpreted for resource verification.
For more detailed comparisons, a mylar overlay was made of some magnified
imagery, and resource patterns were traced thereon (Figure 12). These mylar
maps were brought into the field where the delineated resources were
identified and verified. Registration of the delineated resources to actual
ground entities was accomplished largely by relying on landmarks. In most
cases the entire area was quickly perused and attention was given to
representative resource occurrences; once a resource's appearance was
satisfactory, the identity was extrapolated to likenesses in the scene.
The manual techniques were assumed to result in nearly perfect knowledee
of resource occurrence and identity. This knowledge was used to draw
conclusions as to the accuracy of the digital analysis. A parameter of
accuracy was calculated by selecting, with a random numbers generation, 200
pixel coordinates per scene (approximatly 0.2% of the total) and by
determining the percentage of accurately identified pixels (Figure 13). The
known proportions of resource classes were used as expected proportions for
the 200 sample pixels. A chi-squared test was then applied to the sample
proportions to see if the sample pixels were a fair representation of the
original population. The results agreed with minimum sampling adequacy
requirements for major resources put forth by van Genderen and Lock (1977).
This overall percentage does not necessarily reflect the relative ease or
difficulty of separating various individual resources. Some resources, such
as pavement, nearly always classify easily. Some commentary is directed
toward chronically confused resources, and a confusion matrix is presented for
each classification.
-24-
-------
Figure 12. Hook Lake resource map as manually interpreted and traced on a
mylar overlay of 70-mm color infrared image (original scale
1:38,200 magnified approximately 6X).
-25-
-------
Figure 13. Randomly selected and marked pixels on digitally produced film
image of Hook Lake study area for accuracy check.
-26-
-------
SECTION 5
RESULTS AND DISCUSSION
GROUND RESOLUTION
This section explains the behavior of resource signatures in response to
the application of four?different-sized ground resolution cells; sizes ranged
from 0.25 ra to 14.59 m . Images of the Columbia Generating Station site from
mid-June 1975 were used. These images were photographed on a single flight
with one roll of film. The altitudes above mean terrain level were 1,500,
2,500, and 5,000 ft. The signature behaviors of a few representative
resources are presented in Table 2.
The spectral signatures for each pixel size are not perfectly
symmetrical. Most of the variance was caused by a combination of the
correction algorithm, training-set placement, densitometer idiosyncracies,
phenology, and textural characteristics. Of these variables, training-set
placement is the only one that the investigator can totally control.
Therefore, unless the investigator is capable of redesigning the software and
hardware, some variability in the data might be expected. Direct comparisons
of signature data must be tempered with this consideration.
Although the mean vectors of the training-set signatures do not align
perfectly, the signatures tend to narrow with increasing pixel size (Table 2).
Tne average signature widths,,for each pixel size are. 52.6 for 0.275 m~, 55.4
for 1.32 m", 49.3 for 3.65 nT, and 35.8 for 14.59 nT(Figure 14).
Considering individual resources, the average reduction in signature
widths was 63.4^ for grass, 50.8£ for woods, 44.556 for sand, and 20.5% for
duckweed. The differential signature improvement might be expected because of
the relative extent of the resources. Woods and grasslands covered extensive
areas, whereas sandblows and duckweed colonies were relatively small.
Reducing the ground resolution power served to average and smooth
internal noise in the woods and grassland signatures. The expected
improvement in the internal cell signatures of the smaller resources was
probably balanced somewhat by the pollution of edge pixels (pixels overlying
two resources and having a signature derived from an average of the
reflectance of the two resources). For small resources the edge pixels are
proportionally more influential; avoiding edge pixels while selecting training
sets of small resources is difficult. The overall improved stability of
resource signatures, however, is illustrated by the plot of signature standard
deviations versus pixel size (Figure 15).
-27-
-------
TABLE 2. SIGNATURE BEHAVIOR OF GRASSLAND, WOODS, SANDBLOW, AND DUCKWEED
FROM THE COLUMBIA GENERATING STATION STUDY AREA, JUNE 1976, IN RESPONSE
TO DIFFERENT-SIZED GROUND RESOLUTION CELLS
Resource
Grass
Trees
(a combin-
ation of
three dif-
ferent tree
signatures)
Sand
Duckweed
Band
4
5
6
4
5
6
4
5
6
4
5
6
Pixel
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1.32
3.65
14.59
0.275
1 .32
3.65
14.59
0.275
1.32
3.65
14.59
0 50 100 150 200 255
51
10
116 1.
118 1
1
150
15^
150
130
113 __.
118
1
1
153
15
154
150
135
136
2
44
58 19
180
152
151
148
154
186
168 18
-]
- 170
170
96
I2
219
236
222- 2
JO
6
92
197
94
255
255
255
255
47
-------
Q
111
QC
<
z
o
<
LU
10 11 12 13 14 15
PIXEL SIZE (m2)
Figure 14. Relation between mean signature width (from Table 2) and ground
resolution cell size. Data from three images of the Columbia
Generating Station study area in June 1976 at scales of 1:11,500,
1:19,200, and 1:38,200, and scanning apertures of 50 and 100 y.
-29-
-------
LU
CC
LL
O
LJJ
Q
Q
DC
10 11 i2 13 14 15
PIXEL SIZE (m2)
Figure 15. Relation between standard deviation of resource signatures (from
Table 2) and ground resolution cell size. Data from three images
of Columbia Generating Station study area in June 1976 at scales
of 1:11,500, 1:19,200, and 1:38,200 and scanning aperture of 50
and 100 vi.
-30-
-------
Reducing signature ranges is desirable in the attempt to achieve discrete
resource classes. Since selection of pixel size is related to the discretion
of signatures and to the amount of data available, this selection is an
important consideration for digital analysis.
In addition to the need to achieve discrete resource classes, the other
main criterion influencing a decision on pixel size is the need to resolve
resources. Figure 16 illustrates the resolution success of a portion of the
Columbia Generating Station study area with four pixel sizes. Grassland and
woods border a sandblow in this scene. The configurations of the sandblow and
individual tree crowns are very obvious when a 0.275-m pixel is used. When the
largest pixel is used, however, the edges of the sandblow become rough and
angular; the shape is only approximately correct. The texture of the woods
becomes increasingly smooth with larger pixel size. Although roads are not
present in this scene, a resolving power sufficient for classification of
roads less than J.5 m (25 ft) wide is not attainable as the pixel size of
14.59 m is approached. This is true for any resource having a similar
minimum dimension.
As the pixel size increases, the number of pixels needed to represent a
resource on the ground decreases. Therefore, the number of edge pixels
increases in proportion to the pixels that purely represent a single resource.
These edge pixels detract from a classification effort because they have
irrelevant and erratic signatures. Coping with more edge pixels should be
considered when deciding to increase pixel size.
An additional influence on choosing an optimum pixel size is the
identification of the resources that are the main targets of the
classification. Signatures of larger grossly textured resources are improved
by increasing pixel size (Kan et al. 1975), whereas small resources may be
unresolved. If these smaller resources are deemed critical, the smallest
dimension of these target resources must be the standard for sizing the pixel.
?
For the purpose of this study, the two smallest pix^i sizes, 0.275 m" and
1.32 m", were easily rejected because of unwieldy amounts of data and
signature widths. This elimination left a choice of either 3-65-m or
14.59-ia" pixels. Unfortunately, nothing between these choices was available.
Woods were best classified at_the Columbia Generating Station study area by
using a pixel size of 14.59 m , but the resolution of all resources generally
was much more satisfactory with the 3-65-m pixel.
We decided that resolution of as many resources as possible was the better
choice, even though signature problems would arise. Thus, the loss of some
smaller resources was avoided at the expense of better classification of
grosser features. Therefore, the 3.65-m" pixel was selected for most of this,
study. The^optimum ground resolution cell lies somewhere between the 3.65-m"
and 14.59-31" choices available for this study.
-31-
-------
a. Cell size = 0.275m
b. Cell size = 1.32m
c. Cell size = 3.65m
d. Cell size = 14.59m
Figure 16. Resolving power of various ground cell sizes for a portion of the
Columbia Generating Station study area showing grassland and woods
surrounding a sandblow, June 1976.
-32-
-------
SEASONAL VARIATION
The seasonal development and senescence of vegetation are powerful
influences to spectral reflectance (Olsen and Good 1961, Kiefer 1970, Knipling
1970, Thomson 1972). As the growing season progresses many spectral signature
shifts occur. At any time some resource signatures may be confounded and some
may be discrete. At an earlier or later time the resources that were
confounded may be easily separated, whereas the formerly disjoint signatures
may overlap. Our results support these findings and suggest that if data
files from different dates could be registered, a compilation of unique
signatures could be extracted from otherwise imperfect data sets. No single
date allowed perfect separation of resources.
Seasonal change is obvious in images of the Columbia Generating Station
study area from April, May, June, July, and September. Most striking is the
extensive increase in infrared reflectance illustrated by the dominance of red
dye in the July image. Another seasonal change relates to water levels. The
spring imagery shows scattered open water areas that in later images either
disappear or support floating macrophyte beds.
The effects of seasonal change are probably best illustrated by the
change in individual resource signatures. Table 3 presents the signatures of
woods, grassland, duckweed, and sedge meadow as digitized on images from the
five dates. The quantitative relationships of seasonal signatures are
unreliable, since the original exposure of the various images was not equal
among the dates. The previously described algorithms for correcting and
calibrating data attempt to reconstruct the original exposure values; among
images taken on different dates, however, these original exposure conditions
are bound to vary.
Despite the lack of quantitative reliability, several qualitative trends
are evident. Infrared reflectance, as measured in band 6, increased
considerably after April. Duckweed was the only resource that did not
generally stabilize; it seemed to increase and grow brighter as the season
progressed. Green reflectance, as quantified in band 4, was low in April and
July. In April the vegetation was generally not green enough to predominate
over the bare branches and dead vegetation remaining after winter. Green
reflectance was low in July because in 1976 a drought was affecting plant
growth. The drought persisted through September 1976, but the green
reflectance in the September data was generally of greater magnitude than in
the July data. Several things could account for this: (1) September data
were not calibrated for lack of a film wedge; (2) July film may have been
underexposed; or (3) some rain in September may have been sufficient to green
the vegetation.
The most significant factor in these seasonal trends is the extent to
which they facilitate the distinction of wetland from upland resources. This
capability relates to national interests in wetland mapping and inventory
(Anderson 1972). In April, woods and grassland could be successfully
separated, as shown in the classifications of upland resources in Figure 11.
Lowland resources, duckweed and sedge meadow, had distinct signatures. Woods
could be discriminated from duckweed, and grass and sedge meadow were readily
-33-
-------
TABLE 3. SIGNATURE BEHAVIOR OF GRASS, WOODS, DUCKWEED, AND SEDGE MEADOW
FROM THE COLUMBIA GENERATING STATION STUDY AREA
DURING THE GROWING SEASON
APRIL 1977, MAY, JUNE, JULY, AND SEPTEMBER 1976
Resource
Grass
Woods
Duckweed
Sedge
Meadow
Band
5
6
4
5
5
6
4
5
Season
July
Sept.
April
June
July
Sept.
May
July
Sept.
April
July
June
July
Sept.
July
. I
July
Sept.
April
May
June
July
April
June
July
April
May
July
April
May
June
July
Sept.
April
July
Sept.
c
5
C
55
54
50
42
36
) 1C
78 »
10.
82 <
75 »
» 88
78^-
100
» 75
10
98
80+
#66
85
0 15
107
»-
132
150
134 4
125 <
108
»
13 »
-»
16
132 <
15
148
135
122 0
102
0 13
+ 1
126 »
108
114 » '
^ 120
0 2C
4 166
11
160
» 1«
168
150
151
162
172 »
4 170
164
0
10
164
- 154
0 25
197
214
0
-34-
-------
separated. Woods and sedge meadow had confounded signatures, however, as did
grassland and duckweed.
The May signatures showed some improvement in the separability of the two
upland resources from the wetland resources that were formerly highly
confounded. Unfortunately, the distinction between grassland and sedge meadow
was less complete in May than April. In June, the grassland signature was
almost totally distinguishable from duckweed, but was somewhat confounded with
sedge meadow. The woods were still incompletely distinct from sedge meadow,
and duckweed was still disjoint from woods.
By July both wetland resources were entirely distinct from grassland. Of
the four resources, the only problem was that woods and sedge meadow were
almost entirely overlapped for all three bands. A possible solution to this
problem would be to create a fourth file of textural data. The woods should
have greater texture values than sedge meadow. If successful, this added
textural information could make July data the best for classifying the four
upland and lowland resources. Because woods and sedge meadow are major
resources, their overlap would be a serious problem if textures are
inseparable. Possibly the drought effects enhanced the separation of wetland
and upland herbaceous communities more than would usually be the case for
July.
September provided the poorest data in terms of differentiating
resources. Only duckweed could be discriminated from the other resources;
woods, grassland, and sedge meadow were all confounded. These difficulties
could be due to the uncorrected condition of the data.
On the basis of these results, no optimum time seems to exist for digital
analysis of vegetation. Various resources are best separated at various
dates. This conclusion seems to contradict findings of Olson (1954), Lukens
(1953), and Seher and Tuellar (1973), which indicate that late summer and
early fall imagery are best for detailed vegetation mapping. Carter and
Stewart (1975) suggest that multissason imagery is the best approach for
wetland classification. Aforementioned differences between computer
discrimination of digital information and human qualitative judgments explain
the situation. The rcore simple computer classifiers used in this study do not
rely heavily on the importance of patterns and spatial association, whereas
manual interpretations take advantage of these attributes. In contrast,
digital information affords a more detailed division of tonal change than can
be achieved manually. Therefore, comparisons of optimum time between digital
analysis and manual analysis may be unreasonable.
Reliance on one year's data is inconclusive, especially with all the
inherent uncontrolled variables such as climate, exposure, and registration of
training sets. If the most distinct data bands from available dates could be
registered and combined in one data file, however, a relatively powerful
classification could ensue (Steiner 1972, Hauska and Swain 1975, Svedlow et
al. 1975).
-35-
-------
GENERALITY ATTRIBUTES
The study of generality attributes involved the study of signature
stability and extendability. As mentioned previously, the time and expense of
large-scale classification are closely related to the generality of resource
signatures. The discussion on efficiency following this section will detail
the costs of various steps in the classification process, but we note here
that selecting training sets and identifying resource signatures are among the
most time-consuming and critical steps. If a resource signature, derived from
a training set in one scene, can be generalized for all occurrences of that
resource in the one scene, further selection of training sets of that resource
is unnecessary. Of greater import is determining whether the resource
signature will hold for occurrences of that resource in other images given
some reasonable controls. Use of controls would require that the images be
taken on the same roll of film within a short time span under similar lighting
conditions.
Imagery dated 3 October 1975 was available for two scenes: Hook Lake and
Waubesa Marsh (Figure 1). These scenes portray a mixture of lowland and
upland resources with several resources in common. Because they were
photographed on the same day with one roll of film in a short time span, an
ideal situation was created for testing the extendability of resource
signatures. Tne signatures of representative resources from each scene are
shown in Table 4.
Although some differences could be expected because of such factors as
varying moisture regimes and vegetation density, the signatures for each of
the five resources were surprisingly alike. Especially similar were the
signatures for the two agricultural resources, corn and alfalfa, probably
because they are monotypic, controlled resources with smooth textures.
The variation between the signatures for woods at Hook Lake and Waubesa
could be caused by the amount of shadow included in the training sets. Purely
shaded spots were classified separately from woods; however, some
intermediately shaded pixels were included while delineating the training sets
for woods, and the control over this delineation was arbitrary. As previously
mentioned, the crown tops in the woods were irregularly colored. Thus, the
woods signatures differed by as much as 10$ in one band. Another source of
variation was that the woods signatures, as presented, are a compilation of
several training sets for each scene. A conscientious effort was made to
capture the full spectral range of each resource signature, but the presence
of unclassified pixels indicates that the signatures were never discerned with
perfect accuracy.
The signature of constructions (built-up) for Hook Lake is based on very
few pixels because very few constructed resources occurred in the scene.
Since these constructions were largely narrow roads and small residences,
selecting training sets was difficult. Edge effects may account for the
slightly lower signature values of the Hook Lake constructions.
Sedge meadow was a highly variable resource depending on moisture
conditions and species composition. At Waub?sa Marsh the sedge meadows were
-36-
-------
Table 4. SIGNATURES OF CORN, WOODS, CONSTRUCTIONS, ALFALFA, AND SEDGE MEADOW
IN THE HOOK LAKE AND WAUBESA MARSH STUDY AREAS
OCTOBER 1976
Resource
Corn
Wnnds
Construct ions
Alfalfa
Sedge
Meadow
Band
5
6
4
5
6
4
5
6
4
5
6
4
5
6
Scene
Waubesa
Hook Lake
Waubesa
Hook Lake
Waubesa
Waub es a-
Waubesa
Hook Lake
Waubesa
Hook Lake
Hook Lake
Hook Lake
Waubesa
Hook Lake
Waubesa
Hook Lake
Waubesa
Hook Lake
Waubesa
Hook Lake
Waubesa
3 5
5C
52
1
0 1
r
11
f-f
-|-|
85
95
78 -
3 .
1
95
85
76
30 1i
1 -^-^
1
ii 131
12
108
102
13 13(
12S
106
111
>0 2(
179 «
171
143
144
I42
, -ICO
9
)0 2!
197
183
217
210
>0
l-l
"J
-37-
-------
codominated by tussock sedge (Carex strieta), bluejoint grass (Calamagrostis
canadensis). and sa/eral species of forbs. Tne tussock sedge and forbs have a
lower reflectance than bluejoint. At Hook Lake forbs were absent in the sedge
meadows. Therefore, although the signature at Hook Lake was largely within
the spectral range recorded for Waubesa sedge meadows, it was concentrated at
the higher end.
Overall, this aspect of the study indicated that resource signatures
might be generalized among images when: (1) the images are from the same roll
of film, (2) the images are taken on the same day, (3) the images are
digitized at the same time witn uniform calibrations, (U) training sets are
for nearly identical resources in terms of composition and physiognomy, and
(5) scale is standard. This ability to generalize could be tested by
classifying one scene with the training-set statistics of a different scene.
Although the present study did not go beyond comparing signatures, further
study along these lines may be worthwhile.
The situation often arises in which one or more of the five aforementioned
conditions are not met and yet the generalization of resource signatures is
desired. To test this situation corn and alfalfa data from Hook Lake and
Waubesa Marsh, which had very stable signatures, were compared with corn and
alfalfa data from Waunakee Marsh (Table 5). Tne conditions for generalizing
resource signatures were met except that the data were digitized with
different machine calibrations and were corrected with a slightly different
algorithm.
To test if all resource signatures can be standardized on the basis of
the relationship of signatures for one or two resources, the signature
relationships of corn and alfalfa were calculated for each band. This
calculation was dons by using a simple ratio of the means; for bands 4, 5, and
5 the ratios between the Waunakse signatures and Waubesa-Hook Lake signatures
were 3.75, 0.73, and 0.83, respectively, for corn, and 0.73, 0.74, and 0.83,
respectively, for alfalfa. The similarity of these simple ratios suggests
that standardization should be valid for these resources. In fact, when corn
and alfalfa signatures were corrected by using the corn-signature
relationships for each band, the results were as good as if tue data were
scanned and calibrated in the same way (Taole 5). The variable condition was
successfully standardized.
This approach was not as successful with naturally occurring vegetation
communities, probably because the signatures of native plant communities are
not as stable as cultivated crops. Because of this possible discrepancy, the
signature for woods at Waunakee Marsh was adjusted using the relationship of
corn signatures between Waunakee Marsh and Hook Lake-Waubesa Marsh. The
results are snown in the last entries of Table 5. The unadjusted signatures
for woods at Waunakee Marsh are actually closer to those for Waubesa-Hook Lake
than the "corrected" ones in bands 4 and 5.
The problem with this attempt to generalize resource signatures rests most
likely on the difficulty of accurately portraying natural vegetation
signatures. The concept of standardization is probably valid for perfectly
valid resource signatures. Smooth cultured resources such as crops and
-38-
-------
TABLE 5. MEAN SIGNATURES FROM HOOK LAKE-WAUBESA MARSH STUDY AREAS
AND FROM WAUNAKEE MARSH STUDY AREA BEFORE AND AFTER
STANDARDIZATION BY USING A SIMPLE RATIO OF MEAN CORN SIGNATURES,
OCTOBER 1976
Resource
Corn
Alfalfa
Woods
Band
4
5
6
4
5
6
4
5
6
Scene
Waunakee
Waunakee
(corrected)
Waubesa-HooK L.
Waunakee
Waunakee
(corrected
Waubesa-Hook L .
Waunakee
Waunakee
(corrected)
Waunakee
(corrected)
Waunakee
(corrected)
Waubesa-Hook L.
Waunakee
(corrected)
(corrected)
Waunakee
(corrected)
Waunakee
(corrected)
(
) 5
5
52
0 1(
106
11
90 -
10
3 _-__
85
)0 1E
140
1 1
- - 122
a
102
131
0 2(
41
171
166
179 _
44
146
, 152
)0 2!
197
217
212
)0
-39-
-------
constructions lend themselves to signature identification, and therefore to
standardization, but natural vegetation resources do not. Therefore, the
overall success of such standardization efforts may be limited.
COST-TIKE EFFICIENCY
Perhaps the first question asked by anyone wishing to use digital film
analysis is what it costs. Even if the capabilities of digital analysis are
suited to a particular inventory or mapping problem, the cost of producing a
finished product must be compared to alternative techniques such as manual
photointerpretation or ground surveys (Aldred 1972, Grinnell and Conway 1972).
The following discussion attempts to present the costs of digital analysis on
the EMDAG system. It is important to remember that these costs are for
digitally and manually interpreting a single 70-mm image comprising about 1
square mile and 100,000-pixel data base.
All computer work in this study was performed on a UNIVAC 1110 computer
operated by the Madison Academic Computing Center (MACC) of the University of
Wisconsin. The computer costs reflect the rates charged by MACC for their
services. These costs may vary for other users depending on the rates charged
by their computer service or on the existence of in-house computing services.
The efficiency of computer-user interaction also greatly affects the time and
cost. The speed of assembling and executing computer jobs may vary greatly
for someone keypunching runstream card decks and someone directly wired to the
computer through an interactive color terminal. As described by Ballard and
Eastwood (1977), this method of cost accounting is accurate for the "home
system" but is somewhat inflexible.
Nearly all the computer runs for this project were assembled and executed
on a 30-character-per-second hard-copy interactive terminal. Most output was
from an MACC line printer. The time and expenses for producing a basic
supervised digital classification are shown in Table 5. The time entries
represent the time required for manual operations of compiling and entering
runstreams,and data. Cost entries are in three columns: The first represents
the cost of executing the specific program; the second includes the associated
charges for paper, tape utility, and file utility; and the third equals the
cost of one run multiplied by the typical number of iterations necessary to
achieve a successful classification. Other types of classifications and data
enhancements entail some additional costs (Taole 7).
The information in Taole 6 can be further grouped into general stages of
processing. First, the initial raw-data generation involved digitizing,
checking the file, correcting the data, and selecting a representative subset.
Tne initial creation of a usable digital data set required about 1.5 h and
cost nearly $50. This process was the most cost intensive.
Surprisingly, the manual interpretation and ground survey efforts were
not the most time-consuming steps. Tnis stage of the process required about 5
h, which included the time for tracing and tentatively identifying all
resources shown in Figure 12 and ground checking. The travel costs of this
aspect of tne study were negligible because the locales were nearby. If a
distant area were being surveyed, the expense of visiting the area could be
-40-
-------
TABLE 6. TIME AND EXPENSES OF A BASIC FILM ANALYSIS WHEN UNIVERSITY OF
WISCONSIN EMDAG SOFTWARE PACKAGE AND MACC COMPUTING SERVICES ARE USED
Operation
Scanning 70 mm image
on densitometer at
100 aperture (tape)
File check with
PRINT program
Time (min.)
Each
run Total
20 20
10 20
Cost ($)
Program
incl. pages etc.
$12.00
1.25
Total
Iterations cost
1 $12.00
2 2.50
Create characteristic
curve D vs. logE
Create raw data file
with OPTRON ptrgram
100 tracks
= 150,000 pixels
Calibrate data with
CORRECT program
View data set with SLICE
Reduce data with SUBSET
Overprint presentation
with SLICE
Ground truth
Manual interpretation
Ground survey
Select training sets
(80 sets)
(Additions)
Record training sets
(Additions)
2.00
3.00
2.00
3.00
5
5
10
5
120
180
75
(20)
135
(20)
5 1.50
10 4.00
20 1.20
10 3.00
120
180
120
180
1
2
2
6
1
1
3
3
1.50
8.00
2.40
18.00
(continued)
-------
TABLE 6 (continued)
Operation
Run training set
program TRAIN
' Additions)
Look at signature
histograms with
HSGRAM
(Additions)
Look at signature bar
diagrams with CLASSBAR
Modify training sets
with MODSET
Box classification
with BOX4
(Additions)
For reducing noise
SMOOTH neighborhoods
of nine pixels, two
iterations
Classes assigned colors
with COLOR
File prepared for film
writing with FILM70
Writing and developing
film
Time (rain.) Cost ($)
Each Program
run Total incl . pages etc.
20 40 $1.00
(10) (0.25)
3 9 6.00
(3) (1.50)
3 9 1.00
30 90 0.75
60 100 5.00
(20)
10 20 3.50
20 40 1.00
10 20 2.75
60 120
Total
Iterations cost
3 $1.50
3 9.00
3 3.00
3 2.25
3 15.00
2 7.00
2 2.00
2 5.50
2
Total 13h 11min 19h 3min $48.95 $94.65
-42-
-------
TABLE 7. ADDITIONAL TIME AND EXPENSES FOR DATA ENHANCEMENTS,
TRANSFORMATIONS, AND MORE SOPHISTICATED CLASSIFICATIONS IN DIGITIZED FILM
ANALYSIS WHEN THE UNIVERSITY OF WISCONSIN EMDAG SOFTWARE PACKAGE
AND MACC COMPUTING SERVICES ARE USED
Operation
Enhance data with
FILM RATIO
Enhance data with
TEXTURE
Enhance data with
canonical TRANSFORM
Calculate training set
statistics with STATS
View elliptical signature
dimensions with ROTRAIN
Time (min.)
Each
run Total
5 5
5 5
10 10
5 10
3 3
Cost ($)
Program
incl. pages etc.
$1.75
3.00
2.00
0.25
0.25
Total
Iterations cost
1 $1.75
1 3.00
1 2.00
2 0.50
1 0.25
Load elliptical signatures
into look-up table with
LOADE3 or LOADE4
(30 class) 20 40 3.00 2 6.00
Elliptical classifier
TABCLASS (w/o minimum
distance applied) 10 20 5.00 2 10.00
TABCLASS (w/minimum
distance applied) 10 20 15.00 2 30.00
Total 1h 3min 1h 53min $26.10 $53.50
-43-
-------
more than the entire digital analysis. Another tempering thought is that the
manual efforts were directed at identifying representative resources in the
hopes that the signatures of all resources could be represented. Once a
resource was located on the ground and registered on the digital data base,
little further effort was expended toward locating all other occurrences of
that resource. A totally manual effort would require a more conscientious
effort to accurately locate, register, and map every resource and all
occurrences.
The next general category of processing was the location of training sets
and the determination of signatures. This step was by far the lengthiest,
taking over 9 h, nearly half the total analysis time. This step more than any
other could be expedited by use of an interactive color terminal to eliminate
the necessity of scrutinizing line-printer output.
The classification segment of the process takes only one-tenth the total
time, but represents roughly one-fourth the total expense. With luck, three
iterations of a classification as shown in the cost estimate are not needed.
However, a first classification does not usually characterize all resource
occurrences completely. The remaining time and money are allocated for
preparing a film image of the classification. Three hours and $7.50 were
typically consumed in this endeavor. The costs of film and chemicals were not
included.
With more elaborate and sophisticated methods of data enhancement (such
as ratios, canonical transformations, and texture quantification) or
classification (elliptical classifier and minimum distance classifier), the
expenses of achieving a satisfactory classification may be inflated an extra
$20 to $50. Little additional time is expended for these processes.
Economies of scale heavily influence the total process cost. An economy
of scale can be realized by decreasing the scale of the imagery; the data set
is quartered by halving the scale. Although the efficiency of locating
training sets is enhanced, file maintenance is cheaper, and the paper costs
are reduced, nevertheless, the savings are not very substantial because every
step of tne process must still be performed.
Substantial savings can be realized by the ability to classify an entire
roll of film based on the signatures of representative resources taken from
the first image or two on the roll. Subsequent image classifications may
require signature augmentation; however, the major time factor in digital
analysis is the processing of training sets. Any reduction in this aspect of
the analysis should result in significant savings. Therefore, aerial
photographic missions should be carefully planned so that all desired scenes
are included with exposure and processing variables reduced as much as
possible.
In comparing digital film analysis to ground survey and manual
interpretation, the scale is a very important consideration. For single
images where quantification of resources is not important, digital analysis Is
not competitive with a manual approach. Digital analysis may become more
efficient where very complex resource patterns exist, such as roadways and
-------
housas mixed with vegetation in an urban area. Wh3n quantification of the
areal extent of resources is desired, a manual interpreter must use a
planimeter, grid overlay, or weighing technique to achieve what a digital
analysis inherently accomplishes.
An important consideration for potential users of this technology is that
classification success is not assured. The documented costs of processing
digitized film information may be applied to a wholly unsatisfactory
classification effort. Careful scrutiny of resource signatures after
selection of training sets should decide the issue of whether to proceed with
the highly cost-intensive classification portion of the process.
ACCURACY
The primary goal of a digital film analysis is to separate and classify
selected resources. In most cases the success depends on the level of detail
desired and the complexity of the resource base.
Other researchers in digital image analysis have had various rates of
success. Unfortunately, the methods employed to quantify the accuracy were
seldom, if ever, documented. Doverspike et al. (1955) suggested that color
densities alone would not be sufficient to classify land use successfully; by
comparing signatures manually the success rate averaged only about 55$.
Studies limited to gross-resolution satellite data (Bauer and Cipra 1973,
Goodenough and Shlien 197^, Owen-Jones 1977) approached success rates of
greater than 90$ for classification schemes incorporating simplified resource
categories. Owen-Jones (1977) recognized that detailed digitized film
analysis of vegetation complexes presented problems of signature complexity
that reduced accuracy. Detailed studies of forest species with several bands
of multispectral imagery by Rohde and Olsen (1972) achieved accuracies of
nearly $5% correct identification, but they acknowledged that tree signatures
overlapped badly with those of adjacent fields. Wacker and Landgrebe (1972)
achieved success rates of about 50? in separating a few crop types by using
scanned film and several classifiers.
The scenes analyzed in this study presented fairly complex associations
of vegetation, water, and structures. The highest possiole level of detail
was sought. Most vegetation types were separated on the basis of classical
ecological concepts of association.
The various classification attempts met with mixed success. Even with
properly selected training sets, confounded resource signatures were common.
If these represented major resources, the scene would probably be poorly
classified. Some of the problems were due to physiognomic similarities among
associations. For example, wetland areas with a canopy dominated by forbs
such as aster (Aster sp.), goldenrod (Solidago sp.), and joe-pye-weed
(Eupatorium sp.), and a very dense tussock sedge ground layer would properly
be classified as sedge meadow. Upland old fields and other vegetation types
with a forb canopy were commonly confounded with the sedge meadows described
above. Although the goal was to separate these vegetation types, tne
signatures did not permit separation. Possibly, a finer resolution might
allow identification of separate signatures for the forb species, and the
-45-
-------
associations could be split into their component species. However, as
discussed in the previous section on resolution, this amount of detail is
counterproductive. Resources must be defined such that groups of pixels can
be delineated for training sets. Resolution not only relates to the size of a
single pixel, but also to the required occurrence together of a minimum of
four pixels of a like resource to form a training set. As the
resolution becomes progressively finer, the erratic signatures increase and
the uniformity of association signatures decreases.
On the other hand, if resolution is allowed to become less fine, greater
classification success can be achieved but only for gross classes of
vegetation and other resources. For example, a study of impervious surfaces
would attempt to separate all vegetation as one class distinct from impervious
surfaces or bare soil. Tnis low level of detail would not be satisfactory for
most inventories of general land resources. However, as shown in Figure 11,
screening lowland from upland resources simplified the classification effort
and provided success in this study.
The above discussion is not meant to imply that detailed classifications
of plant associations are impossible. We found that whan the associations are
phenologically distinct they can be successfully separated. Typically this
meant that the associations were either heavily dominated by a single species
(such as cropland) or that the association codominants were distributed in
such a way that a homogeneous signature was reflected.
The latter situation occurred at Hook Lake, an acid bog. The classic bog
community is characterized by distinct zonation of vegetation types. This is
illustrated in Figure 1b where a central stand of tamarack is encircled by
leatharleaf over a sphagnum mat, which is in turn surrounded by an uneven zone
of wiregrass sedge. The bog is bordered by a moat of open water. Cattail are
proximate to the edges of open water, and some mixed macrophyte beds of
pickerel weed, arrowhead, and duckweed occur in the water. The uplands had
assorted cropland and mixed oak hardwoods woodlots, some of which were
confounded with the lowland resources.
The classifications of the Hook Lake scene shown in Figure 1? were the
most successful classifications of a vegetation complex attempted and are
considered representative of the maximum classification accuracy for complex
vegetation types when the descrioed classifiers are used. Accuracy was
measured by direct comparisons between the classifications and a blowup of the
scanned image written back onto film in its scanned form (Figure 13). Random
pixels from the classifications were compared to their counterparts on the
original scanned image, and a decision was made on the correctness of the
classified pixel. This decision was based on the manual interpretation map
and in reference to the original photo.
An index of accuracy was then derived by calculating the percentage of
correctly identified pixels. This general index was presented in the form of
a confusion matrix which showed the classification success of individual
resources and those that were confounded. The results of this tally for the
unsmoothed, fully supervised parallelepiped classification are shown in Table
8. Of the 200 randomly sampled pixels, 147, or nearly 75%, were correctly
-46-
-------
a.
c.
Unsmoothed parallelepiped; green=
cattail, olive=wiregrass, It.
orange=leatherleaf, dk. orange=
woods, beige=emergents*, blue=
water, dk. green=deep water,
yellow=corn, red=alfalfa^, maroon=
tamarack^.
b.
Smoothed parallelepiped; green=
cattail, brown=wiregrass, dk.
orange=emergents*, blue=water, dk.
brown=deep water, yellow=corn, red=
alfalfa-^, maroon=tamarack.
Smoothed elliptical: Same color
key as unsmoothed elliptical.
Unsmoothed elliptical; It. green= c
cattail, dk. green=wiregrass,
orange=leatherleaf, dk. gray=woods,
pink=emergents*, blue=water, dk.
blue=deep water, It. red=alfalfa,
yellow=corn, teal=tamarack.
Figure 17. Classifications of the Hook Lake study area, October 1976.
*Emergents were predominantly arrowhead.
-47-
-------
TABLE 8. CONFUSION MATRIX FOR THE UNSMOOTHED PARALLELEPIPED CLASSIFICATION
OF THE HOOK LAKE STUDY AREA.a
Computer Class
Ground Identity
Cattail
Leatherleaf
Water
Emergents
Tamarack
Wire grass
Upland Woods
Alfalfa
Corn
Edge
Shade
Accuracy percentage
minus commissions
CD
03
C 10
O C
0 O
1/1 T3 f-l -H
MH 13 o; d) (/i
oj o -H p., (/)
-------
classified. One-fourth of the improperly classified pixels resulted from the
confusion of water and shade, both essentially black. If this single pair of
confounded entities were united, the accuracy would be improved to more than
30$.
In view of the importance of separating wetland resources from upland
resources, it is interesting that of the incorrect pixels 66% represented
wetland resources confounded with upland resources, nearly 20$ ware confounded
upland resources, and only 15% were among wetland resources. This suggests
both that higher classification success can be achieved for scenes including
exclusively upland or exclusively lowland resources and that when photographed
together the separation of these resource types is difficult.
Resources having the highest classification success included cattail,
leatherleaf, wiregrass sedge, water, and corn. The most poorly classified
resources included shade, alfalfa, edge, trees, and tamarack. The
difficulties encountered with the latter resources were not surprising because
of aforementioned textural or seasonal limitations. If these limitations
could be overcome, and these "problem" resources better classified, the
accuracy would be improved to greater than 90$.
The accuracy for an elliptical classifier employing a minimum distance
routine for overlapping signatures is presented in Table 9. The overall
classification success was 12%. Among the more easily classified resources
were leatherleaf, water, alfalfa, and corn. Over one-fourth of the
misclassified pixels were transitional or edge pixels. Nearly 80$ (44 of 56)
of the misclassified pixels involved trees. Although the elliptical
classifier more accurately portrayed texturally smooth resources, it did not
make use of textural information and was thus hindered in separating roughly
textured resources.
One approach to eliminating aberrant pixels from a classification is to
apply a smoothing routine to the classification whereby Bach pixel is compared
to its neighbors, much like the texture routines. If the neighborhood of a
pixel is uniformly different from the pixel itself, the pixel is defined as
erratic and is reclassified to match its surrounding neighbors. For extremely
complex scenes where resources may be accurately portrayed by small groups or
even single pixels, the smoothing routines probably detract from the accuracy.
For relatively expansive resources, however, such as crops, woodlands, or
meadows, a smoothed version may be more accurate. This hypothesis was tested
by examining the same 200 pixels usad in the previous accuracy tests and
comparing them to a smoothed version of the same classifications (Figure 17).
Table 10 and Table 11 present the results of tnese comparisons.
The overall accuracy of the parallelepiped classifier was improved to
34.5? for the smoothed version (Table 10). Tree classes (tamarack, shade, and
upland woods) accounted for nearly 70$ of the problems; the accuracy of the
smoothed classification, discounting problems due to trees, would have besa
937$.
Equally good classification accuracy (34.5$, Table 11) was achieved by
the smoothed elliptical classifier (Figure 17). All classes were consistently
-49-
-------
TABLE 9. CONFUSION MATRIX FOR THE SMOOTHED PARALLELEPIPED CLASSIFICATION
OF THE HOOK LAKE STUDY AREA.3
Computer Class
Ground Identity
Cattail
Leatherleaf
Water
Emergents
Tamarack
Wiregrass
Upland Woods
Alfalfa
Corn
Edge
Shade
Accuracy percentage
minus commissions
DO
rt
p
C I/I l/l O 4-1 -H
-H P.XWS -HXE
i i ^H coo) rt -POO
HO ocflhtS^H CnS
oJjifHbO^ibCqt i
-------
TABLE 10. CONFUSION MATRIX FOR THE UNSMOOTHED ELLIPTICAL TABLE LOOK-UP
CLASSIFICATION OF THE HOOK LAKE STUDY AREA.3
Computer Class
Ground Identity
Cattail
Leatherleaf
Water
Emergents
Tamarack
Wiregrass
Upland Woods
Alfalfa
Corn
Edge
Shade
Accuracy percentage
minus commissions
CD
DO
nj
C 1/1
O £
O O
0 "O fn -H
Mj 13 0) 0) 1/1
TO O -H PH (/I
m tn i/i o (4-1 .H
.-H 4-)ASV)S -HXE
i i fn COaJ nj -POO
HO onjM-aMH cm
nj.CfH&QMtQC' i dJCDJ-tt/3
+J4-)DMnj
-------
TABLE 11. CONFUSION MATRIX FOR THE SMOOTHED ELLIPTICAL TABLE LOOK-UP
CLASSIFICATION OF THE HOOK LAKE STUDY AREA.3
Computer Class
Ground Identity
Cattail
Leatherleaf
Water
Emer gents
Tamarack
Wiregrass
Upland Woods
Alfalfa
Corn
Edge
Shade
Accuracy percentage
minus commissions
,x in s -H >, g
i 1 f-l CO03 03 -MOO
HO) (UoS^iTJ^H (303
oSjsfHtyjfnWiC' i tuaJMin
^->-4-> O3 G 03 -H £L
-------
better identified than in the unsnoothad version except for water, shade, and
transitional. Because the resolving power of the smoothed classification is
essentially degraded, the more intricately expressed resources, such as the
three listed, are expected to classify less successfully.
Although the same overall classification accuracy was attained for the
smoothed box classifier and smoothed elliptical classifier, their relative
success with individual resources varied. Wooded resources (tamarack and
upland woods), alfalfa, and mixed emergeats were much better classified with
the elliptical classifier. Water and cattail were more successfully
identified with the box classifier.
-53-
-------
REFERENCES
Akca, A. 1971. Identification of land use classes and forest types by means
of microdensitcmetry and discriminant analysis, p. 147-164. In
Application of remote sensors in forestry, 25th Int. Union For. Res.
Organ. Congr., Washington, D.C.
Aldred, A.M. 1972. Decisions on combining data from several sensors, p.
681-690. In First Canadian symposium on remote sensing. Ottawa, Ont.
Ali, M., and J.K. Aggarwal. 1976. Automatic detection and classification of
infestations of crop insect pests and diseases from infrared aerial color
photographs, p. 4B1-4B11. In Third annual symposium on machine processing
of remotely sensed data. Laboratory for Applications of Remote Sensing,
Purdue Univ.
Anderson, R.R. 1972. Applications of high-altitude remote sensing to coastal
zone ecological studies, p. 191-195. In Operational remote sensing: An
interactive seminar to evaluate current capabilities. Am. Soc.
Photogramm. Houston, Tex.
Anson, A. 1968. Photographic interpretation from color aerial photographs,
p. 381. In J.T. Smith [ed.] Manual of color aerial photography. Am.
Soc. Photogramm. Falls Church, Va.
Avery, T.E. 1966. Foresters guide to aerial photo interpretation. U.S.
Dep. Agric. For. Serv. Agric. Handb. 308. 40 p.
Baldwin, M., H.M. Smith, and H.W. Whitlock. 1947. The use of aerial
photographs in soil mapping. Photogramm. Eng. 13:532-536.
Ballard, R.J., and L.F. Eastwood, Jr. 1977. Estimating costs and performance
of systems for machine processing of remotely sensed data, p. 208-214. In
Fourth annual symposium on machine processing of remotely sensed data.
Laboratory for Applications of Remote Sensing, Purdue Univ.
Bauer, M.E., and J.E. Cipra. 1973. Indentification of agricultural crops by
computer processing of ERT3 MSS data. Laboratory for Applications of
Remote Sensing, Purdue Univ. LARS Inf. Note 030173. 9 p.
Benson, L.A., C.J. Frazee, and F.A. Waltz. 1973. Analysis of remotely sensed
data for detecting soil limitations, p. 3A11-3A19. In First annual
symposium on machine processing of remotely sensed data. Laboratory for
Applications of Remote Sensing, Purdue Univ.
-54-
-------
Bernstein, R. 1976. Digital image processing of earth observation sensor
data. IBM J. Res. Dev. 20:40-57.
Berry, J.K., and J.A. Smith. 1977. An overview of vegetation canopy
modelling for signature correction analyses, p. 194. In Fourth annual
symposium on machine processing of remotely sensed data. Laboratory for
Applications of Remote Sensing, Purdue Univ.
Bryant, N.A., and A.L. Zobrist. 1976. IBIS: A geographic information system
based on digital image processing and image raster datatype, p. 1A1-1A7.
In Third annual symposium on machine processing of remotely sensed data.
Laboratory for Applications of Remote Sensing, Purdue Univ.
Bushnell, T.M. 1932. A new technique in soil mapping. Am. Soil Surv.
Assoc. Bull. 13:74-81.
Carter, P., and W.E. Gardner. 1977. An image-processing system applied to
earth-resource imagery, p. 143-162. In E.G. Barrett and L.F. Curtis [ed.]
Environmental remote sensing 2: Practices and problems. Crane Russak,
New iork.
Carter, V. and W.R. Stewart. 1975. Seasonal color infrared photographs for
mapping inland wetlands on U.S. Geological Survey 7.5-minute quadrangles,
p. 143-161. In Fifth biennial workshop on color aerial photography in the
plant sciences and related fields. EROS Data Center, (Sioux Falls,
S.Dak.).
Cole, M.M., and E.S. Owen-Jones. 1974. Remote sensing in mineral
exploration, p. 49-66. In E.G. Barrett and L.F. Curtis [ed.]
Envirormental remote sensing. Edward Arnold, London.
Colwell, fl.N. 1965. The extraction of data from aerial photographs by human
and mechanical means. Photogrammetry 20:211-228.
Colwell, fl.N. 1958. Remote sensing of natural resources. Sci. Am.
218:54-69.
Cox, T.L., H.G. Hitchcock, and S.G. Weber. 1975. Processing remotely sensed
data for dimensional analysis, p. 1B37-1B44. In Second annual
symposium on machine processing of remotely sensed data. Laboratory for
Applications of Remote Sensing, Purdue Univ.
Curtis, J.T. 1959. The vegetation of Wisconsin: An ordination of plant
communities. Univ. Wisconsin Press, Madison, Wis. 657 p.
Curtis, L.F. 1977. Ground monitoring for airborne and space studies of
land-use and soil conditions, p. 192-215. In E.C. Barrett and L.F. Curtis
[ed.] Environmental remote sensing 2: Practices and problems. Crane
Russak, New York.
-55-
-------
Doverspike, G.E., F.M. Flynn, and R.C. Heller. 1965. Microdensitometer
applied to land-use classification. Photogramm. Eng. 41:294-306.
Driscoll, R.S., P.O. Currie, and M.J. Morris. 1972. Estimates of herbaceous
standing crop by raicrcdensitometry. Proc. Am. Soc. Photogramm.
38:358-364.
Driscoll, R.S., R.N. Reppert, and R.C. Heller. 1974. Microdensitometry to
identify plant communities and components on color infrared aerial photos.
J. Range Manage. 27:66-70.
figan, W.C., and M.fi. Hair. 1971. Automated delineation of wetlands in
photographic remote sensing, p. 2231-2252. In Proceedings of the seventh
international symposium on remote sensing of the environment. Vol. 3.
Univ. Michigan.
Fisher, L.T., R.rt. Kiefer, F.L. Scarpace, and S.L. Wynn. 1977. Remote
sensing, p. 126-131. In Documentation of environmental change related to
the Columbia Electric Generating Station. Inst. Environ. Stud. Rep. 82.
Univ. Wisconsin-Madison.
Goodenough, D., and S. Shlien. 1974. Results of cover-type classification by
maximum likelihood and parallelepiped methods, p. 135-164. In Second
Canadian symposium on remote sensing. Univ. Guelph, Guelph, Ont.
Grinnell, H.R., and M.S. Conway. 1972. The economics of remote sensing of
forest land, p. 691-696. In First Canadian symposium on remote sensing.
Ottawa, Ont.
Hauska, H., and P.h. Swain. 1976. Application of a class of sequential
classifiers to multitemporal remote sensing data, p. 2B-7. In Third
annual symposium on machine processing of remotely sensed data.
Laboratory for Applications of Remote Sensing, Purdue Univ.
Heiber, R.H. 1976. General polygons used to determine training and test
areas in digital remote sensing imagery, p. 1A-29. In Third annual
symposium on machine processing of remotely sensed data. Laboratory for
Applications of Remote Sensing, Purdue Univ.
Hoffer, R.M., P.E. Anuta, and T.L. Phillips. 1972. ADP, multiband and
multiemulsion digital photos. Photogramm. Eng. 38:989-1001.
Hyde, R.F., S.N. Goward, and P.W. Mausel. 1977. ISURLS levels
classification: A low cost approach to multispectral data analysis, p.
322-332. In Fourth annual symposium on machine processing of remotely
sensed data. Laboratory for Applications of Remote Sensing, Purdue Univ.
Kamat, D.S., K.L. Majumder, T.J. Majumder, I.C. Matieda, C.V.S. Prakash, and
V.L. Swaminathan. 1977. Machine processing of aerial data for
agricultural resources inventory and survey experiment, p. 124-134. In
Fourth annual symposium on machine processing of remotely sensed data.
Laboratory for Applications of Remote Sensing, Purdue Univ.
-56-
-------
Kan, E.P., D.L. Ball, J.P. Basu, and R.L. Smelser. 1975. Data resolution
versus forestry classification and modelling, p. 1B24-1B36. In Second
annual symposium on machine processing of remotely sensed data.
Laboratory for Applications of Remote Sensing, Purdue Univ.
Kiefer, R.W. 1970. Effects of date of photography on airphoto interpretation
using color and color infrared films. Proc. Int. Symp. Phot. Navig. (Ohio
St. Univ.) 1970: 100-117.
Knipling, E.B. 1970. Physical and physiological basis for the reflectance of
visible and near infrared radiation from vegetation. Remote Sens.
Environ. 1:155-159.
Kuril, A.D. 1970. Color and IR photos for soils. Photogramm. Eng.
36:475-482.
LeShack, L.A. 1971. ADP of forest imagery. Photogramm. Eng. 37:855-895.
Lukens, J.E. 1958. Color aerial photography for aquatic vegetation surveys,
p. 441-446. In Proceedings of the fifth symposium on remote sensing of
the environment. Univ. Michigan.
Lyon, R.J.P. 1969. Geological remote sensing: A critical evaluation and
prognosis, p. 349-402. In Principe de la detection a distance et
application a 1'etude des resources terrestres. Centre Nationale D'Etudes
Spatiales and Univ. Michigan. Paris, France.
Moore, R.T., M.C.Stark, and L. Cahn. 1964. Digitizing pictorial information
with a precision optical scanner. Photogramm. Eng. 40:923-931.
Nunnally, N.R., and R.E. Witmer. 1970. Remote sensing for land-use studies.
Photogramm. Eng. 36:449-453.
Olsen, C.E., Jr., and R.E. Good. 1951. Seasonal changes in light reflectance
from forest vegetation. Photogramm. Eng. 27:107-114.
Olson, D.P. 1964. The use of aerial photographs in studies of marsh
vegetation. Maine Agric. Exp. Stn. Tech. Ser. Bull. 13. 62 p.
Owen-Jones, E.S. 1977. Densitometric methods of processing remote sensing
data, with special reference to crop-type and terrain studies, p. 101-124.
In E.G. Barrett and L.F. Curtis [ed.] Environmental remote sensing 2:
Practices and problems. Crane Russak, New ₯ork.
Pearson, R.L., and L.D. Miller. 1972. Remote mapping of standing crop
biomass for the estimation of tne productivity of the short grass prairie.
Pawnee National Grasslands, Colorado, p. 1357-1381. In Proceedings of the
eighth international symposium on remote sensing of the environment.
Univ. Michigan.
Reppert, J.N., R.S. Driscoll, and R.C. Heller. 1969. Wildland site and
treatment delineation by microdensitometry using 70 mm color aerial
-57-
-------
photography, p. 251-252. In Proceedings of the American Society of
Photogrammetry-American Cong. Surv. Map. Portland, Ore.
Rohde, W.G., and C.E. Olsen, Jr. 1972. Multispectral sensing of forest tree
species. Photogramm. Eng. 38:1209-1215.
Scarpace, F.L. 1977. Densitometry on color and color-IR imagery, p. 301-318.
In Proc. Am. Soc. Photogramm., 43rd Annu. Meet., Washington, D.C.
Scarpace, F.L., fl.W. Kiefer, S.L. Wynn, B.K. Quirk, and G.A. Friederichs.
1975. Quantitative photo-interpretation for wetland mapping, p. 750-771.
Jn Proc. Am. Soc. Photogramm., 41st Annu. Meet., Washington, D.C.
Scarpace, F.L., and A.W. Voss. 1975. Analysis of multi-layered films, p.
26-34. In Proc. Am. Soc. Photogramm., 40th Annu. Meet., St. Louis, MO.
Seher, S.J., and P.T. Tuellar. 1973. Color aerial photos for marshland.
Photograram. Eng. 38:489-499.
Simonett, D.S. 1969. Remote sensing studies in geography: A review, p.
467-497. In Principe de la detection a distance et application a 1'etude
des resources terrestres. Centre Nationale D'Etudes Spatiales and Univ.
Michigan. Paris, France
Smedes, H.W., H.G. Linnerud, L.D. Woolaver, and S.J. Hawks. 1971. Digital
computer mapping of terrain by clustering techniques using color film as a
three band sensor, p. 2057-2072. In Proceedings of the seventh
international symposium on remote sensing of the environment. Vol. 3-
Univ. Michigan.
Spurr, S.H. 1960. Aerial photographs in forestry. Ronald Press, New York.
333 P.
Steiner, D. 1972. Multispectral-multitemporal photography and automatic
terrain recognition, p. 601-609. In First Canadian symposium on remote
sensing. Ottawa, Can.
Svedlow, M., C.D. McGillem, and P.fi. Anuta. 1976. Experimental examination
of similarity, measures and preprocessing methods used for image
registration, p. 4A9-4A17. In Third annual symposium on machine
processing of remotely sensed data. Laboratory for Applications of Remote
Sensing, Purdue Univ.
Thomson, D.E. 1972. Airborne remote sensing of Georgia tidal marshes, p.
126-139. I_n Operational remote sensing: An interactive seminar to
evaluate current capabilities. Am. Soc. Photogramm. Houston, Tex.
Townshend, J.R.G. 1977. Ground information for the earth resources Skylark,
p. 216-245. In E.G. Barrett and L.F. Curtis [ed.] Environmental remote
sensing 2: Practices and problems. Crane Russak, New lork.
-58-
-------
van Genderen, J.L., and B.f. Lock. 1977. Testing land-use map accuracy.
Photogramm. Eng. 43:1135-1137.
Wacker, A.G., and D.A. Landgrebe. 1972. Minimum distance classification in
remote sensing, p. 577-599. In First Canadian symposium on remote
sensing. Ottawa, Ont.
Whitman, A.I., and K.L. Marcellus. 1973. Textural signatures for wetland
vegetation. Proc. Am. Soc. Photogramm. 1973 (Pt. II): 979-992.
Wiersma, D.J., and D. Landgrebe. 1976. The use of spatial characteristics
for the improvement of multispectral classification of remotely sensed
data, p. 2A18-2A25. In Tnird annual symposium on machine processing of
remotely sensed data. Laboratory for Applications of Remote Sensing,
Purdue Univ.
Williams, D.H., and J.K. Aggarwal. 1976. Computer location of citrus trees
using color aerial infrared transparencies, p. 3B-2. In Third annual
symposium on machine processing of remotely sensed data. Laboratory for
Applications of Remote Sensing, Purdue Univ.
-59-
-------
TECHNICAL REPORT DATA
(Please read Instructions on the reverse before completing}
REPORT NO.
EPA-66C/3-80-054
2.
3. RECIPIENT'S ACCESSION NO.
4. TITLE ANDSUBTITLE
Mapping Vegetation Complexes with Digitized Color
Infrared Film : Wisconsin Power Plant Impact Study
5. REPORT DATE
JUNE 1980 ISSUING DATE,
6. PERFORMING ORGANIZATION CODE
7. AUTHOR(S)
Warren J. Buchanan and Frank L. Scarpace
8. PERFORMING ORGANIZATION REPORT NO.
9. PERFORMING ORGANIZATION NAME AND ADDRESS
Institute for Environmental Studies
Environmental Monitoring and Data Acquisition Group
University of Wisconsin-Madison
Madison, Wisconsin 53706
10. PROGRAM ELEMENT NO.
1BA820
11. CONTRACT/GRANT NO.
Grant R803971
12. SPONSORING AGENCY NAME AND ADDRESS
ENVIRONMENTAL RESEARCH LABORATORY - DULUTH, MN
OFFICE OF RESEARCH AND DEVELOPMENT
U.S. ENVIRONMENTAL PROTECTION AGENCY
DULUTH, MINNESOTA 55804
13. TYPE OF REPORT AND PERIOD COVERED
Final; 7-75 to 7-77
14. SPONSORING AGENCY CODE
EPA/600/03
15. SUPPLEMENTARY NOTES
16. ABSTRACT
Environmental inventories of sites typically include some efforts to map land
cover. These maps are used to describe the existing environment or to monitor
and assess environmental changes. Digitized photo-analysis was investigated in this
context for mapping vegetation complexes. The goal was to assess the utility of this
technique for site analysis by testing several variables.
Some images were successfully classified, but other classification attempts
were abandoned; digitized film analysis was not an entirely reliable mapping
technique. Resource signatures often overlapped. Additional research should
determine if multiseasonal data overlayment, data-enhancement transformations, and
machine adjustments can create the necessary reliability for operational applications,
17.
KEY WORDS AND DOCUMENT ANALYSIS
DESCRIPTORS
b. IDENTIFIERS/OPEN ENDED TERMS
COSATI Field/Group
Vegetation
Mapping
Color infrared film
Electric power generation
Remote sensing of
environmental condi-
tions
Wisconsin power plant
study
02/D
08/B
08/F
18. DISTRIBUTION STATEMENT
RELEASE TO PUBLIC
19. SECURITY CLASS (ThisReport)
UNCLASSIFIED
21. NO. OF PAGES
-72-
20. SECURITY CLASS (This page)
UNCLASSIFIED
22. PRICE
EPA Form 2220-1 (Rev. 4-77) PREVIOUS EDITION is OBSOLETE
-60-
USGPO 660-804 8/80
------- |