.. (I:

‘ 1 Elfin.

 

. 4.1: 5v

514:“; :v
3 .

ffam

.6):
.15.:

H.

855....

z 4.31:?! in:

. »
tn..mm...r...3§ .

EH 0’?
«u’VIvthAn...
;- lil-

u
31.”... 5
“v3... fiflwhfinwmma 5.» M;

‘,..

THESYS

Y'x

 

LIBRARY

”when State
University

 

 

 

This is to certify that the
thesis entitled
MICHIGAN FOREST DEFOLIATION DETECTION

USING MULTITEMPORAL' NDVI SIGNATURES i
DERIVED FROM AVHRR SATELLITE IMAGERY

presented by

Kenneth A. Duda

has been accepted towards fulfillment
of the requirements for

M. A. degree in Geography

eel/WM

Major professor

Date October 2, 1998

0.7639 MS U is an Affirmative Action/Equal Opportunity Institution

 

 

 

PLACE IN REFURN BOX to rem
TO AVOID FINES return on or

- “I.

ove this checkout from your record.
before date due.

MAY BE RECAILED with earlier due date if requested.

 

DATE DUE

DATE DUE

DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

MICHIGAN FOREST DEFOLIATION DETECTION
USING MULTITEMPORAL NDVI SIGNATURES
DERIVED FROM AVI'IRR SATELLITE IMAGERY

by
Kenneth A. Dude

A THESIS
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
MASTER OF ARTS
Department of Geography

1998

ABSTRACT

MICHIGAN FOREST DEFOLIATION DETECTION USING MULTITEMPORAL
NDVI SIGNATURES DERIVED FROM AVI-IRR SATELLITE IMAGERY

by

Kenneth A. Dude

Multitemporal Advanced Very High Resolution Radiometer (AVHFIR) satellite data and
ancillary information stored in a geographic information system (GIS) were used to
assess the effectiveness of employing maximum likelihood (ML) and backpropagation
(BP) artificial neural network (ANN) classifiers to discriminate areas of forest defoliation
caused by gypsy moths in the Lower Peninsula of Michigan and in Roscommon County
during a single year. The AVHFIFI sensor offered comparatively high temporal
resolution. The Normalized Difference Vegetation Index (NDVI) was calculated for each
of three temporal composites corresponding to pre-, peak and post defoliation stages,
then sampled image data were input to the classifiers. Overall accuracies ranged from
19.7% to 78.8%. Based on the Kappa statistic, results were similar for the ML and ANN
methods in the Lower Peninsula, however the ML method markedly outperformed the
BP ANN technique at the county level. and compared favorably with other reported
results. The generalizing nature of the ML approach yielded greater accuracy than did
the attempt by the ANN to seek detail while limited by the coarse spatial resolution of
the imagery, which resulted in heterogeneous pixel content and limited availability of

training samples.

Copyright by
KENNETH ALAN DUDA
1998

My parents. Clifford A. and Ellzabeth J. Dude, provided constant early
encouragement as I sought answers to questions about our world, and they
instilled in me the desire to make the best possible use of those abilities which I
have been given. Always placing a high value on the pursuit and application of
knowledge and the exercise of reason, their unwavering support throughout the
years has contributed to all of my accomplishments, including this work, which I
dedicate to them.

ACKNOWLEDGEMENTS

Foremost. I must acknowledge the insightful guidance of my graduate committee, in
particular my advisor Daniel G. Brown. Ph.D., and also additional readers Bruce w.
Pigozzl, Ph.D., David P. Luech. Ph.D. and Jon B. Burley, Ph.D. Their ready
availability. timely suggestions and continual patience, especially during the latter phase
of the project, helped bring this work to a rapid conclusion.

In addition. I received recommendations concerning the application of artificial neural
networks to satellite imagery classification during visits with Anil P. Jain, Ph.D.,
Computer Science Department, and Fathi M. A. Salem, Ph.D., Electrical Engineering
Department, both at Michigan State University. I also received similar early counsel from
Tom Vogl, Ph.D. and Tom German, both of the Environmental Research institute of
Michigan (ERIM), as well as from Sucharita Gopal, Ph.D., Department of Geography,

Boston University.

The basic data were directly provided by the Center for Remote Sensing (CRS) at
Michigan State University, and I completed some data preprocessing while employed
there. Funding for the CRS project and additional support were provided by Frank Sapio
of the Michigan Department of Natural Resources. Forest Management Division.

I am ewedaIIy thankful for strong bonds of kinship and friendship. capable of
weathering the extended periods of separation necessitated by the demands of this
work. To new friends and old, and especially to my family, i extend my heartfelt thanks
for your valued support, which was generously offered in many ways.

TABLE OF CONTENTS

LIST OF TABLES ................................................................................. viii
LIST OF FIGURES ............................................................................... ix
LIST OF ABBREVIATIONS ................................................................... xii
CHAPTER 1.0 INTRODUCTION ......................................................... 1
1.1 Overview .............................................................................. 1
1 .2 Research Questions .............................................................. 4
CHAPTER 2.0 BACKGROUND .......................................................... 7
2.1 Gypsy Moth Defoliation .......................................................... 7

2.2 Satellite Imagery .................................................................... 15

2.2.1 Spectral Responses ........................................................ 17

2.2.2 AVHRR Characteristics .................................................. 18

2.3 Classification Techniques ....................................................... 24

2.3.1 Maximum Likelihood .................................................. 27

2.3.2 Artificial Neural Networks ........................................... 28

CHAPTER 3.0 DATA ......................................................................... 35

3.1 Study Area ........................................................................... 37

3.2 AVHRR Imagery ................................................................... 47

3.3 Land Cover ........................................................................... 53

3.4 Defoliated Areas .................................................................... 54

CHAPTER 4.0 METHODS .................................................................... 58

4.1 Data Processing ................................................................... 59

4.1.1 Preprocessing ............................................................. 59

4.1.1.1 Raster AVHRR Imagery and Land Cover Data..... 59

4.1.1.2 Defoliated Areas ................................................ 82

4.1.1.3 Sampling ........................................................... 65

4.1.1.4 Roscommon County Scene Extraction ................. 69

4.1.2 Maximum Likelihood Classification .............................. 70

4.1.3 Artificial Neural Network Classification ......................... 73

4.2 Accuracy Assessment ............................................................ 78

CHAPTER 5.0 RESULTS, DISCUSSION AND CONCLUSIONS ............. 81

5.1 Results ............................................................................... 81

5.1.1 Lower Peninsula Classification ................................... 82

5.1.2 Roscommon County Classification .............................. 84

5.2 Answers to Research Questions and Discussion ...................... 88

5.3 Conclusions and Recommendations for Future Research .......... 105

vi

APPENDIX
A Selected World Wide Web Resources ................................................. 112

REFERENCES CITED .......................................................................... 1 16

vii

Table 4.1 .
Table 4.2.
Table 4.3.

Table 5.1 .

Table 5.2.
Table 5.3.

Table 5.4.

Table 5.5.

Table 5.6.

Table 5.7.

Table 5.8.

LIST OF TABLES

Computed linear transformation coefficients of the back transfon'nation....61

Lower Peninsula NDVI summary statistics .......................................... 72
Roscommon County NDVI summary statistics ..................................... 72
Error matrix for maximum likelihood classification of Lower Peninsula

data ........................................................................................ 82
Error matrix for neural network classification of Lower Peninsula data ....... 83

Error matrix for maximum likelihood classification of Roscommon
County data .............................................................................. 85

Error matrix for neural network classification of Roscommon County
data ........................................................................................ 87

Lower Peninsula and Roscommon County defoliation by actual percent
defoliated quartiles .................................................................... 93

Roscommon County maximum likelihood results for training sample
range and for category outside training range .................................. 98

Roscommon County maximum likelihood defoliated accuracy and class
homogeneity ............................................................................. 99

Areal extent and correct/Incorrect classification ratio for key values of
percent defoliation .................................................................... 104

viii

LIST OF FIGURES

WW
figure 1.1 Comparison of pattern recognition and 'inferencing capabilities for
selected artificial intelligence techniques ................................... 3
CW
Figure 2.1 Gypsy moth larva .................................................................... 8
figure 2.2 Global distribution of the gypsy moth........ .................................. 8
figure 2.3 Etienne Leopold Trouvelot ........................................................ 9
Figure 2.4 1990 distribution of the gypsy moth in North America ..................... 11
figure 2.5 25-year projection of gypsy moth spreading .................................. 12
figure 2.6 Annual acreage of Michigan gypsy moth defoliation ....................... 13
figure 2.7 Annual number of counties with Michigan gypsy moth defoliation ...... 14
figure 2.8 NOAA Polar Orbiter TiROS-N .................................................... 16
figure 2.9 2,399 kilometer swath of AVHRR coverage over North America ........ 20
figure 2.10 Artificial neural network architecture employed in this research ....... 29
mm
Figure 3.1 Roscommon County in the State of Michigan ................................ 38
figure 3.2 Monme temperature profile for Houghton Lake, Michigan .............. 40
figure 3.3 Monme heating degree day profile for Houghton Lake, Michigan ...... 41
figure 3.4 Monthly cooling degree day profile for Houghton Lake, Michigan ....... 42
figure 3.5 Monme precipitation profile for Houghton Lake, Michigan ................ 42
figure 3.6 Monthly snowfall profile for Houghton Lake, Michigan ..................... 43

Figure 3.7 1993 timberland ownership in the Northern Lower Peninsula Unit..... 45

figure 3.8 1980 and 1993 areal extent of major foresttypes in the Northern

Lower Peninsula Unit ........................................................... 47
figure 3.9 Lower Peninsula pro-defoliation NDVI ........................................ 50
figure 3.10 Lower Peninsula peak defoliation NDVI ...................................... 50
figure 3.11 Lower Peninsula post defoliation NDVI ...................................... 51
figure 3.12 Roscommon County pre-defcliaticn NDVI ................................... 51
figure 3.13 Roscommon County peak defoliation NDVI ................................ 52
figure 3.14 Roscommon County post defoliation NDVI ................................. 52
figure 3.15 Lower Peninsula forest mask ................................................. 54
figure 3.16 Michigan counties with observed defoliation in 1993 ..................... 56
W90:
figure 4.1 Control points used in resampling ............................................... 62
figure 4.2 Reference sketch map for Roscommon County defoliation cluster ..... 64
figure 4.3 Complete Lower Peninsula reference data ................................... 68
figure 4.4 Sampled Lower Peninsula data .................................................. 68
figure 4.5 Roscommon County complete and sampled reference data ............. 69

WW

figure 5.1 Frequency distribution of proportion defoliated for all Lower
Peninsula defoliated pixels .................................................... 94

figure 5.2 Frequency distribution of proportion defoliated for all Roscommon
County defoliated pixels ........................................................ 95

figure 5.3 Spatial distribution of proportion defoliated for all Roscommon
County defoliated pixels ........................................................ 95

Figure 5.4 Frequency distribution of proportion defoliated for all Rommmon
County defoliated pixels correctly classed as defoliated by ML
classifier ............................................................................ 96

figure 5.5 Subpixel percent of range versus percent defoliation for Roscommon
County ML classification ....................................................... 99

figure 5.6 Correct/incorrect ratio versus percent defoliation for Roscommon
County ML classification ....................................................... 102

figure 5.7 Correct/Incorrect ratio versus percent defoliation for Lower Peninsula
and Roscommon County ML and ANN classifications ................. 104

xi

Al

ANN
AVHRR
80

BP
000
CDA
CIR
CRS
DNR
FAQ
FIA
FTP
GAC
GIS
GLERL
HRV
LAC
MCC
MDNR
ML
MSS
MVC
NCAAS
NCDC
NDVI
NESDIS
NIR
NOAA
NODC
NPV
PAN
PCC
RMSE
SPOT
TD

TM

LIST OF ABBREVIATIONS

Artificial Intelligence

Artificial Neural Network

Advanced Very High Resolution Radiometer
Bhattachanya (or Jeffries-Matusita) Distance
Backpropagation

Charge-Coupled Device

Command and Data Acquisition

Color infrared

Center for Remote Sensing

Department of Natural Resources
Frequently Asked Questions

Forest Inventory and Analysis

file Transfer Protocol

Global Area Coverage

Geographic information System

Great Lakes Environmental Research Laboratory
High Resolution Visible

Local Area Coverage

Midwestern Climate Center

Michigan Department of Natural Resources
Maximum Likelihood

Multispectral Scanner

Maximum Value Compositing

NOAA CoastWatch Active Access System
National Climatic Data Center

Normalized Difference Vegetation index
National Environmental Satellite, Data and Information Service
Near Infrared

National Oceanic and Atmospheric Administration
National Oceanographic Data Center
Neucleopoiyhederosis virsus

Panchromatic

Percent Correctly Classified

Root Mean Squared Error

Systeme Pour I'Observation de la Terra
Transformed Divergence

Thematic Mapper

xii

Chapter 1
INTRODUCTION

1.0 introduction

Through the use of emerging technologies, this work attempted to improve upon
previous efforts to efficiently identify widely-distributed areas of undesirable deciduous
forest leaf-loss, or defoliation, caused by insect feeding. Successful implementation of
the approach proposed in this effort would enable land managers, entomologists,
ecologists, geographers and landowners to rapidly identify areas of leaf-loss as well as
confirm the locations of unaffected regions, reduce associated monitoring costs, assist
in targeting available control funds to those areas mom affected, and base decisions on

known quantitative levels of analytical accuracy.

1.1 Overview

The Forest Management Division of the Department of Natural Resources (MDNR) of
the State of Michigan, represented by Frank Sapio, and the Center for Remote Sensing
(CRS) at Michigan State University. represented by Dr. David P. Lusch, attempted to
develop a rapid and relatively inexpensive technique to delineate the extent of
deciduous forest defoliation caused by gypsy moths (Lymantn’a dispar L.) in Michigan
(Lusch and Sapio, 1994). Current detection methods are labor intensive, time-

consuming, costly, and based on outdated equipment that is nearing the end of its life

expectancy. Specific areas of defoliation are normally first visually identified on airborne
imagery acquired for the entire region, and then hand-sketched onto topographic maps
which are used for later analysis.

Funding was provided to CRS by MDNR to acquire and prepare suitable data for
analysis. The purpose was to investigate the viability of using Earth-orbiting satellite
data in order to detect portions of forested areas in the State of Michigan which had
been defoliated by gypsy moths during one growing season. This investigation by
MDNR/CR8 utilized Advanced Very High Resolution Radiometer (AVHRR) satellite
imagery in conjunction with maps produced by MDNR of areas known to be defoliated,
which were based on the interpretation of airborne video reconnaissance. Initial
attempts did not yield conclusive results, and no further work was planned by the

partnership on this problem (Lusch and Sapio. 1994).

In this report, such Earth-orbiting satellite data and data captured by sensors in aircraft
are referred to as ‘remoteiy sensed,’ and the monitoring and data acquisition activities
are described as ‘remote sensing.’ This terminology is consistent with the usage by
practitioners in the field, though it is truly nonspecific and could equally apply to other
techniques, for example various surface and subsurface geophysical exploration
procedures.

While working at the Center for Remote Sensing at Michigan State University, one of my
assigned tasks involved the preprocessing of spatial data stored in a geographic
information system (GIS). These data were being prepared as part of the

2

aforementioned project undertaken by CRS on behalf of MDNR. At about the same time
period, I read an article by McConnack and Day (1993) ccnceming the application of
various artificial intelligence (Al) techniques to Earth investigations. Of particular interest
was the mention made of the superior pattern recognition capabilities of the artificial
neural network (ANN) approach. The authors noted that the ability of human intelligence
to infer and to recognize patterns is the essence of its. power. A plot of pattern
recognition capability versus inferencing ability was presented for expert systems, fuzzy
logic, genetic algorithms and neural networks, included here as figure 1.1. Genetic
algorithms and especially neural networks were shown to have superior pattern
recognition capabilities and weaker inferencing abilities. Fuzzy logic and especially
expert systems excel at inferencing, but do not periorrn as well at pattern recognition.
This article, and my general interest in Al applications, led to further reading on the

subject, in particular concerning artificial neural networks.

 

 

 

 

 

Neural 1

Network
Pattern 1 Genetic 1
Recognition Algorithm
Capability

Fuzzy Expert
Logic System

Inferenclng Ability

 

 

Figure 1.1 Compariaon of pattern recognition and inferencing capabllltiea for eelected
artificial intelligence techniques. After McConnack and Day (1993).

3

The research described in this thesis evolved from my desire to apply both conventional
remote sensing image classification procedures and the less commonly utilized ANN
approach to a multitemporal change detection problem, in this case involving a forest
management issue affecting many states in the northeastem United States. It was
hoped that insight would be gained concerning the relative strengths and weaknesses of
each method, and also that a practical surveillance solution could be developed for the
detection of affected areas. I further wished to gain familiarity with methods that might
enable the integration of multisensor ‘remoteiy sensed’ data (not limited in this case only
to satellite and aircraft observations) in order to develop characteristic feature

classification signatures.

1.2 Research Questions

The work conducted in this study quantitatively evaluated the ability of two supervised
classifiers to discriminate areas of forest defoliation caused by gypsy moths through the
integration of multitemporal and multispectral AVHRR imagery and ancillary spatial in-
formation stored within a geographic inforrnaticn system. The problem was one involving
two opposite forest states: leaf-on (nondefoliated) or leaf-off (defoliated), with the
possible complicating existence of varying combinations of these two conditions.
Supervised classification employs known outcomes for several representative input

vectors to develop classification solutions for vectors with initially unknown class

assignments.

As explained in greater detail in section 2.2, the selection of the sensor involved a
compromise between temporal and spatial resolution. with questions primarily arising
concerning the adequacy of the spatial component in measuring the targeted features,
given the scale of the defoliation process and the extent during the monitoring period. in
addition, it was necessary for the temporal component to enable image acquisition at
sufficient frequency to adequately sample the process being monitored.

Satellite imagery has been frequently used in land cover studies. The maximum
likelihood (ML) classification technique is one of the most commonly used methods
(Foody, McCulloch and Yates, 1995a; Jensen, 1996). ANNs have been successfully
utilized in several image classification tasks, and have often, but not always, produced
higher levels of classification accuracy than conventional approaches. It was anticipated
that the neural network method adopted for use in this research might yield higher
classification accuracies than the conventional algorithm, by exploiting the widely cited
robust nature of the technique in processing the temporally-composited, band-ratioed,

coarse-resolution satellite data available for this study.

Specific questions addressed through this research include the following:

1) Can the maximum likelihood supervised classification technique be successfully
applied to the problem of detecting defoliated areas?

2) Can the ANN classification technique be successfully applied to the problem of

detecting defoliated areas?

3) How do the accuracy levels of the maximum likelihood and the ANN techniques

compare in this application, given the final parameters utilized?

4) Are the classification accuracy levels for defoliated and nondefoliated pixels similar?

5) At a subpixel level of detail, what insight may be gained from an analysis of pixel
homogeneity?

Chapter 2
BACKGROUND

2.0 Background

To explore the possibility of differentiating between gypsy moth defoliated and nondefoli-
ated forest land cover in Michigan, multichannel and band-ratioed multitemporal AVHRR
satellite data were utilized. Analysis involved the use of commme available remote
sensing, GIS, statistical and ANN software, with data spanning a period of
approximately 3.5 months during the 1993 growing season. Forest defoliation by gypsy
moths, satellite imagery characteristics and classification techniques are summarized in

this Chapter.

2.1 Gypsy Moth Defoliation

Defoliation can cause tree mortality and dramatically alter the character of the affected
woodlot or forest. With some 300 million acres of susceptible forest cover throughout the
United States, the impact on ecological balances, local economies. wildlife habitat. water
quality and flow, and available shade can be substantial. The natural range of the gypsy
moth is in Europe, North Africa and Asia, where it has a widespread distribution.
Introduction to the United States occurred during the 19th century. The larva is shown in
figure 2.1, while a map of the global distribution is provided In Figure 2.2.

 

Figure 2.1 Gypsy moth larva.

Source: USDA Forest Service, Northeastern Forest Experiment Station.
Gypsy Moth in North America http://gypsy.fsl.wvnet.edu:80/gmoth/ 3/31/98

 

 

 

 

 

Figure 2.2 Global distribution of the gypsy moth.
The natural range is shown as dark gray, the region of recent introduction is shown in light gray.
Source: USDA Forest Service. Northeastern Forest Experiment Station.
Modified from Gypsy Moth in the World http: //gypsy. fsl. wvnn 4 'g .... ..... ‘html 3/22/98

 

 

 

The gypsy moth was accidentally introduced to the United States in 1868 or 1869 by
French artist and amateur scientist Etienne Leopold Trouvelot (1827-1895). after he
brought specimens to Massachusetts in 1868 for cross-breeding purposes. His
unsuccessful experiment was an attempt to develop a new strain of ‘silkworm.’ He is
reported to have realized the significance of the initial release from his home in Medford.
near Boston. and advised authorities, but no action was taken at that time (USDA Forest

Service, 1998a). Refer to figure 2.3 for an image of Trouvelot.

s. “\-
X '2‘.
\\.
\
\-

\\\\.\.\\
R‘t"
‘. \xk
\'.\\ “\
“\‘.\‘.

\:::.I\'I.\\t&

u.“

aims... ““-
\
.\
\\ I. \ ‘-
\K‘.

x
4.\
\I.

\ \\\1§
\. ‘Y‘
s. s \
~ are.
I, \;~%
'- \ \\\
\ \\ z‘.

\
\\

Ix;

\ ‘24- “R

i \
x \\\
\k\

\
\\\\
\\

\\.\\.
\
A\\\

-.-.- “R
-.k “shun.

,./
' r

\\

'- $
\ \-
\).t'_\\
i \

 

Figure 2.3 Etienne Leopold Trouvelot.
The individual responsible for introducing the gypsy moth to the North American continent.
Source: USDA Forest Service, Northeastern Forest Experiment Station.
E. Leopold Trouvelot, Perpetraior of Our Problem
http'J/gypsy.fsl.wvnet.edu/gmoth/worldlworld.html 3/22/98
Courtesy of Mary Lea Shane Archives, Lick Observatory. University of California. Santa Cruz.

The gypsy moth is considered to be the primary defoliator of hardwoods in the
northeastem portion of the United States, where its distribution is widespread, as shown
in figure 2.4. A map of the projected spread of gypsy moths over the next 25 years is
presented in Figure 2.5, and this extension is expected to cover a wide band adjoining
the outer boundary of the present distribution. A thorough study assessing United States
forest susceptibility to gypsy moth defoliation was completed by Liebhold, Gottschalk

and Mason (1998), and an excerpt from their summary‘document is included below.

Over-all forest susceptibility was quantified using five measures: average basal area
per acre of preferred species, proportion of stand basal area in preferred species,
proportion of land area covered by susceptible stands (> 20% of basal area in
preferred species), proportion of land area covered by highly susceptible stands (>
50% of basal area in preferred species), and proportion of land area covered by
extremely susceptible stands (> 80% of basal area in preferred species) for each
county. All three [sic] measures yielded maps that indicated similar distributions of
susceptible forests over the cotenninous US. The areas with the highest
concentration of susceptible forests were in the central and southem Appalachians,
the Cumberland Plateau, the Ozark Mountains, and the northwestern lake states.
Comparison of these maps with the know [sic] distribution of individual susceptible
species indicates that oaks are the major component of susceptible forests in the
Appalachian, Cumberland, and Ozark areas but quaking aspen is the major
susceptible species in the northwestem lake states (Liebhold, Gottschalk and
Mason, 1998).

The ten top-rated species, In terms of susceptibility to gypsy moth defoliation, were
determined to be (highest susceptibility first) white oak (Quercus alba), sweetgum
(Liquidambar styreclflua), quaking aspen (Populus tremuloides), northern red oak
(Quercus rubra). black oak (Quercus velutina). chestnut oak (Quercus prinus), post oak
(Quercus stellata). water oak (Quercus nigra), paper birch (Betuia papyrr’fera) and
southern red oak (Quercus falcata). These ten species have a combined total basal

area of 7.714 billion square feet per acre in the 48 cotenninous states in the United

States. A complete list of the top 20 species identified, as well as maps depicting the

10

distribution, in basal square feet per acre. for all of these are provided by Liebhold,
Gottschalk and Mason (1998). In the same work. these authors concluded that, on a
national level, 'the impacts of defoliation and costs of gypsy moth management are

likely to increase in the future.‘

 

 

\

.
G
d eh".

I... _ .- .
setteifi’tfi“
“'e'?t‘ , v.5. r
2:33W" AF-anr ."'

n‘.‘ ..

 

 

 

 

figure 2.4 1990 Distribution of the gypsy moth In North America.
Counties with gypsy moth presence are shaded. The black circle has a radius 01850 kilometers
and is centered at the point where the gypsy moth was introduced to the United States in 1868 or
1869. This spreading has thus occurred over a period of approximately 122 years.
Source: USDA Forest Service, Northeastern Forest Experiment Station
Gypsy Moth Spread in North America http://gypsy.fsl.wvnet.edu/gmothlspread/spread.html
3/22/98

 

 

 

 

 

 

Figure 2.5 25-year projection of gypsy moth spreading.
Source: USDA Forest Service. Northeastern Forest Experiment Station
Gypsy Moth Spread in North America http://gypsy.fsl.wvnet.edu/gmothlspread/spread.html
3/22/98

The distribution of gypsy moths within the State of Michigan now includes every county
in the Lower Peninsula and also portions of the Upper Peninsula. Joria, Aheam, and
Connor (1991) state that there are approximately 2.02 million hectares (just under five
million acres) of forested land in the State of Michigan that are susceptible to defoliation
by the gypsy moth, though USDA forest inventory figures suggest the actual extent is
somewhat larger. as discussed in section 3.1 of this report (Leatherberry, 1993). Oak
and aspen are particularly affected, though over 600 species of trees are subject to leaf

consumption.

As of the late 1980’s, the projected extent of defoliation was expected to reach hundreds
of thousands of hectares per year (Montgomery, 1988). Defoliation in Michigan has
been inconsistent, both in the areal extent and the number of counties affected, as
shown in Figures 2.6 and 2.7, respectively. The peak over the last 19 years occurred in
1992. There has been decreased extent since then, though there was a large increase
in the number of counties affected in 1997 compared to 1996. Although recent Michigan
defoliation acreage has been less than that of the 1992 peak, a new strain of gypsy
moth was discovered in North Carolina that has the potential for dissemination
throughout the Midwest, suggesting the need for enhanced detection capabilities and
continued monitoring efforts. The 1993 study period was the third highest in both areal
extent of defoliation and number of counties affected for the period of record from 1979

to 1997.

 

 

 

 

 

800000
709000.-
600000--
gm...
5400000-
gawooo-
32.0.0...
100000-
c-
YODI'

 

 

 

Figure 2.6 Annual acreage of Michigan gypsy moth defoliation.
Data source: Sapio (1998b).

 

 

Bearer

Counties with Defoliation

 

 

 

 

 

figure 2.7 Annual number of counties with Michigan gypsy moth defoliation.
Data source: Sapio (1998b).

Decreased activity levels in recent years are attributed to the viral disease
Neucleopoiyhederosis virsus (NPV), to the introduced Japanese Entomophaga
maimaiga fungus, to parasitoids which feed in the body of the various life stages of the
gypsy moth, to predators that consume the gypsy moth, as well as to very cold winters
that result in high mortality of the egg masses (lngeils, 1995; Michigan State University
Extension, 1998; Sapio, 1998a). The effectiveness of the fungus is dependent upon the
abundance of moisture and high humidity during early June, with high mortality

occurring when these conditions exist (lngeils, 1998).

Federal and State government gypsy moth suppression costs in Michigan over the
period 1986 to 1995 totalled approximately US$18 million, with a combined (Federal and
State) peak annual expense of US$36 million in 1993 (USDA Forest Service, 1998b). In

addition, some communities initiated locally-funded control programs.

Given the magnitude and impact of the problem, it is desirable to closely monitor the
extent of defoliation. The MDNR has traditionally obtained information through the use
of a color infrared (CIR) video monitoring system employed in air flights over the State.
The typical method of mapping widespread occurrence is by sketching the affected
areas onto topographic maps based on low altitude aerial reconnaissance (Ahem at al.,
1991; Talerlco, Walker and Skratt 1978). Limitations of this approach were summarized
by Joria, Ahearn and Conner (1991 ), and include the subjective and variable skill of
interpreters, the changing condition of monitored areas during the length of time re-
quired to obtain a complete data set over the millions of hectares involved, and the brief
two to three week temporal window of peak defoliation during which field data must be
acquired. This typically occurs during the first three weeks in June (Muchoney and
Haack, 1994). The major source of interpretation error has been tree mortality caused

by defoliation in previous years.

2.2 Satellite Imagery

Satellite imagery has been utilized in an attempt to overcome the limitations mentioned
in the preceding section concerning the traditional method of identifying widespread
areas of leaf loss (Dottavic and WIIliams, 1983). Landsat Thematic Mapper (TM),
Multispectral Scanner (M88), and Systeme Pour l'Observation de la Terra -1 (SPOT-1)
scenes have enabled identification of defoliated regions, as summarized by Muchoney
and Rack (1994). AVHRR data acquired by National Oceanic and Atmospheric
Administration (NOAA) satellites have also been proposed for use In this application
(Lusch and Sapio, 1994). AVHRR acquisition commenced with the launch of TlROS-N

on October 13, 1978, and continues to the present with other operational satellites in the

15

numbered NOAA series (Figure 2.8). These systems are passive, in that, in the visible
and near-infrared (NIR) bands, they rely on reflected solar energy for illumination of the

target, rather than actively emitting and then receiving a pulse as in radar systems.

 

Figure 2.8 NOAA Polar Orbiter TIROS-N.
Precursor to the NOAA AVHRR satellite series.
Source: NOAA http://psbsgi1.nesdis.noaa.gov:8080./NSORS/ML/nsorsz.html 3/16/97

Earth-orbiting sensors have been widely used in land cover studies because of the large
surface areas that can be monitored at the same time, the ability to use several
electromagnetic wavelength bands in classifications, and the ability offered to revisit
locations for temporal comparisons. Sensors acquire information for a narrow range, or

band, of wavelengths, rather than for one specific value.

The general concept of spectral response, which is the fundamental issue concerning
target separability, and information concerning AVHRR data, which are the type used in

this analysis, are briefly reviewed in the next two subsections.

2.2.1 Spectral Responses

Surface materials have differing spectral responses, or degrees of reflectivity, at
different wavelengths, and it is this characteristic that enables multi-band satellite sensor
data to be used to discriminate between features. For example, there is comparatively
little difference between vegetation and sandy loam soil In the visible (blue. green and
red) wavelengths, which are approximately between 0.4 and 0.7 micrometers. However,
there is a broad separation between the percent reflectance curves for vegetation and
sandy loam soil in the near-infrared (NIR) wavelengths between 0.7 and 1.3
micrometers. At these NIR wavelengths, vegetation has approximately 2.5 times the
response and in excess of 60% reflectance. Consequently, these materials would be
easily separable using a NIR wavelength band. Similany, different types of vegetation
have more widely varying levels of reflectance in the near-infrared, while in the visible
wavelengths the values are much less separable. The variation in average percent
reflectance over the visible and NIR wavelengths for certain vegetation classes is
distinctive. The reflectance separation between classes enables the discrimination be-
tween various types of vegetation. For example, deciduous trees characteristically pro-
vide almost twice the average percent reflectance of coniferous trees (Avery and Berlin,

1992).

In this manner, reflectance information acquired by the AVHRR sensor was used in this
research in an attempt to discriminate between defoliated and nondefoliated deciduous
trees, since the spectral response is different for healthy versus damaged leaves, for
leaves at various stages of senescence, and for leaf-on versus leaf-off conditions (Avery

and Berlin, 1992). Note, however, that removal of the canopy foliage allows undergrowth

I7

and soil to contribute more to the reflected signals captured by the satellite sensor. All
incoming solar radiation is either absorbed, transmitted or reflected. At visible
wavelengths (AVHRR channel 1 is red) where high absorptance is typical for vegetation,
plant pigmentation controls the spectral response. At longer wavelengths in the near-
infrared (AVHRR channel 2), absorptance is much reduced and transmittance and
reflectance assume much greater and almost equal importance. The response from
vegetation in the near-infrared is controlled by the internal leaf structures, with the
palisade mesophyll of the dorsiventral structure serving as light ‘plpes’ to the inner air
voids where refraction occurs prior to emittance. Younger leaves have fewer air voids
and a corresponding lesser level of reflectance in the near-infrared than leaves of
greater maturity. The compact leaf structure has fewer air voids and lesser levels of
reflectance. Finally, it should be noted that the NiR reflectance from a canopy of leaves
is greater than from just one leaf due to partial transmittance through leaves to others,

and then reflectance from all (Lusch, 1993).

2.2.2 AVHRR Characteristics

AVHRR data are widely used for environmental monitoring, their primary advantage
being useful spectral bands with high temporal resolution and single scene coverage of
large areas. The most significant shortcomings of this sensor are '(1) the lack of
radiometric calibration coefficients, and (2) the high slant angle which distorts radio-
metric readings at extreme off-nadir angles' (Ehrlich, Estes and Singh, 1994, p. 147).

The lack of atmospheric corrections is another serious shortfall.

18

The AVHRR sensor acquires information for a 2,399 kilometer swath over the Earth’s
surface with each pass, and every day 14.1 satellite orbits are completed around the
Earth at the nominal 833 kilometer altitude. Data are acquired at a 1.1 kilometer ground
resolution at nadir. A diagram of the swath coverage is depicted in Figure 2.9. The
period of each orbit is 102 minutes, and there is a 255° distance between orbits. The
orbit is inclined at 98.9° and there is a 155° scan angle from nadir (Ullesand and Kiefer,

1994).

Full-resolution Local Area Coverage (LAC), and sampled Global Area Coverage (GAC)
data at a nominal spatial resolution of 4 kilometers, are recorded onboard as acquired
and then later transmitted to a ground station during overpass. Command and Data
Acquisition (CDA) stations at Gilmore Creek, Alaska and at Wallops island, Virginia
receive sensor data and use communications satellites to rebroadcast the information to
the Suitland, Maryland facilities of the NOAA National Environmental Satellite, Data and
lnforrnation Service (NESDIS; USGS EROS Data Center, 1998).

AVHRR data are available for the period from 1979 to the present. Current imagery for
Michigan is readily available through the NOAA CoastWatch program, as summarized
by Leshkevich, Schwab and Muhr (1993). In contrast to other imagery, the AVHRR data
were available at no cost for this study, a significant factor in this investigation and also
worthy of consideration in the potential application. The specific products used in this

research are discussed in section 3.2.

19

 

figure 2.9 2,399 kilometer swath of AVHRR coverage over North America.
Source: USGS EROS Data Center. lrnage modified from
httpzlledcwwwcr.usgs.gov/glis/graphicslguidelavhrr/figure1 .gif 3/19/98

Data products of the CoastWatch program are available for four windows covering the
Great Lakes region. The imagery used in this study was from the Michigan-Huron
window, a Local Area Coverage (LAC) which encompasses 40.76 to 46.73 degrees
north latitude and 79.78 to 88.05 degrees west longitude, with a mid-latitude pixel size of
1.3 kilometers. 'Actual resolution is determined by dcose, where d is the spatial
resolution at the equator and 6 is the Iatitude...level 1b data are mapped to a Mercator

projection and resampled to a 512 by 512 pixel grid“ (Leshkevich, Schwab and Muhr,

20

1993, p. 372). A subset was utilized which more closely corresponded to the specific

area of interest.

The AVHRR sensors acquire information in five radiometric bands. For this work, data
from channels 1 and 2, representing 0.58 to 0.68 micrometers in the visible red, and
0.725 to 1.0 micrometers in the reflected infrared, respectively, were combined by CRS

in the following manner to yield the Normalized Difference Vegetation Index (NDVI):

NDVI = W
Band 2 (near IR) + Band 1 (visible red)

NDVI is a commonly used indicator of both the presence as well as the condition of
crops and natural vegetation, and values typically fall between approximately 0.1 to 0.6
for vegetation. Damaged leaves progress through characteristic changes in spectral re-
sponse, and a defoliated tree provides a different response than does a healthy tree in a
normal leaf-on state, as outlined in section 2.2.1. NDVI is thus considered an ap-
propriate indicator of defoliation (Avery and Berlin, 1992). In addition, NDVI provides
some compensation for variations in illumination, the slope of the ground surface,
aspect, and other characteristics affecting the analysis (Llllesand and Kiefer, 1994).
Goward et al. (1991) review AVHRR NDVI measurements and related considerations.
Cihiar, St.-Laurent and Dyer (1991) found that AVHRR NDVI growing season

trajectories offer useful insight on vegetation development.

The use of ratios in classification has often provided better results than when the
channel data were presented separately. since the relationship was provided to the

classifier explicitly and thus this information did not have to be discovered during

21

processing (Bailey and Thompson, 1990; Hammerstrom, 1993b). Green and Cosentino

(1996) also cite as benefits the reduction of topographic effects and the augmentation of
spectral trends.

As the biomass increases through leaf generation during a typical growing season, and
as the maturing leaves develop more air cavities in the mesophyll, the NDVI ratio
increases. Later in the season, chloroplast deterioration occurs and eventually the cell
structure collapses. Visible leaf colors shift from green to yellow to red-brown, and NDVI
decreases. By the end of the growing season, the NIR reflectance has decreased and
the reflectance at red wavelengths has increased, thus reducing NDVI (Avery and
Berlin, 1992; Jensen, 1996). The reduction in biomass at the peak defoliation stage,
caused by gypsy moth feeding, was expected to result in a reduction in measured NDVI.
It was anticipated that these changes would be evident in the observed spectral-
temporal signature, and that discrimination of affected areas would thus be possible. As
described in detail in Chapters 3 and 4, three time periods were selected at which

spectral responses were measured and then employed in data classification.

The 1,300-meter-square spatial resolution of AVHRR imagery in Michigan is profoundly
inferior to either 30-meter Landsat TM data or 20-meter SPOT multispectral High
Resolution Visible (HRV) scenes, but satellites canylng AVHRR sensors offer a daily
revisit frequency for a given location. This is vastly superior to the 16-day revisit period
available for Landsat or the variable 5-day repeat coverage cycle for SPOT sensors in
the mid-latitudes. For Landsat, this time refers to acquisition by the same satellite, and
more frequent revisits are possible using the second sensor platform. For SPOT, which

22

has an orbital revisit period of 26 days, the more frequent acquisition is possible using

the pointable optics for off-nadir imaging.

In addition to the increased temporal resolution, Lillesand and Kiefer (1994) note that
the coarser spatial resolution of the AVHRR sensor yields an increased radiometric
resolution due to the greater amount of total energy arriving at the detectors of the
charge-coupled device (000), which results in higher signal levels and an increased
ability to detect very small differences in incoming radiation. Roesch, Van Deusen and
Zhu (1995) reported, however, that scale effects of the coarse AVHRR spatial resolution
contributed to a bias in the estimation of percent forest cover when target features are
substantially smaller than the pixel size. Small regions of all classes may go undetected,

and smoothing of edges would occur.

The high temporal resolution of AVHRR data provides a solution to the problem of ob-
taining synoptic information over large areas with easily manageable data volumes.
This enables maximum value compositing (MVC) techniques to be applied in order to
generate cloud-free scenes which depict the entire study area at approximately the
same time. This is necessary because there is only a 5 to 10 percent probability of any
overflight of the State of Michigan having minimal (0 to 10 percent) cloud cover (Lusch,
1994).

With MVC, a balance must be established between the conflicting demands of using a
compositing period long enough to minimize cloud presence yet short enough to provide
the temporal resolution required to isolate targeted changes in the landscape. It should

be noted, however, that multitemporal AVHRR MVC may be sensitive to ”residual cloud

23

contamination, [variations in] preceding climatic events, temporal and spatial scales of
analyses, and the composition and spatial organization of the study area“ (Allen, Bara
and Walsh, 1994, p. 52).

2.3 Classification Techniques

Remote sensing image classification represents an attempt to categorize surface
features based on data acquired by sensors positioned at some distance above the

surface.

Land cover classification of remotely sensed data usually leaves inaccuracies in the final
product. These errors can be quantified when a land cover map exists that is considered
to be completely correct, however such reference (or ‘truth’) information is not always
readily available. There may be several concurrent contributing factors to the error.
Temporal variation in the characteristics of the targets, spatial variation in responses
from a given target, atmospheric influences upon the radiafion received at the sensor,
off-nadir viewing angles and sun-target-sensor geometry, variance in sensor radiometric
response and inadequate or variable detector calibration can all affect the data quality
and the resulting classification. Research continues to address the normalization of
these factors which degrade the accuracy of the final classification (Lillesand and Kiefer,

1994).

Classification mchnlques provide varying levels of ability to accommodate such variance
in remotely sensed imagery. The commonly applied Gaussian ML classifier considers

both the variance and covariance of the class spectral vectors when assigning an

unknown pixel to a category, but a normal distribution of the training data is assumed to
exist. Violation of this assumption may result in an unacceptably high level of

mlsclassification (Lillesand and Kiefer, 1994; Paola and Schowengerdt, 1995a).

Chen and Takagi (1993) investigated the use of an ANN to classify NOAA AVHRR data,
and pointed out that it is unrealistic to expect traditional statistical classifiers to perform
well when classes overlap and when information is distributed across many often very
highly correlated bands along only a limited number of major components. Here at al.
(1994, p. 102) note that 'the most severe limitation arises when conventional classifiers
implicitly presume a particular input metric with which to measure likeness in the
multidimensional space of the feature vector.’ Further, these authors indicate that the
absence of these assumptions in ANN classifiers which, in training, adaptively determine
'a metric specific to the problem,‘ allows the method to avoid 'this prejudicing of the

classification process” (Hara etal.,1994, p. 102).

in a detailed comparison of BP ANN and ML classifiers in an urban application, Paola
and Schowengerdt (1995a, p. 995) concluded that, due to the nonparametric nature of
the ANN approach, it “is more robust to training site selection.” This is contrasted with
the maximum likelihood method, which they observe to be “sensitive to the purity of the
class signatures...and performs poorly if they are not pure” (Paola and Schowengerdt,
1995a, p. 995). They also note that
the lack of a statistical assumption in the neural network allows it to produce more
mutually exclusive class discriminant functions, resulting in more detail at the pixel
level and possibly at subpixel levels. The maximum-likelihood classifier, because it
depends on to [sic] the assumption of normal (and potentially high variance)

distributions, can mask this fine detail in the classification (Paola and Schowengerdt,
1995a, p. 994).

Brown, Lusch and Duda (1998, p. 248) confirmed this greater degree of ‘smoothing’
produced by the ML classifier, noting that the ANN technique was better able to identify
the classes which were more difficult to discern, and that “ANNs can be much more

sensitive to slight variations in the inputs.”

Paola and Schowengerdt (1995a) further elaborate by stating that
The neural network, because of its fully-interconnected nature, produces a
fundamentally different type of classification. The warped appearance of the network
class output functions is due to the competitive learning that takes place during
backpropagation training. As one class is trained it suppresses the output value of all
the other classes and, in this way, the classes compete with one another for
‘territory" in the feature space. Thus, the network produces “probability densities“

that have sharper inter-class transitions and do not overlap nearly as much as those
of maximum likelihood (Paola and Schowengerdt,1995a, p. 989).

Because of these reports from current researchers, and other similar observations which
have appeared in the literature, this work employed the ANN classifier in addition to the
conventional maximum likelihood approach. The intent in doing so was to determine
whether the performance characteristics of either would yield greater classification
accuracy, given the data described in Chapter 3 and the methods described in Chapter
4. Openshaw (1994) has argued that this is in fact the only valid jusfification for

experimenting with what he refers to as ‘neuroclasslfication.’

The approach used in this research is consistent with the integrative role suggested by
McKeown (1987) and Foody (1995) for the use of artificial intelligence techniques in
combining a variety of remote sensing, GIS and other input information to produce a

greater understanding of land-related issues.

26

2.3.1 Maximum Ukellhood

The maximum likelihood classification method is a supervised approach, in that known
per-pixel class information is used to develop a characteristic signature, based on the
corresponding input band data, that is then used to classify unknown pixels. The
technique is commme employed in remote sensing applications and has been
thoroughly described in related literature (Dude and Hart, 1973; Jensen, 1996; Lillesand
and Kiefer, 1994; Schowengerdt, 1983). Maximum likelihood is considered the
statistically optimal classifer for such applications (Bischof, Schneider and Pinz, 1992;

Wan, 1990).

Both the variance and the covariance of the input band spectral information are utilized
as the probability, or likelihood, of class membership is assessed for each unknown
pixel to be classed and for each possible assignment category. Probability density
functions are generated for each category and are used in determining which category
has the maximum probability, or is most likely, to be correct for the input vector of each
unknown pixel. Assignment is made to the category of highest probability, with the
underlying assumption that a Gaussian, or normal, distribution of single-band class
training data exists. During implementation, the minimum number of pixels required for
training purposes varies from (ion) to (100-n), where n is the number of the spectral
bands used as input variables (Clark Labs, 1995; Lillesand and Kiefer, 1994).

Details ccnceming the maximum likelihood methods used in this research may be found
in section 4.1.2.

27

2.3.2 Artificial Neural Networks

Within the field commonly described as artificial intelligence, many neural network algo-
rithms have been developed, many versions of software implementations are available
and research interest persists (Sarle, 1998a). Comprehensive summaries of
developments in the field have been prepared by several researchers in recent years
(Hush and Home, 1993; Lippman, 1987; Rumelhart and McClelland, 1986).
Introductory information on the implementation of ANNs has been provided by

Hammerstrom (1993a and 1993b).

ANNs have been successfully utilized in numerous classification applications involving
remote sensing data (Paola and Schowengerdt, 1993; 1995b). The method with the
widest reported application in image classification in recent years is backpropagation
(BP), with several variants available. These methods employ many processing

iterations to arrive at a solution.

Feedforward BP ANNs offer the ability to perform supervised classification. They consist
of a variable number of layered, weighted, and interconnected processing elements
which accept input data, sum and transfer these data using pro-defined functions,
compare output values with reference data, and readjust the weights on the connections
based on a specified ieaming rule. The process is repeated until the network has been
trained, by adjusting the weights, to classify training data such that the output meets a

target error threshold. Variations of feedforward BP algorithms primarily involve

28

modification of the transfer functions and ieaming rules. A diagram depicting the ANN

architecture used in this research is presented in Figure 2.10.

The activation, or transfer, functions are utilized within the processing elements of the
hidden layer to introduce nonlinearity into the ANN computations (Sarle, 1998b). One
advantage of the method is that highly irregular spikes or blades in the error surface are
less likely to occur, thus facilitating the desired descent to the global minimum error.

Input Layer
(3 variables)

Hidden Layer
(3 nodes)

 

Output Layer
(2 classes)

 

[om Classification]

 

Figure 2.10 Artificial neural network architecture employed In this research.
Full interconnection exists between elements in adjacent layers.

The input data and the desired classification output are used in determining the
structure of a BP ANN. The input layer shown in figure 2.10 simply accepts the input
data vectors, and is determined by the selection of variables considered cmcial to the

task, with one ‘node’ or element for each variable. The internal stnicture must be

29

established by trial and error to optimize the classification accuracy (Paola and
Schowengerdt, 1997). The input layer distributes the data to the ‘hidden’ or internal
nodes where a weighted sum of the inputs is calculated, passed through a transfer
function, and then distributed to the elements in the output layer. The output layer
consists of one element for each of the desired categories in the final solution. Here, a
weighted sum of the incoming values is computed, passed through a transfer function,
and a result is generated. in supervised training, the actual result is compared with the
desired result, and the derivatives of the error with respect to the weights are sent back
(back-propagated) to the hidden nodes where the individual contribution to the error is
calculated at the nodes, and weights are then changed in accordance with a pre-
established criterion (Hammerstrom, 1993b). The values of the weights ‘leamed’ in this
manner thus facilitate discrimination between classes. Network architecture optimization
can be approached by either ‘pruning’ an overly large structure or by gradually building a

minimal structure.

One might visualize a bowl-shaped object in three dimensions to represent a highly
simplified error surface which is to be traversed by gradient descent from maximum
error somewhere near the upper lip of the bowl to minimum error at the base. The
starting point is established by the weight initialization. The ‘learning rate’ would control
the vertical descent from initial maximum error to the final minimum error, while the
‘momentum’ term would affect horizontal oscillatory movement at roughly similar

magnitudes of error (vertical axis) along the surface of the bowl.

To elaborate, the ieaming parameter in the BP approach “is analogous to the distance

along the error surface traveled in a single epoch, so that the smaller the learning rate,

30

the smaller the changes in the weights of the network at each epoch“ (Skidmore et al.,
1997, p. 508). One version of a feedforward BP ANN includes a momentum term that is
“added to the learning rate to incorporate the previous changes in weight with the
current direction of movement in the weight space. In other words, inclusion of the
momentum term avoids wild swings across the error surface, while allowing the system

to learn faster" (Skidmore etal., 1997, p. 508).

Momentum controls the magnitude of weight change between epochs, and thus affects
the speed of learning, but not the accuracy. An attempt is made during the processing
routine to limit the number of iterations utilized such that there is a sufficient number for
the network to adaptively ‘learn’ the discrimination functions through the weight update
procedure, while at the same time retaining the ability to generalize on nontraining data.
Because the learning rate is analogous to the incremental decrease in error, small

values for this parameter will require a larger number of epochs, or iterations.

in addition to ieaming rate and momentum, the rate of convergence to minimum error is
also affected by the order in which the training vectors are presented, the complexity of
the regions for which decision boundaries are generated and the network structure
(Huang and Lippman, 1988).

One of the key advantages of ANNs is their distribution-free nature. Prior knowledge
about a class statistical distribution is not required. Further, the ANN automatically
determines the relative reliability and weight associated with each data source input to
the network (Benediktsson, Swain and Ersoy, 1990; Here at al., 1994). Hepner et al.

(1990) found ANNs to be somewhat tolerant of noisy data and also of missing data.

31

Lippman (1987) noted that a multilayer system can generate arbitrarily complex decision
boundaries and can discriminate overlapping classes. The degree of class boundary
complexity. or the ability of the ANN to adapt to the n-dimensional distribution of input
data, is limited by the number of processing elements, or nodes, In the network. A
neural net classifier implemented by Huang and Lippman (1988) was found to be robust
because it discounted outliers. Klimasauskas (1992, p. 51) noted that the effective com-
plexity of BP ANNs varies depending on the complexity of the input data, and that 'the
complexity of the functional relationships and the stability of the resulting solutions have
given neural networks an edge in modeling dynamic systems.“ It has also been reported
that an ANN approach was successful in training on only mixed pixels when attempting

to classify woody vegetation in southern England (Foschi and Smith, 1997).

Perhaps of greatest value, neural networks are better able than conventional classifiers
to detect subtle pattems in the data being classed. Also, the greater flexibility afforded
by their nonlinear nature, the ability to generalize and develop the required model, and
the ability to accept various types of data without violating assumptions are positive
characteristics (Skidmore er al., 1997). ANNs require many training samples
(Benediktsson, Swain and Ersoy, 1990), however, and these must be of good quality
(Skidmore at al. 1997). in addition, the number of training samples selected from each
class establishes the a priori probabilities, which cannot be avoided (Decatur 1989;
Foody etal. 1995b). The varying response to weight initialization can yield substantial
differences in classifications, in parficular when a limited number of processing elements
is specified in the hidden layer. The use of a greater number of nodes can provide more
consistent results (Paola and Schowengerdt, 1997). While Skidmore er al. (1997) offer

the opinion that BP ANNs will not become the dominant method used in remote sensing

32

and GIS classification due to the limitah'ons, they suggest that a hybrid approach
Involving expert systems and rule-based techniques offers promise.

While ML classification is considered the statistically optimal approach (Bischof,
Schneider and Pinz, 1992; Wan, 1990), ANNs have nonetheless outperformed the ML
method. Bischof, Schneider and Pinz (1992) attribute this to the often only minor devia-
tion of the image spectral information from the normal distribution required by ML theory.
Paola and Schowengerdt (1995a) concur, and add that the competitive nature of ANN
training, combined with the ability to consider the training data for all classes while
generating class boundaries, results in more precisely defined boundaries with less
overlap than provided by ML.

The neural network, because of its fully-interconnected nature, produces a
fundamentally different type of classification. The warped appearance of the
network class output functions is due to the competitive learning that takes place
during backpropagation training. As one class is trained it suppresses the output
value of all the other classes, and, in this way, the classes compete with one
another for “territory' in the feature space. Thus, the network produces
“probability densities” that have sharper inter-class transitions and do not overlap
nearly as much as those of maximum-likelihood. These “densities” are warped
around one another and fit together in a jigsaw fashion to form the final decision
regions of the classifier (Paola and Schowengerdt, 1995a, p. 989).

Paola and Schowengerdt (1995a) suggest that the ANN makes use of the training data

with greater efficiency than the ML classifier.

The maximum-likelihood classifier models each class by its own statistics,
independent of all other classes. The neural network, on the other hand,
considers all class training data when defining the “probability density“ for each
class. This is accomplished through the competitive training property discussed
previously. When all the other classes are being trained, the output for a given
class is set low (e.g., 0.1), thus, in effect, providing information about where that
class does not exist in the feature space. This is used in addition to the training
data for the class itself, which indicates where the class does exist. Thus the
network is using the training data more efficiently than maximum-likelihood to

33

determine the natural distributions of the classes and is more robust when faced

with a decrease in training samples (Paola and Schowengerdt, 1995a, p. 991 ).
Previous work with artificial neural networks has been adversely affected by a lack of
reliable procedures for developing optimum network architectures, training data and
sampling considerations, difficulty in reaching the global minima of the error curve with
gradient descent, variable output depending on weight initialization, and slow processing
speeds. Recent research such as that by Foody, McCullough and Yates (1995a;
1995b), Blamire (1996) and, In particular, Paola and Schowengerdt (1993, 1995a,
1995b, 1997) addresses these issues. Genetic algorithms may be used for neural

network parameter optimization (Murray, 1994).

Details concerning the specific BP ANN methods employed in this work are presented in

section 4.1.3.

Chapter 3
DATA

3.0 Data

The digital data used in this work consisted of AVHRR scenes, raster GIS land cover
information, and reference defoliation polygons for the State of Michigan. Those portions
of the dataset within the Lower Peninsula and within the Roscommon County political
boundaries were used for classification. The direct source of all of this information was
the Center For Remote Sensing at Michigan State University, where earlier work had
been conducted using the data.

For the 1993 growing season, multidate AVHRR NDVI composites consisting of two-
band ratios for each of three separate narrow temporal windows, as well as GIS data on
susceptibility to defoliation, were used in classification. Areas of observed defoliation,
hand-sketched by MDNR from airborne video imagery and later digitized, constituted the
reference (‘truth’) data, and was used for comparison against the classification output.
During the 1993 study period, forest defoliation caused by gypsy moths in Michigan was
primarily observed in the northern Lower Peninsula.

The year 1993 was selected for the work reported here because during that time period
the extent of defoliation in Michigan was among the highest levels for the 19 years
included in the period of record (1979 through 1997), thus enabling larger sample sizes

for training purposes. The period of study was thus representative of one of the times of

35

greatest concern to land managers in Michigan, but was not the somewhat anomalous
peak year. The year 1993 immediately followed a season of much greater defoliation,
and preceded several years of lower extent. In addition to these considerations, the
reference information was already available in digital form for 1993, and selecting this
period minimized the amount of preliminary data preparation required before
classification tests could be conducted.

The three temporal windows isolated from the AVHRR data available for the year 1993
were chosen in an attempt to coincide with, firstly, the period of peak defoliation, and
then for times during the growing season immediately prior to and just after this period.
it was anticipated that the spectral response would differ depending on the varying
abundance and condition of deciduous tree leaves present at these three stages. As
noted by Muchoney and Rack (1994), the time of peak defoliation typically occurs from
June 1-21, but there is some latitude possible for later satellite observations. The
composite AVHRR data selected by MDNR/CR8 to correspond to the period of peak

defoliation was acquired in mid-July, somewhat later than the actual peak.

Because the coarse resolution of the AVHRR sensor resulted in many pixels having
both defoliation and nondefoliation content, it was anticipated that a higher proportion of
purer (more thoroughly defoliated) pixels could be obtained for this exploratory
investigation by using a subset of the Michigan data. Roscommon County was selected,
in addition to the Lower Peninsula scenes, for the classification experiments conducted
in this work because the area within his political boundaries of the county included a
large portion of one of the most dense concentrations of defoliated forest in the State. It

was expected that this region would thus yield many of the purer pixels necessary for

36

classifier training, and would also provide some less homogenous pixels for comparative
analysis (Jensen, 1996). in addition, there were relatively few nonforest pixels in this

locale which would be excluded from classification.

3.1 Study Area

As just mentioned, classification was performed on both Lower Peninsula and
Roscommon County data which was extracted from the complete Michigan data set. As
highlighted in figure 3.1, Roscommon County lies in the north-central portion of the
Lower Peninsula. General information regarding the topography, geomorphology, soils,
geology, climate, and forest cover of the State will now be reviewed, with an emphasis
on that region in the north-central Lower Peninsula observed to be defoliated during the
year 1993, and in particular the Roscommon County test area. The primary area of
investigation is thus in the northern portion of the Lower Peninsula, but observed
defoliation extends also to the southern portion of Michigan and to a county in the Upper

Peninsula.

Michigan’s two peninsulas have a varied topography, ranging from the more rugged
Keeweenaw highland in the western portion of the Upper Peninsula to the level lake.
border plains adjacent to lakes Michigan, Huron, Superior, Erie, and St. Clair. Other
physiographic regions represented within the State include plains, roiling plains, hill-
lands, upland plains, and hilly uplands. Most landforms are the result of continental
glaciation in the region, with the last glacial recession occurring approximately 9,500

years before the present. Roscommon County falls in the Houghton-Higgins upland

37

 

 

 

 

 

 

Figure 3.1 Roscommon County in the State of Michigan.
Classification trials used Lower Peninsula and Roscommon County data.

plain and the West Branch hill-land found in the north-central Lower Peninsula, in a
landscape ecosystem identified by Albert (1995) as the Grayling outwash plain.
Landforms in this county include moraines, till plains, and outwash plains. Soil
associations in Roscommon County are predominantly spodosols and histosols.
Spodosols are commonly gray colored soils in which leaching has migrated minerals
and humus from the upper A horizon to the lower B horizon. Histosols have a high
organic content, have undecayed vegetation present, and are typically found in wet
areas (Sommers, 1977; Sommers et al., 1984). There are two large lakes in west-

central and northwest Roscommon County, named Houghton and Higgins, with areas of

38

8,113 hectares (20,032 acres) and 3,888 hectares (9.600 acres), respectively (Sommers
et al., 1984).

Elevations In the State range from 570 feet abover‘nean sea level along Lake Erie in the
southeastern Lower Peninsula to 1,980 feet in Baraga County in the western Upper
Peninsula. Elevations in Roscommon County range frbrn approximately 1,000 to 1,400
feet above mean sea level. The youngest bedrock is in the center of the Lower
Peninsula, with older fonnatlons radiating out from this area. The youngest formations
present in the center of the Lower Peninsula are Jurassic of the Mesozoic Era. The
oldest underlying bedrock in the State includes Lower, Middle and Upper Precambrian in
the western part of the Upper Peninsula. Predominantly Mississippian and
Pennsylvanian rocks of the Mesozoic Era, which ranges from approximately 290 to 360
million years before the present, underlie Roscommon County. These sedimentary
rocks are overlain by 200 to 800 feet of glacial drift (Sommers, 1977; Sommers at al.,

1984).

Michigan, located in the middle latitudes and having prevailing air movement from the
west to the east, has a continental climate. Mean annual temperatures throughout the
state range from less than 40 to in excess of 48 degrees Fahrenheit, with warmer
readings in the south and along Lake Michigan, and the coldest temperatures in the
western Upper Peninsula. The hills and valleys present the opportunity for relatively
colder localized nighttime temperatures in surface depressions. The annual precipitation
ranges from less than 28 to more than 36 inches. Highest rainfall occurs in the
southwest portion of the Lower Peninsula and in the northwest and north-central areas
of the Upper Peninsula (Sommers, 1977).

39

Average monme temperature, heating, cooling, precipitation and snowfall information is
available for the period from 1961 through 1990 for the Houghton Lake Station located
in west-central Roscommon County. The NOAA National Climatic Data Center (NCDC)
Midwestern Climate Center (MCC) identification number for this station is 203932. The
satellite data used in this research, described in detail in section 3.2, covered the
periods mid-May to mid-June, mid-July, and late August. The 1961-1990 average
monthly high and average monthly low Fahrenheit temperatures for these three periods,
as indicated Figure 3.2, are approximately 70/45, 82/54, and 75/50 degrees,

respectively.

 

I‘DUGIfl‘IllLLAKEfiflSW, MI (203932)
flown“: 181-199 Extrema: 1940-195

,17 :r ij _Q. ‘ .. 1 ‘ _ T r , ft . .._-,~
.- .

 

0;“
83:88:38.:
r F—f 1 ° '*

I

l

Trimaran-‘0 (F)
l
\
S
\
l
’1‘
/

 

 

 

 

1. # -—_____r. If”- -'
O .. - - ..
-1. .
-2. i ,
.3. r g ..
Jan Han flay Jul Sep Nov
Fob fipr Jun Flog Oct Dec
North of Your
- " ‘ ~ -* Jfi'r‘ ' .. ._ -- ‘ - -, fliduskrn
I: Record High “Haw High—~80. Lou 2-- Record Lou] at...“
. ...... . JV. .. ., . - . 7 7 fl - ~. mm

 

 

 

Figure 3.2 Monthly temperature profile for Houghton Lake, Michigan.

Source: Midwestern Climate Center, Champaign, Illinois, a cooperative program of the NOAA
NCDC and the Illinois State Water Survey Division of the lllinois DNR.
htth/mccsws.uiuc.edu/Summary/Gif/203932.TempM.gif 3/23/98

40

Average monthly heating degree days at Houghton Lake, summarized in Figure 3.3,
range from above 1400 in January to a low of about 50 in July. Average cooling degree
days are graphically presented in Figure 3.4, and occur from April through October, with

a peak in July.

Figure 3.5 shows that the average monthly maximum precipitation at Houghton Lake
ranges from a low of just over one inch in February to a high of over 3.5 inches in
September. The months from April through November have higher average monthly

precipitation levels than other months.

Snowfall occurs at the Houghton Lake station from October through April, with an
average monthly peak of over 12 inches in January and a low of less than one inch in

October. Figure 3.6 depicts this distribution.

 

HOUGHI‘0N_LAKE_6_HSW, MI (203932)
averages: 131-1990

 

 

 

Figure 3.3 Monthly heating degree day profile for Houghton Lake, Michigan.

Source: Midwestern Climate Center, Champaign, Illinois, a cooperative program of the NOAA
NCDC and the Illinois State Water Survey Division of the lllinois DNR.

hflp llmm‘ sws nim‘ J ,.'3if/203932. HddM. gif 3/23/98

41

 

iniKi‘flUUAKEfiJSH, ”1 (203932)
Mega: 131-18!)

 

 

 

Figure 3.4 Monthly cooling degree day profile for Houghton Lake, Michigan.

Source: Midwestern Climate Center, Champaign, Illinois, a cooperative program of the NOAA
NCDC and the lllinois State Water Survey Division of the Illinois DNFi.

http-Ilm ewe nine ‘ "‘ ,.'3if/203932.CddM.gif 3/23/98

 

 

I'DUGI‘II'W_LAKE_6_WSW, HI (203932)
nuances: 1361-1990

Precipitatim (in. )

 

Minnow
cum
Center

 

 

 

Figure 3.5 Monthly precipitation profile for Houghton Lake, Michigan.

Source: Midwestern Climate Center. Champaign, Illinois, a cooperative program of the NOAA
NCDC and the lllinois State Water Survey Division of the Illinois DNR.

http Ilmm «we nim‘ ‘ -_ ,,'3if/203932. PrecM. gif 3/23/98

 

42

 

I'DWJAKLSJSW, ”I (203932)
Moses: 1331-1390

    

 

Ii the: torn
Clinch
Center

 

Source: Midwestern Climate Center, Champaign, Illinois, a cooperative program of the NOAA
NCDC and the lllinois State Water Survey Division of the Illinois DNR.
http'llmm sws uirm J "‘ ,.'3if/203932.Sn0wM.gif 3/25/98

 

Michigan is in a forested region broadly characterized by Barbour and Billings (1988) as
containing mixed deciduous species, yet there is considerable compositional variety
present. They show forest associations in the southern portion of the Lower Peninsula to
be Beech-Maple, with the remainder of the State falling in their Conifer-Deciduous
association class. The climate found throughout Michigan favors forest growth, yet there
is a north-south and a more subtle east-west transition in typical species, as noted by

Harman (1984):

Moving south to north through the Lower Peninsula, one notices an increase in the
number of evergreen species, particularly in the vicinity of Clare. These species
(white spruce, balsam fir, jack pine) are boreal in distribution but extend south into
northern Michigan on upland sites where they are usually only minor associates of
other native hardwoods (though jack pine occurs extensively on sandy uplands).
Apparently sensitive to warm, dry summers and neutral or basic soils, and unable to
compete with more temperate-climate associates, they are limited primarily to the

43

northern Lower Peninsula - north of a line from Bay City to Muskegon (the tension
zone) - and to the Upper Peninsula, where cooler summers and lower evaporation
are typical along with the acidic, sandy soils on which these trees are most
competitive. Warmer summer weather with higher evaporation rates combined with
lea favorable soils seem to restrict their occurrence in southern Michigan (Harman,
1984. p.87).
The Roscommon County study area in the north-central Lower Peninsula is north of the
‘tension zone’ referred to by Harman, and is also north of Clare. The north-south
species transition in the Lower Peninsula is noted by Harman to be the apparent result
of climatic influence, but the narrowness of the transition zone may be related to the
soils distribution. In the northern region of the Lower Peninsula, coarse, sandy soils are

more prevalent, whereas to the south of this tension zone he notes heavier glacial till

soils are more prevalent.

Harman goes on to state that

A less conspicuous forest gradient exists from the central to the western portions of
the Lower Peninsula. Woodland composition in the central Lower Peninsula is a
mosaic of beech-maple mixed with some oak-hickory communities that results from
the patchwork of different soils. Westward through the state, however, the more
drought-sensitive beech-sugar maple forests become more common, and north of
about Muskegon, they are mixed increasingly with yellow birch and hemlock. ...This
westward increase of beech, sugar maple, and other drought-sensitive associates
may be related to climatic modifications by Lake Michigan (Harman 1984, p.87).

1993 data for the northern Lower Peninsula show that 52% of the timberland was
privately owned by individuals and farmers (Figure 3.7). Of the balance, 40% was
publicly held, and 8% was owned by corporations (Leatherberry, 1993).

 

Other Miscellaneous
public corporations
1 it 8%

Individuals

&farmers
52%

    

National
Forest
1 2S

 

 

 

Figure 3.7 1993 timberland ownership in the Northern Lower Peninsula Unit.
Source: MDNR Forest Management Division.
http://www.dnr.state.mi.us/www/lmd/fia/nlp/fianlp.html 3/23/98

In the Roscommon County area for which classification experimentation was conducted,
presettlement forests were primarily Pine-Oak (Sommers etal., 1984). In the northern
Lower Pensinsula, "The deciduous members of the oak-hickory species become fewer,
and upland pin oak may be the dominant species. Dry sites that have been heavily
disturbed are dominated by red maple, bigtooth and quaking aspen, and some pines
(both red and jack), but recurrent fires eradicated the pines on some sites. Extensive
jack-pine plains characterize sandy sites that have been burned less often or those sites

that are a bit less xeric (excessively dry)” (Herman, 1984, p.88).

In 1993, the fifth Michigan forest inventory was completed. This wodr was identified by
the USDA as the 1993 Michigan Forest Inventory and Analysis (FIA), and details were
included for the Northern Lower Peninsula Unit (Leatherberry, 1993) which

encompasses the majority of all observed defoliation. The Northern Lower Peninsula

45

Unit includes those Michigan counties on the tier from Oceana through Bay and all other

counties to the north of this tier in the Lower Peninsula.

The inventory was carried out under the responsibility of the North Central Forest
Experiment Station of the USDA Forest Service. Figure 3.8 indicates that in the
Northem Lower Peninsula Unit where gypsy moth defoliation was concentrated, Maple-
Birch, Aspen, and Oak-Hickory were the most abundant forest types, with a combined
total areal extent in that year of approximately 4.7 million acres, and the addition of Elm-
Ash-Soft Maple and other hardwoods increases this forested total to approximately 5.5
million acres. This 5.5 million acres represents 73.50% of all forested land in the Unit.
Other forest types specifically mentioned in a USDA summary include Jack Pine, Red
Pine, and other softwoods, as shown in Figure 3.8. The inventory indicated that 65.36%
of the 11.3451 million acres of land in the USDA Northem Lower Peninsula Unit were

forested when the 1993 Michigan FlA was conducted (Leatherberry, 1993).

46

 

 

01h" 80300068

Red pine
E} 1993
Jack pine I “30
g Other hardwoods
! EIm-ssh-soit maple
o
it

OCR-NONI! ”'5‘“‘3'1'3‘3:I'2-2-:-:-:-:-

eeeeeeeeeeeeeeeeeee
eeeeeeeeeeeeeeeeee
eeeeeeeeeeeeeeeeee

eeeeeeeeeeeeeeeeeee
eeeeeeeeeeeeeeeeeee

eeeeeeeeeeeeeeeeeee
eeeeeeeeeeeeeeeeeee

Aspen

eeeeeeeeeeeeeeeeeeeeeeeeeeeee
eeeeeeeeeeeeeeeeeeeeeeeeeeee

eeeeeeeeeeeeeeeeeeeeeeeeeeeee
eeeeeeeeeeeeeeeeeeeeeeeeeeee

Maple-birch

 

 

 

 

     

o (is {a {.5 5.0 is
flea (rt-on acres)

 

 

Figure3.8 1980and1993arealextsntofmalorforssttypsslnthsuorthernLower

Peninsula Unit.
Source: MDNR Forest Management Division.
http://www.dnr.state.mi.us/www/fmd/fia/nlplfianlp.html 3/23/98

3.2 AVHRR imagery

The most recent two months of AVHRR data for the study area at the time of collection
by CBS were available through the NOAA Great Lakes Environmental Research Labo-
ratory (GLERL), located in Ann Arbor, Michigan. Archived data were available through
the NOAA National Oceanographic Data Center (NODC). The activities of the current
NOAA CoastWatch Active Access System (NCAAS) include archiving and distributing
CoastWatch AVHRR data. Detailed information concerning these facilities and data is

available via the World Wide Web as indicated in Appendix A.

47

 

As part of previous work at the CBS, and as described by Liebens (1994), ten 1993
AVHFiFl scenes were downloaded from NOAA in Ann Arbor (then-recent dates) and
Washington (then-older dates). Selection was based on completeness of geographic
coverage and degree of cloud cover. The images were selected with the goal of
providing cloud-free composites for three times: pre-defoliation, concurrent with peak
defoliation, and post defoliation. The pre-defollation composite image approximately
covered the period mid-May to mid-June, and consisted of the following five julian dates:
136, 146, 157, 162, and 163. The peak defoliation composite was produced from
scenes on julian dates 193, 196, and 197, which corresponded to mid-July. The post
defoliation time composite for late August consisted of scenes 243 and 244. image
acquisition time was 1300 hours for julian date 157, 2100 hours for julian date 193, and
2000 hours for all other dates, producing sun-angle differences for julian date 157 and
193 compared with the other eight scenes. All images were geometrically registered to
scene 157 using a linear method. Registration error was stated to be on 'the order of
0.1 - 0.2 pixels' (Liebens, 1994). Error in the range from 0.5 to 1.0 pixels or less is
normally considered acceptable (T ownshend at al., 1992). The NDVI was calculated for
these images, and rescaled to values of 0 through 255. The cloud cover was removed
through maximum value compositing, in which the highest pixel values of the scenes in
each narrow temporal window were used to generate the final composite. This elimi-
nated the clouds, which had very low NDVI values.

Although not specifically stated in the CRS literature regarding this topic, it appears as
though both NOAA-10 and NOAA-11 satellites were used, based on the overpass times

just mentioned. NOAA-10 was launched September 17, 1986 and remains in service.

48

NOAA-11 was launched September 24, 1988 and went out of service on September 13,
1994 (USGS EROS Data Center, 1998). Both were classed as third—generation

operational meteorological satellites.

Uncorrected, die combined usage of NOAA-10 and NOAA-11 data would introduce
variations due to differing sensor response characteristics. Also, the difference in
overpass times results in different sun angles, which in turn contribute to differences in
scene brightness. The different satellite orbital alignments introduce varying ground

resolution (coarser farther off-nadir) and varying lengths of atmospheric travel paths.

Figure 3.9, Figure 3.10 and Figure 3.11 provide a visual impression of the Lower
Peninsula pre-, peak and post defoliation NDVI values, respectively. Though the three
images appear somewhat similar, generally moderate values are observed in the earlier
pre-defoliation stage, while higher NDVI ratios existed later in the growing season during
the period of peak defoliation, with slightly declining values at post defoliation. This
relationship is evident at a greater level of detail in the Roscommon County subsets
presented as Figure 3.12, Figure 3.13 and Figure 3.14 for pre-, peak and post
defoliation phases, respectively. Summary statistics for the defoliated and nondefoliated

pixels at the two regional levels are presented later in section 4.1.2.

49

 

Figure 3.9 Lower Peninsula pre-defollatlon NDVI.
Dark green is high, yellow is moderate, and dark red is low. Mid-May to mid-June, 1993.

u: 1...... “M '“L-L... ..4 ._.A¢Lu

Figure 3.10 Lower Peninsula peak defoliation NDVI.
Dark green is high, yellow is moderate, and dark red is low. Mid-July, 1993.

50

 

Figure 3.11 Lower Peninsula post defoliation NDVI.
Dark green is high, yellow is moderate, and dark red is low. Late August, 1993.

 

Figure 3.12 Roscommon County pro-defoliation NDVI.
Dark green is high. yellow is mid-range, dark red is low. Mid-May to mid-June. 1993.

51

 

Figure 3.13 Roscommon County peak defoliation NDVI.
Dark green is high, yellow is mid-range, dark red is low. Mid-July, 1993.

 

Figure 3.14 Roscommon County post defoliation. NDVI.
Dark green is high, yellow is mid-range, dark red is low. Late August, 1993.

52

3.3 Land Cover

Since the forest defoliation which is to be detected must occur in forested areas, the
classification procedure would benefit from the knowledge of whether the area
represented by a pixel is or is not forest land cover. More specifically, it would be
beneficial to exclude from consideration those pixels in nonforested portions of the study

The CBS at Michigan State University compiled a land cover map of the state based on
the visual interpretation of custom-enhanced Landsat imagery. Both M88 and TM data
from the period 1979 to 1986 were interpreted. These scenes were optically enlarged to
match stable-base copies of USGS 19 x 29 quadrangies (1 :250,000-scale). A 40 hectare
minimum mapping size was employed. The hand-drawn land cover polygons were
digitized and then rasterlzed at a cell size of 333.33 meters. Subsequently, these data
were aggregated using a 3 x 3 model filter to the final one kilometer x one kilometer cell
size using custom software which resolved ties by deferring to the lower numbered of
the two land cover categories (i.e. 1-Urban would take precedence over 2-Agriculture, et

cetera).

In previous work by staff of the CRS, this map was registered to the AVHRR scenes and
forested portions of the region were isolated and used to create a binary image, shown
in Figure 3.15. In preprocessing operations within a geographic information system, this
binary image was used to exclude all nonforested pixels from the data later presented
for classification.

53

 

 

 

 

 

Figure 3.15 Lower Peninsula forest mask.
Forested areas are indicated in black.

3.4 Defoliated Areas

Reference information, consisting of observed areas of gypsy moth defoliation in 34
Michigan counties, was acquired by MDNR in 1993 using an airbome video system.
Defoliated areas were sketched onto 1:100,000 topographic maps by MDNR following
interpretation. These defoliated areas were later digitized for each individual county and

were available for this study as vector CMAP files.

The most seriously affected portion of Michigan is the northern half of the Lower
Peninsula, as shown in Figure 3.16, which depicts counties affected during the study
period. Due to a very small occurrence on Bois Blanc Island, Mackinac County in the
Upper Peninsula is indicated as a defoliated area in Figure 3.16. However, due to the
small size of the affected area, this was lost in the vector to raster conversion. There
were no occurrences noted for other portions of the Upper Peninsula in the ‘truth’ data
provided by MDNR. Thus, all known defoliated areas in the raster data, which served as
the basis for processing operations, were in the Lower Peninsula of Michigan. Small
areas were defoliated in Oakland and St.Clair counties in southeastern Michigan. The
extent along Lake Michigan was as far south as Muskegon County, but no defoliation
was noted in Antrim, Charlevoix and Leelenau counties, which are located in the
northwestern Lower Peninsula. Defoliation presence was expressed in two ranges: 75 to
100% and 50 to 75%, with digital information identified to indicate either of these two

states, or the absence of defoliation.

As noted by Congalton (1991), the accuracy of the ‘truth’ data is often not known with
certainty. Such was the case in this investigation, although due care and standard
professional practices were exercised in the acquisition, interpretation, digitizing and
preprocessing of the information. As is customary, it was assumed that the reference
data were correct. However, due to the time lapse between acquisition and use, field
verification was not possible in connection with this research. Recall that defoliation

totals may be subject to misinterpreted mortality from previous years.

55

 

 

 

 

 

Figure 3.16 Michigan Counties With Observed Defoliation in 1993.
Based on MDNR interpretations of airborne video. Black areas indicate counties where defoliation
was observed. The Roscommon County classification test area, which also experienced
defoliation, is shown in white. Nondefoliated counties are depicted as green.

A statistical summary generated by CMAP software for the statewide CMAP vector files
indicated that in this format there were a total of 41 1,763.2 acres within defoliation
polygons, with a mean area of 349.54 acres, yet with a standard deviation of 862.02
acres. At full conversion, this total acreage would fill 986 of the 1.3 kilometer-square
raster cells, given 417.6 acres per cell. The largest of the 1,178 polygons was 9,774.6
acres in size, which would cover 23.41 complete 1.3 kilometer-square AVHRR pixels.
Twenty-two ’slivers' less than 0.3 acre in size were deleted in preprocessing, so the

smallest vector defoliation polygon was 0.3 acres. in the original CMAP vector files,

56

areas interpreted to be 75 to 100% defoliated were classed as 1, and areas interpreted
to be 50 to 75% defoliated were classed as 2. There were a total of 1,942 of the 1.3
kilometer-square raster cells classed as defoliated in Michigan, which indicates 328,198
hectares (810,995 acres) defoliated. This exceeds the amount indicated by the vector
’truth’ data. Additional details ccnceming the difference in vector and raster defoliation

areas are provided in section 4.1.1.2.

57

Chapter 4
METHODS

4.0 Methods

Data preprocessing involved several steps utilizing a number of software packages to
prepare the digital files for input to classifiers. Following classification, common

measures of remote sensing accuracy assessment were calculated and compared.

The hardware and software utilized for this work are commercially available or are
freeware, and were provided by the Department of Geography and the Center for
Remote Sensing and GIS at Michigan State University, and by the Institute for Parallel
and Distributed High Performance Systems at the University of Stuttgart. This research
employed Sun Ultra 1 and Ultra 2 workstations and Pentium and 486-class personal
computers to run CMAP, lDRiSl, EXCEL, SYSTAT, ERDAS, PCl EASl/PACE.

FOXPRO, ARC/lNFO and SNNS software for the various tasks.

The general processing requirements were to merge previously digitized county vector
reference information into a statewide coverage, convert this to raster format, coregister
the result with three precomposited raster AVHRR NDVI files, sample the data, and
generate Lower Peninsula and single-county classifications of the data by the use of a
convenfional maximum likelihood classifier and also by employing a less conventional
artificial neural network approach. Accuracy was then assessed by generating an error,
or confusion, matrix, calculating percent correctly classified (FCC) and Kappa values,
and comparing these for the different classification methods and spatial extent. Two

measures of signature separability were used to determine the statistical similarity

58

between the defoliated and nondefoliated classifier training information. Finally, Kappa

values were compared with results achieved by other researchers.

4.1 Data Processing

Processing is logically divided into two phases, the preprocessing work and the actual
classification of the properly formatted, assembled, and sampled data. Michigan Lower
Peninsula data were initially prepared, and a Roscommon County subset was then
selected, or windowed, from these for additional analysis.

4.1.1 Preprocessing

The preprocessing operations converted both the raster and vector data into a form
suitable for classification, with varying amounts of preliminary work required on the
digital files received. These files included the AVHRR imagery, land cover information
and boundary files, and a map showing known areas of defoliation occurrence during

1993, the period under investigation.

4.1.1.1 Raster AVHRR imagery and Land Cover Data

Imagery was received from CBS as raster files in lDRlSl format containing NDVI values,
scaled from 0 through 255, as the pixel information content. Three temporally
composited scenes were provided, one each for the pro-defoliation, the peak defoliation,
and the post-defoliation stages. As detailed in section 3.2, these composite periods
correspond to mid-May to mid-June, mid-July and end-August, respectively. The method

59

of calculating these NDVI values was reviewed in section 2.2.2. The forest mask had
been applied to all of these scenes, thus excluding nonforefied areas from
consideration. The forested portion was approximately 38% of the entire Lower

Peninsula of Michigan, at 1.3 kilometer-square resolution.

Using six corresponding points from the AVHRR scenes and from a Michigan state and
county boundary file, which properly overlayed the defoliation file, the three AVHRR
scenes were resampled to the 0.325 kilometer-square preliminary spatial resolution of
the reference data, and later aggregated to 1.3 kilometer-square resolution. The
reference information was in the Michigan State Plane (Zone 2) coordinate system.

A linear (first order) mapping function was used to minimize the adverse effect of any
inadvertent poor choices in points and due to the limited number of readily identifiable
control points on the data received. Nearest neighbor resampling was performed, thus
placing the NDVI value of the closest, or nearest, cell into the new output cell. This
approach does not alter the data values, but only moves them to a new location. It is the

preferred method for resampling.

Following standard procedures for this type of operation, those correspondence points
having the highest residuals were omitted, while retaining a sufficient minimum number
distributed around the defoliated area. A distribution of six points encircling the primary
region of defoliation remained from a total of 19 points identified initially, as depicted in
Figure 4.1. The residuals for these six individual points ranged from a low of 0.20 to a
high of 0.94 pixels. Six points are twice the mathematical minimum requirement. The
conventional level of acceptable root mean square error (RMSE) for such an imagery

resampling procedure is one-half pixel (T ownshend et al., 1992). The final overall

RMSE, using only these six points and expressed in input image units, was 0.58, which
approximates the desired one—half pixel level. The first-order polynomial coefficients of
the back transformation, from new cell to old cell, are provided in Table 4.1, and are
used in the following equations:

’=bo+b,x+b,y
y’=bo+b,x+b,y

where x’ and y’ represent the original input, and x and y represent values in the rectified
image, and on, b, and b2 are the coefficients listed in Table 4.1.

Table4.1 Computed llneartransformation coefficientsofthebacktransformatlon
Truncated. Values were carried internally in the ldrisi software module to 20 significant figures.

Coefficient X Y
b0 ‘ - 321 .68107582 14308523385
b1 0.00023916 0.00000417

b2 - 0.00000238 0.00023674

61

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

l_

 

 

 

 

 

 

 

 

l | |

—-

 

 

 

 

 

Figure 4.1 Control points used in resampling.
Black control points enlarged for visibility. Areas of defoliation are red.

4.1.1.2 Defoliated Areas

Vector files representing the actual extent of gypsy moth defoliation within each affected
Michigan county, as well as a state and county boundary file, were received in CMAP
format. In CMAP, these individual county files were merged to provide a single file
providing comprehensive information. This file was edited, cleaned and built, and slivers
less than 0.3 acres were deleted. This reference information was then exported from
CMAP to lDRlSl vector format. This product was classed to distinguish between areas
of 75-100% defoliation (class 1), areas of 50-75% defoliation (class 2), and areas of no
observed defoliation. These regions were based on MDNR interpretations of airborne

video reconnaissance acquired in 1993, and hand-sketched onto 1:100,000 topographic

62

maps, which served as the basis for digitizing. Figure 4.2 is an example showing a large

defoliation occurrence in southeastern Roscommon County.

in lDRlSI, the vector reference files were converted to raster format, with the final cell
size chosen to coincide with the 1.3 kilometer-square mid-latitude pixel size of the
NOAA CoastWatch Michigan-Huron AVHRR scene that was the source of the spectral
information (Leshkevich, Schwab and Muhr, 1993). The vector-to-raster conversion first
utilized cells of size 0.325 kilometer by 0.325 kilometer (to produce a total of 16 pixels
per AVHRR cell), which were then aggregated to 1.3 kilometer-square cell size, with
each cell having an averaged value in the range 0 to 1 to represent the degree of
defoliation. Prior to aggregation, in order to specify the proportion a cell was defoliated,
the cells identified as 75-100% defoliated in the original vector data were assigned the
midpoint value of 0.875 as the proportion defoliated in the raster format. Similarly, the
pixels identified as 50—75% defoliated in the CMAP data were assigned the midpoint
value 0.625 as the proportion defoliated. Thus, no call in the aggregated raster data was
identified as 100% defoliated, and the maximum defoliation content was 0.875, whereas

the minimum was zero.
This approach provided subpixel reference information that was used in sampling and

later analysis. Since there are 169.0 hectares (417.6 acres) in a 1.3 by 1.3 kilometers

cell, the area of 1/16 of such a cell is 10.56 hectares (26.1 acres).

63

 

 

 

Figure 4.2 Reference sketch map for Roscommon County defoliation cluster.

Red is 75-100% and dark green is 50-75% defoliation. The southeast comer of Roscommon
County is at lower right, with Gladwin County below and Ogemaw County to the right. Base map
is USGS Houghton Lake 1:100,000 30 X 60 minute quadrangle.

The onscreen information within the IDRISI software and a related technical reference
manual both state that the ‘Polyras’ command used for the vector to raster conversion in
this work functions as described in the following quotation: “Conversion is done by
determining the intersection points of the polygon boundary and lines representing the

centers of image rows. A boundary causes a cell to be included in the polygon if its

position is more than halfway through that cell” (Eastman, 1992, p. 138-139). This
procedure generated considerably more raster defoliation than was found in the original
vector format. As the preceding quotation implies, the cause was apparently due to
many cases where only partial polygon content within a raster cell forced the raster pixel
to be classed as defoliated. The total area mapped with some defoliation in the original
vector CMAP files was 166.6346 hectares (411.7632 acres), in contrast to the 328,198
hectares (810,996 acres) represented by the 1,942 raster cells at 1.3 kilometer-square
spatial resolution which were identified as defoliated and used as reference information
in classifications during this research. However, many 1.3 kilometer-square pixels had
lower defoliation content than the vector polygons, which were estimated only at either

87.5% or 62.5%.

4.1.1.3 Sampling

In order to develop characteristic NDVI signatures for the three temporal windows
represented by the AVHRR composite images, samples of the defoliated and the
nondefoliated cells were generated.

The defoliated sample was based on within-pixel total defoliated area (defoliation being
the primary target to be identified), and all pixels having a defoliation content greater
than or equal to 50% were used for training. The need for training pixels possessing
greater purity was discussed in Chapter 3. The nondefoliated sample consisted of a
random selection of nondefoliated pixels falling near (up to 1.3 kilometers or one pixel
away from) all defoliated pixels with gmater than 0% defoliation (not only those with
greater than or equal to 50% defoliation). Nondefoliated pixels that were adjacent to

65

defoliated pixels were thus randomly selected. with an attempt made to provide a total
number of nondefoliated cells approximately equal to the number of sampled defoliated
pixels, in conformance with common practices (Blamire, 1996). This approach favors the
ANN method in that the borders of the decision boundary are highlighted, in contrast to
the standard approach in maximum likelihood training in which samples are taken from
within the class region (Bishof, Schneider and Pinz, 1992; An and Chung, 1994).
Congalton (1991) recommends a stratified random sampling approach when few

samples are extracted from each class.

The complete Lower Peninsula image depicting defoliated and nondefoliated cells is
shown in Figure 4.3. There were 1,942 defoliated and 21,295 nondefoliated pixels. This
image was sampled as described and the resulting product is depicted in Figure 4.4. In
the sampled Lower Peninsula image that was presented to the classifiers for training,
there were 372 defoliated pixels and 294 nondefoliated pixels. The Roscommon County
analysis was based on data from the full and sampled Lower Peninsula data to reflect
only the areas within the county. No separate resampling was done for Roscommon
County. The Roscommon County complete and sampled images are presented in
Figure 4.5. in the complete county image, there were 160 defoliated and 595
nondefoliated pixels. in the sampled Roscommon County image, there were 55
defoliated pixels and 23 nondefoliated pixels. Figure 4.5 may be compared with Figure

4.2, a reference sketch map of the southeastern portion of Roscommon County.

An and Chung (1994) have reported that better results are not guaranteed by larger

training sample sizes, citing Wann, Heidiger and Greenbaun (1990) as further

corroboration. Lawrence (1991) stressed that examples should be presented in
proportions that are fairly even.

NDVI values corresponding to the sampled pixels were then extracted from composites
coinciding with the three temporal defoliation stages and used to generate characteristic
signatures which were later employed to classify other cells. Two techniques
implemented in EASl/PACE were utilized in order to assess the degree of statistical
similarity between these training signatures so as to discern whether it was reasonable
to anticipate sharply defined spectral class partitioning by the classifiers.

These signature separability measures were the Transformed Divergence (TD) and the
Bhattacharrya or Jeffries-Matusita Distance (BD). While the BD method is considered to
have a stronger theoretical basis, both measures produce similar real numbers between
0 and 2. Values in the range from 0 to 1.0 are considered to be very poorly separable,
and values in the range from 1.0 to 1.9 are poorly separable. Good separability may be
anticipated when the calculated result is in the range from 1.9 to 2.0. Lower numbers
reflect a greater degree of statistical similarity, and may be indicative of either high
within-class training site variance or improper spectral band combinations (Richards,
1986). Both the TD and the 80 measures were calculated for the Lower Peninsula and
the Roscommon County defoliated and nondefoliated signatures.

67

 

 

 

 

 

Fl are 4.3 Complete Lower Peninsula reference data.
Black areas are defoliated. Green areas are nondefoliated. White is masked.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

I‘l
T
I ll

Fi ure 4.4 Sampled Lower Peninsula data.
Black areas are defoliated. Green areas are nondefoliated. White is masked.

 

 

 

 

 

 

 

 

 

 

-
I I
Complete Data Sampled Data

Figure 4.5 Roscommon County complete and sampled reference data.
Black areas are defoliated. Green areas are nondefoliated. White is masked.

4.1.1.4 Roscommon County Scene Extraction

For classification by the maximum likelihood method and by the BP ANN technique, a
subset of the full Lower Peninsula dataset was utilized. Because of the relatively coarse
spatial resolution of the AVHRR sensor, most pixels were ‘mixed,’ or nonhomogenous.
These cells with a mixture of both defoliation and nondefoliation content were
anticipated to be difficult to discriminate, and consequently a region was selected
through visual inspection of the digital files and the 1:100,000 topographic reference
sketch maps where, due to more concentrated and extensive known defoliation, it was
anticipated that there would be a greater proportion of ‘purer’ pixels than would be found

overall in the Lower Peninsula.

69

This region is approximately enclosed within the political boundaries of Roscommon
County, and was extracted by specifying the upper left and the lower right corner cell
row and column numbers, based on the full Lower Peninsula scene. A vector map of the
county boundaries for the State was placed over a raster NDVI scene and comer raster
pixels for the region to be extracted were selected that fell under the intersection of
county boundaries at the four vector county comers. Roscommon County scenes were

extracted from all files used in classification.

4.1.2 Maximum Likelihood Classification

Using data prepared as described in the preceding section, maximum likelihood
classification was conducted using EASl/PACE commercial software (PCl Inc.,
Richmond Hill, Ontario, Canada) on Lower Peninsula information as well as on
Roscommon County data. The classification technique was briefly summarized in

section 2.3.1.

Defoliated and nondefoliated training signatures were created for the statewide and
county tests, with equal a pn'ori probabilities employed in those instances when a pixel
to be classified was within overlapping portions of class signatures in the feature space.
The three composited NDVI images for the pro-defoliation, peak defoliation, and post
defoliation stages were the ‘bands’ used to provide the spectral information for the
classification, with both a Lower Peninsula set and a Roscommon County set used in
separate tests. The classifier was forced to assign all defoliated and nondefoliated pixels
based on the Mahalanobis distance. with the exception of only those pixels within the

70

specified regions which were masked as nonsusceptible to defoliation. With the
Mahalanobis measure of divergence, a lesser distance between a pixel vector and a
class vector indicates greater probability of class membership, as reviewed by Jensen
(1996). A file representing the most likely class of each pixel was generated following

Summary NDVI statistics for the Lower Peninsula full and sampled defoliated and
nondefoliated cells are presented in Table 4.2. Similar data for Roscommon County are
included in Table 4.3. Note that the highest mean NDVI value for defoliated cells within
each group occurs at peak defoliation in every case except for the Roscommon County
defoliated cells, which had a high post-defoliation figure. The Roscommon County peak
defoliation mean NDVI values for defoliated cells exceeded those for the Lower

Peninsula.

The signatures presented in Tables 4.2 and 4.3 reveal a deviation from the anticipated
reduction in NDVI at peak defoliation in the defoliated pixels. This postulated reduction
in NDVI was based on the controlling processes ouflined in sections 2.2.1 and 2.2.2. it
was a matter of concern that defoliated and nondefoliated class discrimination based on
such similar signatures would likely prove to be very difficult to accomplish. This
situation may reflect the inability of the coarse spatial resolution to adequately detect
insect-induced transformations in leaf state due to combined returns from multiple
reflectors within a very large sensor ground cell. Rather than indicating presence or
absence of defoliation, the resulting detector responses summarized in Tables 4.2 and
4.3 may instead primarily represent a broad measurement of typical seasonal growth

and later incipient senescence at the spatial resolution of the satellite. Though a

71

thorough investigation of this anomaly is beyond the scope of the current endeavor,
additional details regarding signature separability are provided in sections 2.2.1, 4.1.1.3,
5.1.1 and 5.1.2. Section 5.2 includes a discussion of the impact of mixed pixels in this

research.

Table 4.2 Lower Peninsula NDVI summary statistics
Mean, standard deviation.

CONDITION STAGE FULL SAMPLE
Defoliated Pre-defoliation 135.8, 17.4 135.5. 17.0
Peak defoliation 147.5, 17.9 146.6, 16.4
Post defoliation 144.5,18.1 142.5, 21.6
Nondefoliabd Pre-defoliation 132.2, 20.6 133.0, 19.0
Peak defoliation 145.4, 21 .4 146.0, 22.1
Post defoliation 142.1 . 20.4 142.2. 22.2

Table 4.3 Roscommon County NDVI summary statistics

Mean, standard deviation.

CONDITION
Defoliated

Nondefoliated

STAGE

Pro—defoliation
Peak defoliation
Post defoliation

Pre-defoliation
Peak defoliation
Post defoliation

72

FULL

128.9, 16.5
154.9. 13.4
153.4, 9.3

125.3. 1 3.7
149.0. 1 8.4
146.2. 1 7.1

SAMPLE

123.9, 16.7
1 52.3. 14.0
154.1 , 8.9

125.6, 10.8
152.4, 12.0
149.9. 8.8

4.1.3 Artificial Neural Network Classification

The classification of imagery by artificial neural networks was reviewed in section 2.3.2.
Two software packages were used which offered the ability to create and employ

artificial neural networks in the supervised classification of spectral information.

EASl/PACE was used for the ANN classification reported in this work. EASI/PACE
includes an implementation of a simple feedforward BP- supervised classifier with a
momentum term. This is the only neural network classification algorithm currently
supported in the package. One advantage this software possesses is that mapping the
results is easily and automatically accomplished by the module, thus facilitating visual

analysis.

The Stuttgart Neural Network Simulator (SNNS) was also employed in this research for
early comparisons. SNNS is offered free of charge by anonymous File Transfer Protocol
(FTP; Stuttgart Neural Network Simulator version 4.1, 1995). This software compiles on
a Sun workstation and on other computer systems, is fairly well documented, includes a
wide array of algorithms, provides onscreen network analysis and visualization tools,
and is supported to a limited extent through an email discussion list. SNNS was known
to be capable of producing and processing the small to medium-sized network anticipat-
ed for this study (Lutzy and Dengel, 1993).

While SNNS offers more speed, control of processing, graphical network analysis

capabilities and flexibility in algorithm selection than does EASl/PACE, it lacks a simple

method for quickly mapping output information. Therefore, the results reported here

73

were generated by the EASIIPACE algorithm, which is a variant of standard feedforward
backpropagation that includes a ‘momentum’ term. The two packages were found to
provide nearly identical output in previous research, given the use of an appropriately
configured BP with momentum algorithm in SNNS (Brown, Lusch and Duda, 1998).

The sampled data described in section 4.1.1.3 was used for training the simple BP with
momentum ANN for the Lower Peninsula and county data sets. The trained networks
were then assigned the task of classifying all pixels in the respective Lower Peninsula
and Roscommon County scenes, with the exception of those masked as being
nonsusceptible to defoliation. Tables 4.2 and 4.3 provide summary statistics for NDVI in
the Lower Peninsula and Roscommon County, respectively, as briefly discussed in

section 4.1.2.

The BF network architecture was established in part by the number of variables used as
the basis for classification, in this case three (pm-defoliation, peak defoliation, and post
defoliation NDVI). These three variables each were assigned to one input element in the
network. The architecture was also fixed by the number of classes to be defined, in this
case two (defoliation and nondefoliation). These two classes were each assigned to one
output element in the network. Thus, the number of processing elements in the input
and output layers were established by the data. The number of layers between these
two outer layers and the number of processing elements within each internal layer are
typically determined heuristically. These internal layers are referred to as ‘hidden.’ The
use of too many internal processing elements results in the memorization of the training
vectors, while too few hidden nodes render the network unable to learn through weight

updates (Hammerstrom, 1993b)

74

The Neural Network Frequently Asked Questions (FAQ) list maintained by Warren S.
Sarle of SAS institute, inc. offers the following comments regarding hidden layer

architecture optimization:

Some books and articles offer “rules of thumb" for choosing a topology - Ninputs
plus Noutputs divided by two, maybe with a square root in there somewhere - but
such miss are total garbage. There is no way to determine a good network topology
just from the number of inputs and outputs. it depends critically on the number of
training cases, the amount of noise, and the complexity of the function or
classification you are trying to Ieam. There are problems with one input and one
output that require thousands of hidden units, and problems with a thousand inputs
and a thousand outputs that require only one hidden unit, or none at all.

Other rules relate to the number of cases available: use at most so many hidden
units that the number of weights in the network times 10 is smaller than the number
of cases. Such rules are only concerned with overfitting and are unreliable as well.
All one can say is that if the number of training cases is much larger (but no one

knows exactly how much larger) than the number of weights, you are unlikely to get
overfitting, but you may suffer from underfitting (Sarle, 1998c).

Preliminary evaluation of alternative parameter specifications was conducted using
SNNS software. As suggested by the preceding quotation, parameter optimization could
quite easily have become the subject of a separate, detailed research project. The areas
investigated prior to completing the classifications reported in this work included
variations in backpropagation algorithms, number of input units, number of hidden
layers, number of nodes within each hidden layer, number of output units, ieaming rate,
momentum and number of iterations (epochs). Output was evaluated based on logged
training and testing errors, through visual interpretation of training and testing error
curves, and through the use of the “Analyze” tool within SNNS, which provided the
number of patterns classified correctly, incorrectly and those not classified. The tested

75

parameter values were within ranges recommended by the software developers and

other researchers.

The algorithms evaluated were Quickprop and backpropagation with momentum. The
inputs were expanded from the basic pre-defoliation, peak defoliation and post
defoliation NDVI values to also include ratios and differences of these figures. Due to
the coarse spatial resolution of the sensor compared with the areal extent of the targets,
a moving window of averaged NDVI values, or ‘texture’, was not added to the inputs.
Architectures with both one and two hidden layers were evaluated, containing

between 2 and 50 nodes per single hidden layer and 5 to 30 nodes for each of the two
hidden layers. Architectures with both one and two output nodes were W. The
ieaming rate was assigned within the range of 0.1 to 1.0, and momentum was restricted
to within 0 to 1.0, as suggested in SNNS documentation. ANN training was conducted
for up to 1,000 iterations, based on inspections of the graphed descent of the error
cunres. in addition to these modifications in ANN parameters, alternative data sampling
techniques were also evaluated during this phase. A markedly superior combination of
these key elements of the backpropagation ANN approach was not found through these
preliminary trials, when assessed as described in the preceding paragraph.
Consequently, EASIIPACE configuration recommendations, all within the range of
tested parameter values, were adopted for use in the final classifications.

Confirmation that ANN parameters have been specified such that the best possible
solution was realized is not attainable, and researchers must simply assess whether the
resulting accuracy is sufficient for the given application (Garson, 1991; Jain, 1994).

Caution must be exercised in assigning processing elements to the hidden layer, since

76

an overabundance could add complexity that serves no useful purpose, thus limiting

performance (Eliot, 1993).

The simple fully-interconnected network configuration used in producing the reported
classifications was of the form 3 input elements, 3 internal elements in one hidden layer,
and 2 output elements. The momentum utilized was 0.9 and the ieaming rate was set to
0.1, with a run length of 1000 processing iterations. Network training times for the small
file sizes were on the order of 0.25 hour or less on a Sun Ultra 2 workstation. Minimum
error thresholds were ignored in order to assign every pixel to the most activated class.
An output image was generated for each classification indicating the most likely class for

each pixel.

While the number of input variables presented for final classification could have been
increased through NDVI band differencing or ratioing or through the use of non-ratioed
AVHRR channel 1 and channel 2 data, this work restricted ANN (and maximum
likelihood) input to the simple case of pre-defoiiation, peak defoliation and post
defoliation NDVI ratios. in part, this was to assess whether such input was sufficient to
satisfactorily resolve the decision boundaries. in addition, such an approach attempted
to avoid the negative effects of added dimensionality while exploiting the advantages
cited in secfion 2.2.2 concerning the use of ratios. The implications of the ‘curse of
dimen'sionality' (described by Bellman, 1961) in ANN applications have been
summarized by Jenna Sinkkonnen to refer to the ail-consuming application of network
resources in the representation of “irrelevant portions of the space” due to the
‘exponential growth of hypervolume as a function of dimensionality“ (Saris, 1998b).

77

4.2 Accuracy Assessment

There is a lack of general guidelines for selecting a method for assessing classification
accuracy (Congalton, 1991; Foody, McCullough and Yates, 1995a), though common
approaches include percent correctly classified and Kappa. The percent correctly
classified overestimates classification accuracy because an adjustment for agreement
due to chance is not applied. Kappa, though addressing chance agreement, has been
shown by Foody (1992) to constantly underestimate the overall accuracy of the classi-
fication being evaluated. Further, Congalton (1991, p. 45) has noted that “only simple
random sampling completely satisfies' an assumption on which Kappa analysis is
based. Another method, Tau, has been proposed as an improvement over Kappa (Ma
and Redmond, 1995). in the manner of Foody, McCullough and Yates (1995a), the
percent correctly classified is presented in this work. Further, a crosstabulation of the
classed output versus reference information was performed to enable the creation of an
error matrix and the calculation of the more conservative Kappa statistic, as described
by Jensen (1996). Producer and user accuracies were thus also available for

comparison.

The error matrix compares the pixel assignment for each class in the reference and the
classification output. The percent correctly classified is the total number of cells classed
correctly divided by the total number of pixels classified, with the result multiplied by
100. PCC thus may range from zero to 100%. Producer accuracy reflects the percent
correct for each class based on the reference image. User accuracy considers the
percent correct for each class based on the classification output. The Kappa statistic

uses the total number of pixels under consideration, the sum of correctly classed pixels,

78

and the products of row and column totals to produce the amount by which the
classification exceeds random assignment, which would have a value of 0%. The
Kappa for purely nonrandom assignment would be 100%. Tests were conducted to
assess whether the Kappa values for each classification were significantly different than
zero. The results should be interpreted with caution, however, because by taking every
defoliated cell for accuracy assessment, the variance was underestimated because the
cells were not independent of each other. A significance test was performed to assess
the level of difference between ML and ANN error matrices for the Lower Peninsula.
Such a test for the two classifications of Roscommon County data was not possible due

to the presence of only values equaling zero in one row of the ANN matrix.

Muchoney and Haack (1994) used SPOT HRV and panchromatic (PAN) data acquired
on June 15, 1987 and July 4, 1988 in a comparison of the ability of several classifiers to
detect change caused by gypsy moths in the forest leaf state of a small locale in
Virginia. Though their methods differ from those employed in this study and the results
are thus not directly comparable, Muchoney and Haack’s published summary statistics

were contrasted with the results of this investigation.

Readers should be aware that Hudson and Ramm (1987) summarized the occurrence
of several erroneously published equations for the calculation of the Kappa statistic, and
provided the correct version to the remote sensing community. They further noted that
the correct formulation was presented by Fleiss, Cohen and Everitt (1969), and that the
measure was originally conceived by Cohen (1960). Jensen (1996) presented a correct

formulation, algebraically equivalent to that of Fleiss, Cohen and Everitt (1969) and

79

identical to the correct formulation noted by Hudson and Ramm (1987) to have been

published by Bishop, Feinberg and Holland (1975).

80

Chapter 5
RESULTS, DISCUSSION AND CONCLUSIONS

5.0 Results, Discussion and Conclusions

This work represented an attempt to utilize readily available AVHRR satellite imagery to
detect Michigan forest defoliation caused by gypsy moths by employing multitemporal
signatures and ancillary GIS data in supervised classification exercises. The maximum
likelihood technique, a conventional approach, and BP ANN, a somewhat nonconventional
method for this application, were the supervised classifiers chosen for the task. The analysis
integrated spectral, temporal, and spatial information in the classification and provided a test
of the potential to automate and improve upon the standard MDNR defoliation assessment,
which is based upon human interpretation of airborne video reconnaissance, by using the

commme applied NDVI indicator of vegetation existence and condition.

In this Chapter, the results obtained in all trials are reported, research questions are

answered, the implications are discussed and conclusions are drawn.

5.1 Results

As described in section 4.2, the accuracy of each classification result was quantitatively
assessed through the construction and interpretation of error matrices and the calculation of
the percent correctly classified and the Kappa statistic. The results obtained for each region

and classifier are summarized in the two following sub-sections.

81

5.1.1 Lower Peninsula Classification

All Michigan Lower Peninsula data were sampled and presented for maximum likelihood
and artificial neural network classification, using the methods presented in Chapter 4.
Complete data for the region were presented for classification based on signatures
developed from the samples. A summary of both trials is presented here.

Table 5.1 Error Matrix for Maxlrnum Likelihood classification of Lower Peninsula Data
Reference image

Defoliated Nondefoliated now TOTALS
Defoliated 1449 15015 16464
8.8%
Nondefoliated 493 6280 6773
92.7%
COLUMN TOTALS 1942 21295 23237
74.6% 29.5% 33.3% PCC
0.9% Kappa

In Table 5.1, results are presented for the maximum likelihood classification of Lower
Peninsula data. It can be seen that 74.6% of the cells that were actually defoliated were
correctly classed, yet only 8.8% of the pixels classed as defoliation were really defoliated.
The majority of cells were incorrectly assigned to the defoliated class, indicating a high rate
of commission error, and resulting in this difference in accuracy levels from the producers
and users perspectives. While only 29.5% of the actual nondefoliated cells were correctly
classed, 92.7% of those identified as nondefoliated were actually in that class. Overall,
33.3% of the nonmasked pixels in Michigan were correctly classified, however the low
Kappa of 0.9% suggests that chance played a large role in the classification accuracy. The

Kappa value was significantly different than zero at a 95% confidence level.

82

Table5.2 ErrorMatrleorNeuraiNetworkClasslflcatlonofLowerPeninsulaData

Reference image

Defoliated Nondefoliated now TOTALS

Defoliated 1799 18511 20310

8.9%

Nondefoliated 143 2784 2927

95.1%

COLUMN TOTALS 1942 21295 23237
92.5% 13.1% 19.7% PCC
1 .1% Kappa

The data in Table 5.2 reveal that, in the Lower Peninsula, the ANN classifier was able to
correctly identify 92.6% of those pixels known to be defoliated, but only 8.9% of the cells
classed as defoliated were actually members of that class. Again, a high rate of commission
error is observed. Only 13.1% of the truly nondefoliated pixels were correctly assigned, yet
in the classed image 95.1% of cells identified as nondefoliated were correct. The overall
19.7% correctly classified was low, and the Kappa value of 1.1% again indicated much
agreement due to chance. The Kappa value was significantly different than zero at a 95%

confidence level.

At a 95% confidence level, the Lower Peninsula ML and ANN classifications were not
significantly different. Both the maximum likelihood and ANN classifiers correctly classed a
substantial percentage of the known defoliation, while being far less successful in identifying
the nondefoliated pixels. This occurred because both techniques erroneously assigned over
60% of cells to the defoliated class. While the maximum likelihood method had an overall

percent correctly classified equal to 1.7 times that of the ANN approach, both were shown

83

by the Kappa statistic to have exceeded agreement due to chance by only a very small

margin.

The signature separability measures TD and BD, described in section 4.1.1.3, both indicate
a very poor separation between the defoliated and nondefoliated classifier training
signatures. The TD value was 0.074 and the BD measure was 0.071. The signatures are
statistically very similar, and thus do not contain suitable information to enable

discrimination between the classes.

5.1.2 Roscommon County Classification

As discussed in section 4.1.1.4, it was anticipated that classification accuracies would be
adversely affected by the preponderance of mixed pixels in the statewide coverage, due to
the coarse spatial resolution of the spectral sensor. For purposes of comparison, the
Roscommon County area was selected due to the widespread occurrence of known
defoliation within the county boundaries. All Roscommon County data were sampled and
presented for maximum likelihood and artificial neural network classification in a manner
similar to that performed for the Lower Peninsula. Complete data for the region were
presented for analysis based on signatures developed from the training samples. The
results from each trial are summarized here.

Tabie5.3 ErrorMstrleorMaxlmum LikelihoodClasslficadonofRoscommonCountyData

Reference image
Defoliated Nondefoliated ROW TOTALS
Defoliated 88 164 252
34.9%
Nondefoliated 72 431 503
85.7%
COLUMN TOTALS 160 595 755
55.0% 72.4% 68.7% PCC
22.7% Kappa

As detailed in Table 5.3, the maximum likelihood classification of the Roscommon County
data resulted in the successful identification of actually defoliated cells in 55.0% of the
cases, and 34.9% of the cells classed as defoliated were really members of that class. A
greater level of accuracy was achieved in the correct identification of truly nondefoliated
pixels, at 72.4%, and 85.7% of the pixels classed as nondefoliated were actually in that
category. The overall value of 68.7% correctly classified cells in the county was partially the
result of chance, yet the 22.7% Kappa statistic confirms that considerable agreement was
produced by reasons other than chance. The Kappa value was significantly different than

zero at a 95% confidence level.

The 22.7% Kappa statistic of the Roscommon County maximum likelihood classification is
consistent with results reported by Muchoney and Haack (1994), which ranged from 20.6%
to 45.4% and averaged 31.6%. They evaluated five classifiers for the detection of gypsy
moth defoliation in a 148 kilometer2 (57 miles”) study area in Virginia. For comparison, the
area of the Roscommon County scene was 1,571.7 kilometel’ (606.8 miles”). They provided
summary data for methods termed merged PCA, PCA eigenimage (density level slicing),
image differencing, spectral-temporal, and post-classification. The maximum likelihood

decision rule was applied in the merged PCA, spectral-temporal, and post-classification

85

approaches. Their work was based on June 15, 1987 peak-defoliation SPOT HRV-XS data
and July 4, 1988 peak defoliation SPOT HRV-XS and PAN information, and employed
1:24.000 US. Forest Service aerial sketch maps as reference information, digitized and
converted to raster format. Defoliation categories for 1988 were heavy defoliation from 61%
to 100% defoliated, moderate defoliation from 31 to 60% defoliation, and nondefoliated in
the range 0% to 30% defoliation. Categories for 1987 were only identified as defoliated or
nondefoliated. Consequently, they combined the 1988 data to codings similar to 1987,
including moderate defoliation in the defoliation category. The minimum defoliation polygon
mapped in their work was approximately 4 hectares (10 acres).

Significantly, the results achieved in the single-year multitemporal maximum likelihood
classification of Roscommon County 1,300 meter-square AVHRR data were comparable to
those produced using the much finer resolution two-year multitemporal SPOT HRV-XS 20-
meter-square multispectral data, which was resampled to 10 meters based on panchromatic
data. The Kappa statistic for the maximum likelihood classification of Roscommon County
data was very similar to that produced by Muchoney and Haack using spectral-temporal and
merged PCA classifiers, which had values of 20.6% and 24.9%, respectively. Their highest
Kappa value was 45.4% for image differencing, and their lowest Kappa was the 20.6%
figure for spectral-temporal. The average overall accuracy among the five classifiers in the
Virginia study was 64.2%, with the spectral-temporaland merged PCA classifiers having

61 .1%.and 63.3%, respectively. A high of 74.2% for image differencing and a low of 60.7%
for merged PCA were realized. This compares with the overall percent correctly classified
value of 68.7% for the Roscommon County maximum likelihood trial.

86

stle5.4 ErrorMatrleorNeuralNetworkClessificatlonofRoscommonCountyData

Reference Image
Defoliated Nondefoliated ROW TOTALS
Defoliated o o 0
0.0%
Nondefoliated 160 595 755
78.8%
COLUMN TOTALS 160 595 755
00% 100.0% 78.8% PCC
0.0% Kappa

in the ANN classification of Roscommon County data, total confusion exists with regard to
the identification of defoliated areas, given the input data and selected classification
parameters. In this trial, the error matrix in Table 5.4 shows that every pixel was assigned
to the nondefoliated class, resulting in 100.0% correct identification of the known
nondefoliated cells. There were 78.8% of classed nondefoliated cells placed in the proper
category. Not one defoliated cell was classed correctly with reference to either the ‘truth’
image or the classed image, and the level of accuracy in the classification of the targeted
defoliated pixels was thus zero. While overall 78.8% of all cells were classed correctly, the
wholesale assignment of every pixel to the dominant nondefoliated class resulted in a
Kappa statistic equal to 0.0%, indicating pure chance agreement. Note that there was an
unequal weighting of defoliated cells in the training sample, which would be interpreted by
the ANN classifier as biased a priori information, as mentioned in section 2.3.2 and section

4.1.1.3.

A significance test could not be perlonned to assess the level of difference between ML

and ANN classifications for Roscommon County due to values of zero in the ANN error

87

matrix. The 22.7% Kappa resulting from the ML method indicated superior performance

compared to the ANN approach which had a Kappa value of 0.0%.

For the defoliated class that is of particular interest in this research, the maximum
likelihood method returned superior results in both producer and user accuracy. The ANN
PCC was 10.1% higher than that of the maximum likelihood method because all pixels
were assigned by ANN to the dominant nondefoliated class, though the Kappa statistic was
far superior for the maximum likelihood classification. One similarity between the maximum
likelihood and ANN approaches was that each realized a higher PCC in the Roscommon

County exercise than was achieved in the Lower Peninsula classification.

As also found for the Lower Peninsula data, the signature separability measures TD and BD
both indicate a very poor separation between the defoliated and nondefoliated classifier
training signatures used for Roscommon County data processing. The TD value was 0.252
and the BD measure was 0.232. While these values represent an improvement over those
in the Lower Peninsula analysis by a factor of approximately 3.3 to 3.4, the signatures are
nonetheless statistically very similar, and thus do not provide sufficient information to enable

discrimination between the classes.

5.2 Answers to Research Questions and Discussion

1) Can the maximum likelihood supervised classification technique be successfully
applied to the problem of detecting defoliated areas?

At the county level, the maximum likelihood method used in this research was demonstrated

to generate a 55% producer classification accuracy for the primary defoliated target, with

88

approximately 35% of pixels classed as defoliated actually being defoliated. The results
compare favorably with those achieved in a study by other researchers (Muchoney and
Haack, 1994). However, the relatively low accuracy would most likely prove insufficient for a
monitoring program upon which land management decisions would be based. On a Lower
Peninsula basis, the low 8.8% user accuracy of cells classed as defoliated would no doubt
be unacceptable to land managers. Additional details related to this question are provided in
answer to the fifth research question. it is apparent that overlap exists in the class
probability density functions.

2) Can the ANN classification technique be successfully applied to the problem of
detecting defoliated areas?

Based on this trial, and given the previously described efforts to optimize architecture and
leaming parameters, it must be concluded that the defoliated targets cannot be satisfactorily
discriminated from the nondefoliated pixels by the ANN approach. While high producer
accuracy was achieved at the Lower Peninsula level, the low user accuracy and Kappa
statistic indicate substantial misciassification and dependence on chance agreement. A
more intensive investigation of the BP ANN method in this particular application may very
well enhance between-class boundary definition sufficiently to enable improved

classification, though the very poor training signature separability remains a critical concem.

3) How do the accuracy levels of the maximum likelihood and the ANN techniques
compare In this application, given the final parameters utilized?

As one might surmise from the answers to the two preceding questions, this work shows
that, given the methods adopted and the spectral information presented, the maximum
likelihood classification technique was better able to identify those pixels confirmed by the

89

interpretation of airborne video reconnaissance to contain trees defoliated through the
action of gypsy moths. In other recent work, Skidmore et al. (1997) also concluded that the
ANN method failed to yield significant advantages over more common classification

techniques in mapping forests using Landsat TM and GIS data.

There were similarities between the methods in the general degree of percent correctly
classified when the results were stratified by the areal extent of the classed region. Both
techniques had lower levels of accuracy, expressed as PCC, in the Lower Peninsula
exercises. This accuracy increased by a factor of two to four in the Roscommon County
tests. At the Lower Peninsula level, both techniques generated defoliated producer
accuracy in excess of 70%, and defoliated user classification accuracy was nearly identical
for the two methods. Kappa values, though low, were very similar in the Lower Peninsula
classifications. The inability of the ANN method to identify defoliation in the Roscommon
County test, and the comparatively superior overall performance of the maximum likelihood
technique in the same trial, led to the greatest difference in results between the two

classification methods.

Using the Kappa statistic as the judging criterion, the 22.7% achieved by the maximum
likelihood classification of Roscommon County data was the best outcome realized in these

trials.

4) Are the classification accuracy levels for defoliated and nondefoliated pixels
similar?

The two classification methods returned varying results depending on the method by which

accuracy was assessed.

in the Lower Peninsula tests, the ANN approach provided superior producer accuracy for
the target defoliated cells, while the maximum likelihood technique had an advantage in
terms of producer accuracy in the discrimination of nondefoliated areas. Both methods
generated similar results for each category when user accuracy was evaluated. As a
percentage of the reference image, the ANN technique applied to the full Lower Peninsula
data set displayed a greater ability to accurately identify defoliated pixels, by a margin of
18% (92.6% - 74.6%). As a percentage of the classed image for the same situation, the
results were nearly identical, with only 0.1% greater accuracy for the ANN approach (8.9% -
8.8%). As a percentage of the reference image for Lower Peninsula data, the maximum
likelihood technique more accurately classed nondefoliated pixels by a margin of 16.4%
(29.5% - 13.1%). As a percentage of the classed image for this situation, accuracy was

similar, with a difference of only 2.4% (95.1% - 92.7%), in favor of the ANN method.

As a result of the wholesale assignment of all pixels to the Roscommon County
nondefoliated class by the ANN technique, the accuracy rate for the nondefoliated pixels
was far superior to that for the defoliated pixels. which were classed with an accuracy of
zero. The margin of difference in ANN classification accuracy between defoliated and

nondefoliated cells as a percentage of the reference image was 100.0%. As a percentage of

91

the classed image, this difference was 78.8%. In both cases, the margin was in favor of the
nondefoliated class.

The least difference between defoliated and nondefoliated producer and user accuracy
levels among the four trials was realized for both accuracy viewpoints by the maximum
likelihood Roscommon County classification. In this case, there was a difference of 17.4%
between the producer accuracies for defoliated and nondefoliated classes. There was a
50.8% difference between the user accuracies for defoliated and nondefoliated classes. The
greatest polarization in terms of producer accuracies occurred in the ANN classification of
Roscommon County data, with a difference of 100%. The greatest difference when
considering user accuracies occurred in the ANN classification of the Lower Peninsula data,
though the 86.2% value was similar to those for the three trials other titan the Roscommon

County maximum likelihood test.

5) At a subpixel level of detail, what insight may be gained from an analysis of pixel
homogeneity?

Due to the ‘fuzziness’ of the resulting spectral information, pixels containing a combination
of the desired class and another class, or several other classes, are difficult for any hard
classifier to properly discriminate and assign when partitioning the feature space. When
considering the effect of mixed pixels in these classifications, several additional questions
may be raised in connection with the work described, which will be answered in this section

in relation to the primary issue of pixel homogeneity, or lack thereof.

92

This discussion takes a different approach from that of Foody (1996), who investigated the
relationship between mixed pixels and the output values of an ANN classifier after spatially
degrading finer resolution imagery to approximate AVHRR data. He found that, though the
output activation values did not have a strong relationship to the pixel content, it was

possible to rescale these figures to markedly improve the relationship.

What is the abundance of mixed pixels In the Lower Peninsula and Roscommon
County data sets?

it is certain that most defoliated pixels in the entire data set are mixed. Table 5.5 shows the
Lower Peninsula and countywide abundance of defoliated pixels for each of four ranges of
true defoliation percentages. Only 19.1% of all Lower Peninsula defoliated pixels (372 of
1,942) had true defoliation levels greater than or equal to 50%, which was used as the
selection criterion for the defoliated training sample. In Roscommon County, 34.4% of the

defoliated pixels had a percent defoliated greater than or equal to 50%.

Table 5.5 Lower Peninsula and Roscommon County defoliation by actual percent
defoliated quartiles

Actual Percent Lower Lower Roscommon Roscommon

Defoliated Peninsula Peninsula Co. Number of Co. Percent of
Number of Percent of All Defoliated All Defoliated
Defoliated Defoliated Pixels Pixels
Pixels Pixels

75 to 100 78 4.0 11 6.9

50 to 75 294 15.1 44 27.5

25 to 50 473 24.4 30 18.7

0 to 25 1097 56.5 75 46.9

COIUIIII 701818 1942 100.0 160 100.0

93

How do the frequency dlstrlbutlons of the per-plxel proportion defoliated compare for
all defollated pixels In the Lower Peninsula and In Roscommon County, and do these
differ from the distribution for the correctly classified defoliated pixels In the best
outcome realized among the four classlficatlons performed?

Figure 5.1 and 5.2 provide a clear visual impression that the majority of pixels had low
levels of defoliation in both the Lower Peninsula and the Roscommon County data sets. In
contrast to the Lower Peninsula histogram in Figure 5.1, the frequency distribution in Figure
5.2 for Roscommon County appears somewhat bimodal, and differs markedly from that for
the Lower Peninsula, though the highest frequencies are still in the lowest proportions
defoliated. The spatial distribution of the information presented in Figure 5.2 is shown in

Figure 5.3, and it can be seen that the primary region of high percent defoliation occurs

near the center of the large cluster of defoliated pixels.

 

 

0.03 QED 0.18 0.27 0.3 0.5 0.54 0.63 0.72 0.81 OR)

Figure 5.1 Frequency distribution of proportion defoliated for all
Lower Peninsula defoliated pixels.

94

 

 

0CD one 0.18 0.27 0.3 0.46 0.64 0.63 0.72 0.81 can

Figure 5.2 Frequency distribution of proportion defoliated for all
Roscommon County defoliated pixels.

 

 

 

 

Figure 5.3 Spatial distribution of proportion defoliated for all
Roscommon County defoliated pixels.
Red is high percent defoliation, yellow is mid-range, and green is low.

95

Figure 5.4 shows the frequency distribution for the actual proportion of defoliation for those
Roscommon County pixels correctly classed as defoliated by maximum likelihood. The
bimodal distribution apparent in Figure 5.2 for all defoliated pixels in the Roscommon

County data remains evident, and is slightly more pronounced.

 

 

0.CD one 0.18 0.27 0.3 0.45 0.54 0.53 0.72 0.81 090

Figure 5.4 Frequency distribution of proportion defoliated for all Roscommon County
defoliated pixels correctly classed as defoliated by ML classifier.

Roscommon County was selected for detailed analysis because it was one of the few
counties to experience widespread defoliation during 1993, and thus a higher percentage of
pixels were anticipated to be more thoroughly defoliated than in most other counties. This
higher level of pixel defoliation saturation is seen when comparing the percent defoliation

frequency distributions for the Lower Peninsula and Roscommon County. In this sense,

96

Roscommon County was not representative of all counties in the entire State of Michigan.
Counties having arbitrary political boundaries, this was not considered significant since the
task was solely to develop characteristic spectral signatures in order to enable the
identification of defoliated sites through multitemporal NDVI input vectors. Such a scene
encompassing a dense yet extensive defoliation cluster with a higher frequency component
at maximum pixel purity levels served the research purpose since the data were screened
to include only broadly similar forested areas. Selecting only the intensely defoliated
Roscommon County cluster would, however, have certainly yielded too few pixels for
classifier training and evaluation purposes (Jensen, 1996). The less mixed, or more
homogenous, pixels were desired in order to optimize classifier training, as is customary in
supervised classification efforts (Jensen, 1996; Schowengerdt. 1983).

The Roscommon County maximum likelihood results will serve as the focus for analysis in
the following questions ccnceming the relationship between accuracy and pixel
homogeneity. Based on the Kappa statistics presented in section 5.1, this classification was
clearly superior to the other three tests, and a detailed review of the accuracy at subpixel
levels may shed light on how this was achieved. Recall that the Kappa values indicated

negligible agreement beyond chance for the other classifications.
Were cells outside the percent defoliated training range also accurately classified?

In the training, only pixels having greater than or equal to 50% defoliation were used in the
defoliation sample. Furtheranalysis, summarized in Table 5.6, revealed that 60.2% of the
pixels correctly classed as defoliated in Roscommon County had a true defoliation level of
less than 50%. In addition, 60.0% of all the defoliated pixels with a true defoliation level of

97

greater than or equal to 50% were correctly classed. while 50.5% of all the defoliated pixels
with a true defoliation level less than 50% were correctly classed. The maximum likelihood
classifier thus did correctly identify many of these more highly mixed pixels outside the
range of those presented during signature development, though the Kappa statistic did

indicate chance assignment played a role in the overall accuracy realized.

Table 5.6 Roecomrnon County maximum likelihood results for training sample range and
tor category outside training range

ActualPercentDefollstlon PercentovaerallCorrectly PercentCorrectofAll

Range Classed Defoliation for Defoliated Pixels in Range
I“IMO
Greaterthanorequaito50% 35/88239.8% 35/55=60.0%
Lessthan50% 5338860296 53/105=50.5%

How does classification accuracy vary for differing pixel mixture levels of correctly
and Incorrectly classed defoliation, and are the less mixed, or purer, defoliated plxels
Indeed more accurately classed?

The relationship between the classification accuracy for defoliated pixels and the degree of
pixel class purity (which may also be termed the mixture level) is shown for the Roscommon
County maximum likelihood classification in Table 5.7 and Figure 5.5. As anticipated, the
percentage of pixels correctly classified as defoliated increases as the level of actual
defoliation increases. Larger percentages of pixels were incorrectly classed as the level of

actual defoliation decreased.
Correctly classed defoliated pixels represent a larger percentage of the pixels in the highest

percent defoliation range (75 to 100%) than incorrectly classed defoliated pixels (81.8%
versus 18.2%, respectively). incorrectly classed defoliated pixels represent a larger

98

percentage of the lowest percent defoliation range (0 to 25%) than correctly classed

defoliated pixels (53.3% versus 46.7%, respectively), though by a much smaller margin than

the previous situation (63.6% and 6.6%, respectively).

Table 5.7 Roscommon County maximum likelihood defoliated accuracy and class

homogeneity

 

 

 

 

 

Actual Defoliation Defoliation Defoliation Defoliation Defoliated Correct
Percent Correctly Correctly Incorrectly Incorrectly Pixels In Divided by
Defoliation Classed Classed Classed Classed Range Incorrect
Ran
(Mixture Number of Percent of Number of Percent of Total Ratio
Level) Pixels Range Pixels Range (Percent)
75 to 100 9 81.8 2 18.2 11 (6.9) 4.50
50 to 75 28 63.6 16 36.4 44 (27.5) 1.75
25 to 50 16 53.3 14 46.7 30 (18.7) 1.14
0 to 25 35 46.7 40 53.3 75 (46.9) 0.88
Column 88 NA. 72 MA. 160 (100.0) N.A.
Totals
90
80 if
§ 70 it
I
E 60 4—
.E
a 50 1' —e— Correctly Gassed
E 40 4 —I—- incorrectly Gassed
.6 30 .-
E 20 ~
3.
10 i
0 f i .L t
o 20 4o 60 80 100
Percent Defoliation

 

 

 

Figure 5.5 Subpixel percent of range versus percent defoliation for Roscommon

County ML classification.

The four midpoints of the percent defoliation ranges were used to plot the percent defoliation.

99

The percent of pixels in a defoliation range (y-axis) for the correctly classed pixels increases
with increasing actual percent defoliation (x-axis) (Figure 5.5), while the percent of pixels in
a range for the incorrectly classed pixels decreases with increasing actual percent
defoliation. The crossover point, or point at which the correctly classed and the incorrectly
classed curves intersect, occurs at approximately 25.5% actual defoliation, as scaled from
Figure 5.5. A vertical line above and below this point and intersecting the x-axls separates
two regions where either the correctly classed curve (on the right side) or the incorrectly
classed curve (on the left side) dominates the percent defoliation range (y-axis). The
correctly classed defoliated pixels with greater than 25.5% actual defoliation possess a
larger percent of the defoliation range than the incorrectly classed defoliated pixels. The
incorrectly classed defoliated pixels with less than 25.5% actual defoliation claim a larger

percent of the defoliation range than the correctly classed pixels.

What are the Implications concernlng the sensitivity of the classifier, and what do the
classification results suggest ccnceming the area of detectable defoliation wlthln
pixels?

it is noteworthy that the preponderance of correct classifications generated by the maximum
likelihood classification technique begins to occur at a level of defoliation on the order of
only 1/4 pixel in areal extent. Recall that the training sample consisted entirely of pixels with
greater than or equal to 1/2 pixel defoliation, and that this portion represented only 34.4% of
the total defoliated pixels. The pixels in this same training range of 50% to 100% defoliation
represented only 39.8% of all correctly classed defoliated pixels, as noted in Table 5.6.
Thus, 60.2% of the correctly classed defoliated pixels were outside the sampled defoliation

range presented to the classifier for signature development. The maximum likelihood

100

classifier thus exhibited some ability to identify nonhomogenous pixels, yet there was a
corresponding penalty in that 65.1% (164/252) of pixels classed as defoliated were actually
nondefoliated, as found through interpretation of the error matrix in Table 5.3. The foregoing
indicates the existence of overlapping decision boundaries.

When considering the application of this general AVHRR-based approach to gypsy moth
defoliation detection, and in related fieldwork, it could be useful to know what areal extent of
defoliation occurrence is required to begin achieving more correct classifications than

incorrect classifications for defoliated pixels.

Percent defoliation is equivalent to the areal extent of defoliation within a pixel divided by
the total area of the pixel. For the 1.3 kilometer-square mid-latitude spatial resolution of the
AVHRR pixels employed in this exercise (1 .3’ = 1.69 kilometer” area), the 25.5% actual
pixel defoliation crossover value (value of x at the intersection of the correct and incorrect
curves in Figure 5.5) represents an area of 43.1 hectares (106.5 acres). Thus, correctly
classed within-pixel defoliated areas greater than 43.1 hectares (Le. greater than 25.5%
actual defoliation) possess a larger percentage of the defoliation range than the incorrectly
classed defoliated pixels. This percentage difference steadily increases with greater pixel
percent defoliation from the crossover point to the maximum x-axis value, or highest percent

defoliation, as seen in Figure 5.5.

Another way to consider this accuracy information in relation to pixel mixing is to note that
the ratio of correct to incorrect is 1.0 at the intersection of the correct and incorrect curves
and greater than 1.0 to the right of the intersection. It is less than 1.0 to the left of the

intersection, which occurs at 25.5% defoliation. Figure 5.6 illustrates this relationship

101

between the correctfincorrect ratio and the percent defoliation. At the maximum percent
defoliation, the correctfincorrect ratio is 4.50 (81.8%l18.2%), up from 0.88 (46.7%/53.3%)

for the lowest percent defoliation range.

 

 

 

 

O .L .
12.5 37.5 62.5 87.5
Percent Defoliation

 

 

 

 

Figure 5.6 Correct/Incorrect ratio versus percent defoliation for Roscommon
County ML classification.

The curve is nonlinear. The steepness of the curve between the 62.5% and 87.5%
defoliation values indicates the relatively rapid improvement in classification accuracy within
this zone as percent defoliation increases. The 62.5% and 87.5% values represent range
midpoints within the 50-75% and 75-100% ranges, which also reflect the full range of the

defoliated training sample used in the classification. At lower levels of defoliation content,

102

the improvement in classification accuracy was less pronounced for equal changes in
percent defoliation. This general relationship also existed in the maximum likelihood
classification of the Lower Peninsula data, but was not observed in the artificial neural
network classifications at either regional level, as depicted in Figure 5.7. For reference, the
percent defoliation levels in this figure and in Figure 5.6 correspond to the units of area
provided in Table 5.8. Recall, however, that due to the method of converting the vector
ranges of MDNR-interpreted defoliation percentages observed on video imagery, the

maximum possible percentage of defoliation is 87.5.

The foregoing has provided insight into the limitations imposed by the coarse resolution
AVHRR receiver, given the areal distribution of target occurrences. If one considers that
pixel class purity is the single greatest determinant of classification accuracy, the use of a
finer spatial resolution sensor would be expected to improve classification accuracy,
assuming cloud-free scenes could be acquired in the desired temporal windows at suitable
wavelengths. For example, the Landsat Thematic Mapper (TM) imagery at similar
wavelengths to those used by the AVHRR sensor has a spatial resolution of 30 meter-
square (0.03 kilometer-square), in contrast to the 1,300 meter-square (1.3 kilometer-
square) AVHRR resolution. The resulting 0.0009 kilometei‘ area of a TM pixel is very small
when compared with the 1.69 kilometer2 AVHRR pixel, and 1,877.78 TM pixels are used to
cover the same surface area as just one AVHRR pixel. However, note that Muchoney and
Haack (1994) did not achieve substantially better results using relatively fine resolution

SPOT data.

103

 

 

 

 

 

 

 

.9
i
a: 10 i + M m
8 + M M.
FDS m
6 + nos ML
4 i
2 "V
0 .
12.5 37.5 62.5 87.5
Percent [bfolletlon

 

 

 

Figure 5.7 Correct/Incorrect ratio versus percent defoliation for Lower Peninsula and
Roscommon County ML and ANN classifications.

In the legend, Ml denotes Lower Peninsula data, and ROS signifies Roscommon County data. NN
is the ANN classification, and ML is the maximum likelihood classification.

Table 5.8 Areal extent and correct/Incorrect classification ratio for key values of percent
defoliation

AVHRR Percent Roscommon County Defoliation Area Defoliation Area
Defoliation (midpoint Maximum Likelihood km' acres
of range, except for Correct/incorrect

25.5 and 100.0) Ratio
100.0 NA. 1.69 417.6
87.5 4.50 1 .48 365.4
62.5 1.75 1.06 261.0
37.5 1.14 0.63 156.6
25.5 1.00 0.43 106.5
12.5 0.88 0.21 52.2

1.0 kilometer” = 247.10538 acres. 1 hectare = 2.4710538 acres.

The use of 30-meter-square pixels instead of 1,300-metersquare pixels would, of course,
result in a large increase in file size and processing time, as well as increased costs. For
coverage of the Lower Peninsula of Michigan, on the order of 3 billion pixels would be
needed prior to aggregating to produce a map of percent defoliated with approximately 188
million 30-meter TM pixels. For comparison, there were just under 100,000 pixels in the
aggregated Lower Peninsula AVHRR image. The acquisition of a suitable number of cloud-
free scenes at a sufficiently high temporal resolution could be problematical in this region,

however.

5.3 Conclusions and Recommendations for Future Research

Overall, among the four different classifications performed, the accuracy of the spectral-
temporal classification of gypsy moth defoliation ranged from 19.7% to 78.8%. Producer
accuracy for the defoliated class reached a high of 92.6% in the Lower Peninsula ANN
test, and the highest user accuracy for the same class was 34.9%, generated in the
Roscommon County maximum likelihood experiment. However, classification accuracy in
only one of the four trials exceeded chance agreement by more than a negligible margin,
as determined by Kappa tests. This result, 22.7%, was achieved by the Roscommon
County maximum likelihood classification, and was consistent with values reported by
Muchoney and Haack (1994) for several different classification techniques using spectral

information with a much finer spatial resolution.

Considerable misciassification resulted from the use of nonseparable training signatures
which were in turn caused by severely mixed coarse resolution spectral pixels, rendering

the output unsuitable for use in contributing to the management of the gypsy moth forest

105

defoliation detection problem. The applicability of the work was thus limited at least in part
by a mismatch between the smaller typical target sizes and the large ground resolution of
the sensor. This occurred even though a study period was selected that provided extensive
clusters of defoliation in Michigan, compared to most other years for which data were
available. The sensor selection involved a choice between conflicting needs for high
temporal as well as high spatial resolution, given adequate spectral and radiometric
resolution and minimal costs of data acquisition. With limited training data and poorly
separable training signatures, the generalizing nature of the maximum likelihood technique
yielded greater accuracy in the Roscommon County trial than did the attempt by the
artificial neural network to seek detail with insufficient supporting information. The percent
correctly classified values for the two classifiers were similar at each regional level, though

both produced higher percentages in the Roscommon County analysis.

The Roscommon County maximum likelihood accuracy results indicate that, while
problems involving spatial resolution existed, the NDVI values representing the spectral
component did provide information that enabled the identification of the targets. The use of
individual channel data, in addition to NDVI, may facilitate improved definition of the class

boundaries.

it is unlikely that an alternate sampling method would produce improved defoliation or
nondefoliation recognition accuracy. At the Lower Peninsula level, 19.2% of all defoliated
pixels were used in training, and these possessed the highest percentages of defoliation
content, or the greatest purity. In the Roscommon County investigation, 14.4% of all
defoliated pixels were employed in training, and these were also at the highest levels of

percent defoliation.

106

In addition to the nonhomogenous nature of the pixel class content, another limiting factor
encountered with this data was the paucity of defoliated training cells. For the ANN
technique, this resulted in a limited amount of training pixels per weight in the architecture,
even though a minimal number of internal, or hidden, processing elements were specified.
For the defoliated target, this ratio was equal to 24.8 for the Lower Peninsula and only 1.5
for Roscommon County. A very slightly higher Kappa statistic resulted for the Lower
Peninsula ANN classification. The sampling method favored the ANN approach by
selecting nondefoliated training cells at the spatial class boundaries instead of in the midst
of similarly classed pixels as is typical in maximum likelihood analysis.

While suboptimal ANN parameter specification may have contributed to reduced overall
classification accuracy, the similarly low level of accuracy achieved by the maximum
likelihood method suggests that the realized outcomes were primarily influenced by other

causes.

Analysis of subpixel information provided confirmation that, as anticipated, correct
defoliation classification was directly, though nonlinearly, proportional to the percent of
each cell that was defoliated. Correct results rose dramatically when the defoliation content
exceeded 62.5%, as evidenced by the county level study that was configured to provide a
greater abundance of more highly defoliated pixels than were available on a Lower

Peninsula basis. Correct classification of cells with low percent defoliation did occur.

Given the finding from this thesis research that greater accuracy might be achieved by
minimizing ground mixing through the use of a finer resolution sensor, why, then, did such

an approach fail in other research to produce accuracies much greater than those realized

l07

with the coarse resolution AVHRR sensor? Under peak defoliation leaf-off conditions, lower
NDVI values were expected due to the removal of the high-NDVI leaf component of the
spectral reflectance signal. This would facilitate the identification of affected regions. Yet,
with the absence of a dense canopy, one might surmise that any reduction in NDVI through
leaf loss may have been countered by the greater contribution in the near-infrared of the
undergrowth to the reflectance received at the sensor, as discussed by Spanner et al.
(1990). This would thus inhibit multispectral classification of defoliation, even with data
acquired at a fine spatial resolution. However, visual inspection of Michigan CIR imagery
indicated minimal NIR contribution from the understory, and this explanation seems
doubtful (Lusch, 1998). This question remains unanswered, yet may involve differences in
methodology, and also the tradeoff between boundary pixel influence and spectral variance
as a function of spatial resolution as considered by Markham and Townshend (1981) and
Woodcock and Strahler (1987).

The influence of soil moisture variability on NDVl may play some role in partially explaining
the lack of substantial improvement in classification accuracy when a finer spatial
resolution sensor was employed by other researchers. Todd and Hoffer (1998) investigated
the impact of changes in vegetation cover, soil texture and soil moisture content on NDVI,
and confirmed that, as anticipated, increases in the percentage of vegetative cover yielded
increases in NDVI. They observed that, for equal percentages of vegetation, NDVI levels
increased for soils with a greater moisture content. “NDVI values associated with low
reflecting soils were elevated compared with those associated with high reflecting, brighter
soils' and ‘differences in soil type were less important than differences in soil moisture, in

their impact on NDVI“ (Todd and Hotter, 1998, p. 918). They also found that the effects of

108

soil on the NDVI ratio were most pronounced when cover density was in the range from

20% to 80%.

The coarse spatial resolution of the AVH RR imagery, the lack of radiometric and
atmospheric calibration, variations in time and platform of acquisition, maximum value
compositing, and timing of observations compared with forest defoliation stages were key
sensor concerns in this endeavor. The suitability of the network architecture, adequacy of
the training set, and the hazard of local error minima were the primary issues in the ANN
classification. Compliance with Gaussian requirements was a factor in the maximum
likelihood tests. The reference data against which both classifier outputs were judged was
assumed to meet accuracy requirements, yet verification was not possible. There was a
substantial lapse of time between AVHRR data collection and the acquisition of information
that formed the basis for the forest mask applied in preprocessing. The required vector to
raster conversion resulted in an increase in the extent of classed defoliation, and at the
county level, there was a sizable difference in the number of training pixels for each class.
Misregistration error between the reference information and satellite imagery did exist,
though this was within generally accepted limits. The targets in this detection exercise
were of variable, and often small, extent, especially when compared with the ground pixel

size of the multispectral acquisition sensor.

Ehrlich, Estes and Singh (1994) have stated that the lack of radiometric calibration in the
AVHRR sensor is the primary shortcoming of this platform. Vermote and Kaufman (1995)
offered a method for the absolute calibration of AVHRR channels 1 and 2. Further, the
negative influence of atmospheric variability and differing sun angles were left uncorrected

in the fine-resolution study by Muchoney and Haack (1994) and in this work, leaving the

109

two studies comparable in this regard. A spectral adjustment in the multitemporal images,
as performed by Mikkola (1996) using the multiple regression method of Olsson (1993),
could render the frequency distributions of digital numbers in all scenes more directly

comparable, and this is recommended for any future research.

Given the issues mentioned in the preceding paragraphs, this research attempted to fully
exploit the generally reported robustness of the artificial neural network approach, and, to a
certain extent, tested the limiting assumptions and suitability of the maximum likelihood
classifier. It is recommended that the Forest Management Division of the Michigan
Department of Natural Resources and other interested parties seek alternate approaches
to the automated identification of scattered occurrences of forest defoliation caused by
gypsy moths, as were typical during the 1993 season. These may include either the use of
additional techniques to enhance multitemporal AVHRR data quality or the use of a sensor
with finer spatial resolution than employed in this study. The former approach may remain
limited by the contrast between the areal extent of the target and the sensor ground cell,
and by the resulting aggregated spectral information. The latter option may raise additional
concerns involving temporal resolution, and did not yield greatly superior accuracies when

attempted by Muchoney and Haack (1994).

110

APPENDIX

111

Appendixa
Selected World Wide Web Resources
Related to .

Michigan Forest Defoliation Detection Using Multitemporal
NDVI Signatures Derived From AVHRR Satellite Imagery

This listing is also available Online at

http:llwww.ssc.msu.edul~geo/stu/dudalwwwgyp.html
Accessto all sitesverified on April 2, 1998

Mutations

Michigan Department of Natural Resources (MDNR)
http://www.dnr.state.mi.usl

MDNR Forest Management Division (FMD)
http://www.dnr.state.mi.us/www/fmd/fmdhome.html

Michigan State University (MSU) Department of Geography
http://www.ssc.msu.edu/~geo/

MSU Center for Remote Sensing and Geographic lnfonnation Science
httpzllwww.crs.msu.edul

Great Lakes Environmental Research Laboratory (GLERL)
http'J/www.glerl.noaa.gov/cw/cw.html

National Oceanic and Atmospheric Administration (NOAA)
httpzllwww.noaa.gov/

NOAA National Oceanographic Data Center (NODC)
http://www.nodc.noaa.gov/

NOAA NODC Midwestern Climate Center
httpzllmcc.sws.uiuc.edul

112

mm

1993 Michigan Forest Inventory 8. Analysis
http:llwww.dnr.state.mi.uslwww/fmd/fia/surveyun.html

Climate Summaries for the Midwest - Station Choices in Michigan
http:l/mcc.sws.uiuc.edu/Summary/Michigan.html

The Atlas of Michigan Project
http:llwww.msu.edu/~atlasl

Wishes

Gypsy Moth in North America
hwy/gypsy.fsl.wvnet.edu:80/gmottv

Michigan's Gypsy Moth Education Program
http:llwww.ent.msu.edulgypsyed/

MDNR FMD Forest Health Page
http:llwww.dnr.state.mi.us/wwwffmd/pest/forhealhtml

Evaluation of Forest Susceptibility to the Gypsy Moth Across the Coterminous
United States
http:llgypsy.fsl.wvnet.edu/gmoth/suscept/suscept.html

WW

Advanced Very High Resolution Radiometer
http:l/edcwww.cr.usgs.gov/glislhyper/guide/avhrr

The Remote Sensing Tutorial - An Online Handbook
http:l/code935.gstc.nasa.govl'l'utorial/Start.html

Satellite Environmental Monitoring of the Great Lakes: A review of NOAA's
Great Lakes CoastWatch program.
http:llwww.glerl.noaa.govlctrb/ctrb_0081 1 .htrnl

NOAA CoastWatch Active Access System (NCAAS)
http://www.nodc.noaa.gov/NCAASIncaas-home.htrnl

113

Classification

PCl Maximum Ukellhood Classification
http:llwww.pci.on.ca/cgi-binlpcihlp/MLCIALGORITHM

What is an Artificial Neural Network?
http:/Iwww.emsl.pnl.gov:2080/proj/neuronlneurallneural.ann.htrnl

Neural Network Frequently Asked Questions (FAQ)
ftp:/lftp.sas.com/pub/neural/FAO.html

Sommers

EASIIPACE and lmageWorks from PCl Geomatics Group
http:llwww.pci.on.ca/

Stuttgart Neural Network Simulator (SNNS)from the University of Stuttgart - Version 4.1
Manual
http:l/www-ra.informatik.uni-tuebingen.de/SNNSIUserManual/UserManualhtml

IDRISI from Clark Labs
http:llwww.idrisi.clatku.edul

SNNS FTP Software Download Site in lpub/SNNS/ as file SNNSv4.1.tar.z
fth/ftpjnfonnatikuni-stuttgartde/

114

REFERENCES CITED

115

REFERENCES CITED

Ahem, F. J., J. Sirois, W. D. McColl, R. P. Gauthier, T.T. Alfoldi, W. H. Patterson, and T.
A. Erdle, 1991. Progress toward improving aerial defoliation survey methods by

using electronic imagers. Photogrammetric Engineering & Remote Sensing. 57(2):
1 87-1 93.

Albert, Dennis A., 1995. Regional landscape ecosystems of Michigan, Minnesota, and
Wisconsin: A working map and classification (Fourth Revision: July 1994). North
Central Forest Experiment Station, Forest Service, United States Department of
Agriculture, General Technical Report NC-f 78, 250 p.

Allen, Thomas R., Thaddeus J. Bara, and Stephen J. Walsh, 1994. Observed biases in
the conterrninous U.S. AVHRR satellite data set: A North Carolina case study.
Geocarto lntemational, 9(2): 52-62.

An, Ping, and Chang-Jo F. Chung, 1994. Neural network approach for geological

mapping: technical background and case study. Canadian Journal of Remote
Sensing, 20(3): 293-301.

Avery, Thomas E., and Graydon L Berlin, 1992. Fundamentals of Remote Sensing and
Airphoto Interpremtion. Macmillan, New York, 472 p.

Bailey, David, and Donna Thompson, 1990. How to develop neural-network
applications. Al Expert, 5: 38-47.

Barbour, Michael G., and William Dwight Billings (eds), 1988. North American terrestrial
vegetation. Cambridge University Press, New York.

Bellman, R., 1961. Adaptive Control Processes: A Guided Tour. Princeton University
Press.

Benediktsson, Jon A., Philip H. Swain, and Okan K. Ersoy, 1990. Neural network
approaches versus statistical methods in classification of multisource remote

sensing data. IEEE Transactions on Geoscience and Remote Sensing, 28(4): 540-
552.

Bischof, H., W. Schneider, and A. J. Pinz, 1992. Multispectral classification of Landsat-
images using neural networks. lEEE Transactions on Geoscience and Remote
Sensing, 30(3): 482-490.

Bishop, Y. M. M., S. E. Feinberg, and P. W. Holland, 1975. Discrete Mulfivanate
Analysis - Theoryand Practice. MIT Press, Cambridge, Mass. 575 p.

Blamire, P.A., 1996. The influence of relative sample size in training artificial neural
networks. Intemational Joumal of Remote Sensing. 17(1): 223-230.

116

 

Brown, Daniel G., David P. Lusch, and Kenneth A. Duda, 1998. Supervised
classification of types of glaciated landscapes using digital elevation data.
Geomorphology, 21: 233-250.

Center For Remote Sensing, 1991. CMAP Computer Mapping, Display, and Analysis
Version 2.1 User's Manual. Michigan State University, East Lansing.

Chen, Tao, and Mikio Takagi, 1993. The classification of NOAA images using multilayer
neural network. Procedings of the 6th Australian Joint Conference on Artificial
intelligence. Melbourne, November 1 6-1 9: 409-414.

Cihiar, J., L. St-Laurent and J. A. Dyer, 1991. Relation between the normalized
difference vegetation index and ecological variables. Remote Sensing of the
Environment, 35: 279-298.

Clark Labs, 1995. The ldrisi Project onscreen manual, ldrisi for Windows Version
1.01.002. Clark University, Worcester, MA.

Cohen, J., 1960. A coefficient of agreement for nominal scales. Educational and
Psychological Measurement, 20(1): 37-46.

Congalton, Russell 6., 1991. A review of assessing the accuracy of classifications of
remotely sensed data. Remote Sensing of the Environment, 37: 3546.

Decatur, Scott Evan, 1989. Application of neural networks to terrain classification.
Procedings of lntemational Joint Conference on Neural Networks, Washington,
0.0., 1: 283-288.

Dottavio, C. Lisette, and Darrel L. Williams, 1983. Satellite technology: An improved
means for monitoring forest insect defoliation. Journal of Forestry. January.

Duda, Richard O., and Peter E. Hart, 1973. Pattern classification and scene analysis.
John Wiley 8: Sons, New York, New York, 482 p.

Eastman, J. Ronald, 1992. IDRISI Technical Reference, Version 4.0 rev. 1. Clark
University, Worcester, Massachusetts, 213 p.

Ehrlich, D., J. E. Estes, and A. Singh, 1994. Applications of NOAA-AVHRR 1 km data
for environmental monitoring. lntemational Joumal of Remote Sensing, 15(1): 145-
161.

Eliot, Lance B., 1993. Prefilter your neurons. AI Expert, 8:9-11.

Fleiss, J. L., J. Cohen, and B. S. Everitt, 1969. Large sample standard errors of Kappa
and weighted Kappa. Psychological Bullefin. 72(5): 323-327.

Foody, Giles M., 1992. On the compensation for chance agreement in image

classification accuracy assessment. Photogrammetric Engineering & Remote Sens-
ing, 58(10): 1459-1460.

117

Foody, Giles M., 1995. Land cover classification by an artificial neural network with
ancillary information. international Journal of Geographical lnfonnation Systems.
9(5): 527-542.

Foody, Giles M., 1996. Relating the land-cover corhposition of mixed pixels to artificial
neural network classification output. Photogrammetric Engineering 8 Remote Sens-
ing, 62(5): 491 -499.

Foody, Giles M., Mary B. McCullough, and William B. Yates, 1995a. Classification of
remotely sensed data by an artificial neural network: Issues related to training data
characteristics. Photogrammetric Engineering 8 Remote Sensing, 61 (4): 391 -401.

Foody, G. M., M. B. McCullough, and W. B. Yates, 1995b. The effect of training set size
and composition on artificial neural network classification. Intemafi'onal Journal of
Remote Sensing, 16(9): 1707-1723.

Foschi, Patricia 6., and Deborah K. Smith, 1997. Detecting subpixel woody vegetation
in digital imagery using two artificial intelligence approaches. Photogrammetric
Engineering & Remote Sensing, 63(5): 493-500.

Garson, G. David, 1991. Interpreting neural-network connection weights. AI Expert, (6):
47-51.

Goward, Samuel M, Brian Markham, Dennis G. Dye, Wayne Dulaney, and Jingli Yang,
1991. Normalized difference vegetation index measurements from the Advanced
Very High Resolution Radiometer. Remote Sensing of the Environment, 35: 257-
277.

Green, Kass, and Brian Cosentino, 1996. Using satellite imagery to detect and monitor
forest change. Geo info Systems, 6(2): 22-30.

Hammerstrom, Dan, 1993a. Neural networks at work. IEEE Spectrum, 30(6): 26-32.

Hammerstrom, Dan, 1993b. Working with neural networks. lEEE Spectrum, 30(7): 46-
53.

Hara, Yoshihisa, Robert G. Atkins, Simon H. Yueh, Robert T. Shin, and J. A. Kong,
1994. Application of neural networks to radar image classification. IEEE
Transactions on Geoscience and Remote Sensing, 32(1): 100-109.

Harman, Jay, 1984. Environmental significance of a Great Lakes location. Chapter 5 in

Sommers, Lawrence M., Joe T. Darden, Jay R. Harman, and Laurie K. Sommers,
1984. Michigan, a Geography. Westview Press, Boulder, Colorado, 254 p.

118

Hepner, George E, Thomas Logan, Niles Ritter, and Nevin Bryant, 1990. Artificial
neural network classification using a minimal training set: Comparison to
conventional supervised classification. Photogrammetric Engineering and Remote
Sensing, 56(4): 469-473.

Huang, W.Y., and R. P. Lippman, 1988. Comparisons between neural net and
conventional classifiers. Proceedings of lntemational Joint Conference on Neural
Networks, San Diego, 4: 485-493.

Hudson, William, and Carl W. Ramm, 1987. Correct formulation of the Kappa coefficient
of agreement. Photogrammetric Engineering and Remote Sensing, 53(4): 421 -422.

Hush, Don R., and Bill G. Home, 1993. Progress in supervised neural networks:
What's new since Lippmann? IEEE Signal Processing Magazine, 10(1): 8-39.

Ingells, Norris, 1995. Gypsy moth numbers are down. Lansing State Journal.
January 4.

Ingells, Norris, 1998. Attack gypsy moth with vigor. Lansing State Journal,
March 18, 1-2 D.

Jain, Anil, 1994. Personal conversation.

Jensen, John R., 1996. introductory Digital image Processing: A Remote Sensing
Perspective, second edition. Prentice Hall, Upper Saddle River, New Jersey, 316 p.

Joria, Peter E., Sean C. Aheam, and Michael Connor, 1991. A comparison of the SPOT
and Landsat Thematic Mapper satellite systems for detecting gypsy moth defoliation
in Michigan. Photogrammetric Engineering a Remote Sensing. 57(12): 1605-1612.

Klimasauskas, Casimir C., 1992. Neural networks: an engineering perspective. lEEE
Communications Magazine, 30(9): 50-53.

Lawrence, Jeannette, 1991 . Data preparation for a neural network. AI Expert, 6(11):
34-41.

Leatherberry, Earl C., 1993. Forest statistics for Michigan’s Northern Lower Peninsula
Unit, 1993. Resource Bulletin NC-157. US. Department of Agriculture, Forest
Service, North Central Forest Experiment Station. St. Paul, MN, 48 p. Also
available via MDNR Forest Management Division [Online]
http:llwww.dnr.state.mi.usiwww/fmdifia/nIp/fianlp.html 3/23/98.

Leshkevich, George A., David J. Schwab, and Glenn C. Muhr, 1993. Satellite
environmental monitoring of the Great Lakes: A review of NOAA's Great Lakes
CoastWatch program. Photogrammetric Engineering & Remote Sensing, 59(3):
371 -379. Also [Online] http:llwww.glerl.noaa.govlctrb/ctrb_00811.html

Liebens, Johan, 1994. Procedures used in gypsy moth project. Unpublished
memorandum. January.

119

Liebhold, Andrew, Kurt Gottschalk, and Doug Mason, 1998. Evaluation of Forest
Susceptibility to the Gypsy Moth across the Coterminous United States. USDA
Forest Service, Northeastern Forest Experiment Station. [Online]
httpvlgypsyfsl.wvnet.edu/gmoth/suscept/suscept.html 3122/98.

Lillesand, Thomas M., and Ralph W. Kiefer, 1994. Remote sensing and image
interpretation, third edition. John Wiley & Sons, Inc., New York.

Lippman, Richard, 1987. An introduction to computing with neural nets. lEEE
Acoustics, Speech, and Signal Processing Magazine, 4: 4-22.

Lusch, David P., 1993. Spectral reflectance characteristics of vegetation, soil, and
water. Handout I in Lusch, DP. and Hudson, W.D., 1993, introduction to digital
remote sensing, Fort Worth, Texas. USDA Soil Conservation Service, National
Employee Development Office, 267 p.

Lusch, David P., 1994. Unpublished notes.
Lusch, David P., 1998. Personal communication.

Lusch, DP. and F.J. Sapio, 1994. Time-Series Analysis Applied to Mapping Gypsy
Moth Defoliation Using Multi-Date AVHRR Imagery. USDA, Forest Service,
State Cooperators Meeting. St. Paul, MN, March 2, 1994.

Lutzy, Ottmar, and Andreas Dengel, 1993. A comparison of neural net simulators.
lEEE Expert, 8(4): 43-51.

McConnack, Michael D., and Ron Day, 1993. How artificial intelligence impacts E & P
productivity. World Oil, 214(10): 81-87.

McKeown, David M., Jr., 1987. The role of artificial intelligence in the integration of
remotely sensed data with geographic information systems. lEEE Transactions on
Geoscience and Remote Sensing, 25(3): 330-348.

Ma, Zhenkui, and Roland L. Redmond, 1995. Tau coefficients for accuracy assessment
of classification of remote sensing data. Photogrammetn‘c Engineering 8 Remote
Sensing. 61(4): 435-439.

Markham, B. L., and J. R. G. Townshend, 1981. Land cover classification accuracy as a
function of sensor spatial resolution. Proceedings of the 15" lntemational
Symposium on Remote Sensing of Environment. Ann Arbor, Michigan, p. 1075-

1 090.

Michigan State University Extension, 1998. Built-in controls of gypsy moth - the 3 P's.

Michigan’s Gypsy Moth Education Program. [Online]
http:llwww.ent.msu.edu/gypsyed/docsienemies.html 3/31l98.

120

Mikkola, K., 1996. A remote sensing analysis of vegetation damage around metal
smelters in the Kola Peninsula, Russia. lntemational Journal of Remote Sensing.
17(18): 3675-3690.

Montgomery, B.A. (editor), 1988. Gypsy Moth in Michigan 1987. Extension Bulletin E-
2142, Cooperative Extension Service, Michigan State University, East Lansing.

Muchoney, Douglas M., and Bany N. Haack, 1994. Change Detection for Monitoring
Forest Defoliation. Photogrammetric Engineering 8 Remote Sensing, 60(10): 1243-
1251.

Murray, Dan, 1994. Tuning neural networks with genetic algorithms. Ai Expert, (9): 26-
31.

Olsson, H., 1993. Regression functions for multitemporal relative calibration of thematic
mapper data over boreal forest. Remote Sensing of the Environment. 46: 89-102.

Openshaw, Stan, 1994. Neuroclassification of spatial data. Chapter 3, in Neural Nets:
Applications in Geography, Bruce C. Hewitson and Robert G. Crane, eds. Kluwer
Academic Publishers, Boston.

Paola, Justin D., and Robert A. Schowengerdt, 1993. A review and analysis of neural
networks for classification of remotely sensed multispectral imagery. Research
institute for Advanced Computer Science, NASA Ames Research Center. Technical
Report 93.05, (NASA-CR-f 94291 ), June.

Paola, Justin D., and Robert A. Schowengerdt, 1995a. A detailed comparison of
backpropagation neural network and maximum-likelihood classifiers for urban land
use classification. lEEE Transactions on Geoscience and Remote Sensing, 33(4):
981-996.

Paola, Justin D., and Robert A. Schowengerdt, 1995b. A review and analysis of
backpropagation neural networks for classification of remotely sensed multispectral
imagery. lntemational Journal of Remote Sensing. 16(16): 3033-3058.

Paola, Justin D., and Robert A. Schowengerdt, 1997. The effect of neural-network
structure on a multispectral land-useiland-cover classification. Photogrammetric
Engineering 8 Remote Sensing, 63(5): 535-544.

Richards, J. A., 1986. Remote Sensing Digital image Analysis. Springer-Venag, Berlin.

Roesch Jr., Francis A., Paul C. Van Deusen, and Zhiliang Zhu, 1995. A comparison of
various estimators for updating forest area coverage using AVHRR and forest
inventory data. Photogrammetric Engineering 8 Remote Sensing, 61 (3): 307-311.

Rumelhart, D.E., and J. L. McClelland, 1986. Parallel distributed processing:
Explorations in the microstructure of cognition. MIT Press.

121

Sarle, Warren 8., ed., 1998a. Neural Network FAQ, Part 5 of 7: Free Software and Part
6 of 7: Commercial Software [Online] ftp://ftp.sas.comlpub/neuraVFAQ.html 3/22/98.

Sarle, Warren 8., ed., 1998b. Neural Network FAQ, Part 2 of 7: Leaming [Online]
ftp:l/'itp.sas.comlpub/neuraliFAQ.html 3/21/98.

Sarle, Warren 8., ed., 1998c. Neural Network FAQ, Part 3 of 7: Generalization [Online]
ftp'J/ftp.sas.comlpub/neurallFAQ.html 3/21/98.

Sapio, Frank, 1998a. Entomophaga maimaiga! Better late than neverl Michigan
Department of Natural Resources, Forest Management Division. [Online]
http:l/www.dnr.state.mi.uslwww/fmdipestlentom.html 3/17/98

Sapio, Frank, 1998b. Unpublished intemal documents. Michigan Department of Natural
Resources, Forest Management Division. Lansing, Michigan.

Schowengerdt, R. A., 1983. Techniques for image Processing and Classification in
Remote Sensing. Academic Press, New York, 249 p.

Skidmore, A. K., B. J. Tumer, W. Brinkhof, and E. Knowles, 1997. Performance of a
neural network: Mapping forests using GIS and remotely sensed data.
Photogrammetric Engineering 8 Remote Sensing, 63(5): 501 -514.

Sommers, Lawrence M., ed., 1977. Atlas of Michigan. Michigan State University Press,
East Lansing, Michigan, 242 p.

Sommers, Lawrence M., Joe T. Darden, Jay R. Herman, and Laurie K. Sommers, 1984.
Michigan, a Geography. Westview Press, Boulder, Colorado, 254 p.

Spanner, Michael A., Lars L. Pierce, David L. Peterson, and Steven W. Running, 1990.
Remote sensing of temperate coniferous forest leaf area index: The influence of
canopy closure, understory vegetation and background reflectance. lntemational
Journal of Remote Sensing, 11(1): 95-111.

Stuttgart Neural Network Simulator version 4.1, 1995. Provided via binary FTP transfer
from host: ftp.infonnatik.uni-stuttgart.de in lpub/SNNS as file SNNSv4.1.tar.Z.

Talerico, R.L., J. E. Walker, and T. A. Skratt, 1978. Quantifying gypsy moth defoliation.
Photogrammetric Engineering 8 Remote Sensing, 44(11): 1385-1392.

Todd, Stella W., and Roger M. Hotter, 1998. Responses of spectral indices to variations
in vegetation cover and soil background. Photogrammetric Engineering 8 Remote
Sensing. 64(9): 915-921.

Townshend, John R. G., Christopher 0. Justice, Charlotte Gurney, and James

McManus, 1992. The impact of misregistration on change detection. lEEE
Transactions on Gewcience and Remote Sensing, 30(5): 1054-1059.

122

USDA Forest Senrice, 1998a. E. Leopold Trouvelot - Perpetrator of Our Problem.
[Online] http:l/gypsy.fsl.wvnet.edu:80/gmoth/trouvelot/trouvelot.html 3/21/98.

USDA Forest Service, 1998b. Gypsy Moth Digest - Suppression Costs. [Online]
http:llwww.st.wvnet.edu/fl1p/gm_digest/costs.html 3/20/98.

USGS EROS Data Center, 1998. Advanced Very High Resolution Radiometer. [Online]
http:l/edcwww.cr.usgs.gov/glis/hyperlguide/avhrr 3/21/98.

Vennote, E., and Y. J. Kaufman, 1995. Absolute calibration of AVHRR visible and near-
infrared channels using ocean and cloud views. lntemational Journal of Remote
Sensing, 16(13): 2317-2340.

Wan, Eric A., 1990. Neural network classification: A Bayesian interpretation. IEEE
Transactions on Neural Networks, 1(4): 303-305.

Wann, M., T. Hediger, and N. N. Greenbaun, 1990. The influence of training sets on
generalization in feed-forward neural networks. Proceedings of lEEE Fourth
lntemational Conference on Neural Networks. 3: 137-142.

Woodcock, Curtis E., and Alan H. Strahler, 1987. The factor of scale in remote sensing.
Remote Sensing of Environment, 21 : 311-332.

123

 

IlullllljljlIljlllljllljljjll