\HHHH H W

 

This is to certify that the
dissertation entitled

CLASSIFICATION ACCURACY ANALYSIS OF SELECTED
LAND USE AND LAND COVER PRODUCTS IN A
PORTION OF WEST-CENTRAL LOWER MICHIGAN

presented by

Kin Man Ma

has been accepted towards fulfillment
of the requirements for the

PhD. degree in Geography

 

 

@WDCL/L/

"’Wr6fiessor’s Signature
//rfl 7

Date

MSU is an Affirmative Action/Equal Opportunity Institution

 

'flrr"??? a, .7"?
..._l L23""t,-“\I 1‘{
.r ! o”.
Itaciisdan otate

I}...' 1,‘ .. ‘1.“
’x--5I'.!'.:..ISL}/

 

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

UVA.

im 1: 2813
931 16

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/07 p:/ClRC/DaleDue.indd-p.1

CLASSIFICATION ACCURACY ANALYSIS OF
SELECTED LAND USE AND LAND COVER PRODUCTS
IN A PORTION OF WEST-CENTRAL LOWER MICHIGAN

By

Kin Man Ma

A DISSERTATION
Submitted to
Michigan State University

in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY
Department of Geography

2007

ABSTRACT
CLASSIFICATION ACCURACY ANALYSIS OF
SELECTED LAND USE AND LAND COVER PRODUCTS
IN A PORTION OF WEST-CENTRAL LOWER MICHIGAN
By

KinMan Ma

Remote sensing satellites have been utilized to Characterize and map
land cover and its changes Since the 1970s. However, uncertainties exist in
almost all land use and land cover maps classified from remotely sensed
images. In particular, it has been recognized that the spatial mis-registration of
land cover maps can affect the true estimates of land use / land cover (LULC)
changes. This dissertation addressed the following questions: what are the
Spatial patterns, magnitudes, and cover-dependencies of Classification
uncertainty associated with West-Central Lower Michigan’s LULC products and
how can the adverse effects of Spatial misregistration on accuracy assessment
be reduced? Two Michigan LULC products were Chosen for comparison: 1998
Muskegon River Watershed (MRW) Michigan Resource Information Systems
LULC map and a 2001 Integrated Forest Monitoring and Assessment
Prescription Project (IFMAP). The 1m resolution 1998 MRW LULC map was
derived from US. Geological Survey Digital Orthophoto Quarter Quadrangle
(USGS DOQQS) color infrared imagery and was used as the reference map,
since it has a thematic accuracy of 95%. The IFMAP LULC map was co-
registered to a series of selected 1998 USGS DOQQS. The total combined root
mean square error (rmse) distance of the georectified 2001 IFMAP was 112.20m.
A spatial uncertainty buffer of at least 1.5 times the rmse was set at 20m so

that polygon core areas would be unaffected by spatial misregistration noise. A

new Spatial misregistration buffer protocol (SPATIALM_BUFFER) was developed
to limit the effect of Spatial misregistration on Classification accuracy
assessment. Spatial uncertainty buffer zones of 20m were generated around
LULC polygons of both datasets.

Eight-hundred seventeen (817) stratified random accuracy assessment
points (AAPS) were generated across the 1998 MRW map. Classification
accuracy and kappa statistics were generated for both the 817 AAPS and 604
AAPS comparisons. For the 817 AAPS comparison, the overall classification
accuracy was 68.79% (kappa=0.627). When the 817 AAPS were overlaid onto
the 2001 IFMAP, 213 AAPS within the 20m spatial uncertainty buffer zone were
removed. The remaining 604 AAPS were used to assess the map accuracy and
results Showed that overall classification accuracy was 78.64% (kappa=0.742).
The residual, Spatial registration noise caused an overall thematic accuracy
underestimation of nearly 10%. Therefore, this SPATIALM_BUFFER method was
effective in reducing the effects of spatial misregistration on the accuracy
assessment of LULC maps.

However, even after removing misregistration noise from consideration, 102
out of 604 AAPS were still misclassified (16.9%). The following land cover
classes had the largest number of misclassified AAPS; grassland (12),
urban/built up (10), coniferous forest (10), non-forested wetlands (10), and
agriculture (9). Therefore, thematic misclassifications were still land-cover
dependent and can be influenced by classification system, radiometric,

temporal and phenological issues.

Cepyright by
KIN MAN MA
2007

Dedicated to my wife, Hyunsook Lee Ma,
and my children, Lindsay ShiShi and Joshua Inwoo

ACKNOWLEDGEMENTS

My doctoral dissertation committee has been instrumental in my
success. Dr. Jiaguo Qi, for his tremendous encouragement and endless
patience during this long process; Dr. David Lusch, for wonderful insights,
suggestions, encoUragement, and editorial insights; Dr. Jack Liu, for his
encouragement and humor; and Dr. Richard Groop, for his simple questions
and his willingness to serve on the committee at a late stage in the dissertation.

Great thanks to the following institutions who provided the various data
images for this dissertation. 1) Wexford, Mecosta, and Montcalm County
Extension Offices; 2) Michigan State University: Center for Global Change and
Earth Observations (CGCEO), Remote Sensing and Geographic Information
Science Research and Outreach Services (RS 8!. GIS), and Department of
Geography; 3) Grand Valley State University (GVSU)’S Annis Water Resources
Institute; 4) Michigan Department of Natural Resources (MDNR), Forest,
Mineral, and Fire Management Division. Also Special thanks to Frank Sapio,
Sherman Hollander, and Michael Donovan at MDNR for their assistance.

I joined the faculty of the Geography and Planning at GVSU back in
August 2001 and the following department colleagues have been very helpful
during the writing and thinking about the dissertation, Edwin Joseph, Laurie
Gasahl, Jeroen Wagendorp, Elena Lioubimtseva, and Wanxiao Sun. Other
GVSU administrators and colleagues across campus have also been very helpful
in technical, writing, and academic support: 1) Erika King, former Dean of
Social Sciences, 2) Ellen Schendel, Meijer Center for Writing and Michigan

Authors, 3) Catherine Frerichs, Pew Faculty Teaching and Learning Center, and

vi

4) Wendy Wenner, Dean of College of Interdisciplinary Studies, and 5) Nancy
Crittenden, Writing Consultant.

Also, tremendous thanks and appreciation to people at University
Reformed Church, East Lansing, MI: Kisoon Kim, Brinkman family, Mark
Whalon, Bruce Crew, Shu-Fen Kao, Tom and Stacey Bieler, Irv and Romelia
Widders, Pastor Tom and Joan Stark, John and Paula Holmes; Graduate
InterVarsity Christian Fellowship: Ralph and Utami DiCosty, Ted Watson,
Randy Gabrielse, Brian Hossink, Jenn V. Linton; MSU Geography fellow
students: Geoffrey Duh, Brian Becker, Ranjeet John, Stacey Nelson, Eraldo
Matricardi; First Reformed Church of Grandville, MI: Pastor Doug
VonBroanhorst, Barbara Boertje DeVries, Angie Walker, Pastor Christopher
Wolf, Pastor Mike Painter and family, and small group members: the Anderson,
Hasper, Dulirnarta, and Westrick families for their constant prayer support and
encouragement.

New York family: Ding th Ma, Lai Yung Ma, and Tony Ma for their
support and financial help at various times during this dissertation process;
Seoul, Korea family, mother-in-law (Kwak, Hae Ok) who has brought good food
and laughter to our home, Hyunja’s family, and also to Hyunjung, who always
brings a smile to my face.

My children, Lindsay (ShiShi) and Joshua InWoo, thanks for their
patience and perseverance for many late nights working on the dissertation.

Hyunsook: Greatest appreciation and thanks to my dear wife who has
had to manage our household of young Children during many, many long nights

at the office. She has challenged me with her Sharpening criticism and

encouragement to pursue the most efficient methods for conducting the
research.

The Lord God sustains and strengthens. May this research further the
knowledge of the natural world around us and help us to value and sustain it
for His purposes of preserving our land of Michigan and beyond. This
dissertation has been a challenging feat. May this quote express my reverence

for the Lord, God Almighty and his help in making this work possible.

“Unless the LORD builds the house, its builders labor in vain. Unless the
LORD watches over the city, the watchmen stand guard in vain. In vain you rise
early and stay up late, toiling for food to eat—for he grants sleep to those he
loves.” (Psalm 127: 1-2, NIV)

TABLE OF CONTENTS

LIST OF TABLES .............................................................................................. xi
LIST OF FIGURES ........................................................................................... xii
KEY TO ABBREVIATIONS ............................................................................... xiii
CHAPTER I
INTRODUCTION ............................................................................................... 1
1.1. Remote Sensing of Land Use and Land Cover ......................................... 1
1.2. Michigan’s Land Use and Land Cover ..................................................... 2
1.3. Classification Accuracy and Error .......................................................... 3
1.4. Research Questions ................................................................................ 4
1.5. Significance of the Study ........................................................................ 5
CHAPTER II
LITERATURE REVIEW ...................................................................................... 6
2.1. Remote Sensing of Land Cover ............................................................... 6
2.2. Classification Methods and Analysis ....................................................... 7
2.3. Classification Accuracy and Uncertainty ................................................ 9
2.3.1. Spatial Accuracy Assessment Protocol ............................................ 10
2.3.2. Spatial Accuracy and Resolution .................................................... 11
2.4. Michigan LULC Products ...................................................................... 13
CHAPTER III
RESEARCH DESIGN AND METHODOLOGY ................................................... 15
3.1. Study Area ........................................................................................... 15
3.2. Data Sources ........................................................................................ 15
3.2.1. 1998 Muskegon River Watershed LULC Map .................................. 16
3.2.2. 2001 IFMAP LULC Map .................................................................. 20
3.2.3. US. Geological Survey Digital Orthophoto Quarter Quadrangles... 21
3.3. Research Methods ................................................................................ 21
3.3.1. LULC Codes, Image Processing and Buffer Zone Method Protocol... 22
3.3.1.1. LULC Code System and Re-Coding Procedure .......................... 22
3.3.1.2. Image Processing of the 1998 MRW LULC map ........................ 25
3.3.1.3. Image Processing of the 2001 IFMAP LULC map ....................... 29
3.3.2. Choice of Accuracy Assessment Points ........................................... 32
3.3.3. LULC Map Comparisons and Error Matrix Evaluation .................... 33
3.3.4. Classification Uncertainty .............................................................. 36

CHAPTER IV

RESULTS AND ANALYSIS ............................................................................... 37
4.1. LULC Classification Comparison .......................................................... 37
4.1.1. 2001 IFMAP LULC Map .................................................................. 37
4.1.2. 1998 MRW LULC map .................................................................... 40
4.2. 2001 IFMAP/ 1998 MRW Error Matrix Analysis and Buffer Zone Method
Development ............................................................................................... 42
4.2.1. Error Matrix Analysis ..................................................................... 42
4.2.1.1. Misclassifications of Urban class AAPS ..................................... 48
4.2.1.2. Agriculture misclassified as Grass ............................................ 54
4.2.1.3. Misclassifications of Grass Class AAPS ..................................... 54
4.2.1.4. Shrubland misclassified as Deciduous Forest .......................... 58
4.2.1.5. Deciduous Forest misclassified aS Forested Wetlands .............. 58
4.2.1.6. Misclassifications of Coniferous Forest Class AAPS .................. 61
4.2.1.7. Forested Wetland misclassified as N onforested Wetlands ......... 62
4.2.1.8. Non-forested Wetlands misclassified as Forested Wetlands ...... 65
4.3. Issues Affecting Classification Accuracy Assessment ............................ 66
4.3.1. Classification System Issues ........................................................... 66
4.3.2. Scale Issues ................................................................................... 67
4.3.3. Radiometric Issues ......................................................................... 67
4.3.4. Temporal and Phenological Issues .................................................. 68
CHAPTER V
CONCLUSIONS AND
FUTURE RESEARCH ...................................................................................... 70
5.1. Conclusions ......................................................................................... 70
5.1.1. Spatial Patterns, Magnitudes, and Land-Cover Dependencies of
West-Central Lower Michigan LULC products .......................................... 70
5.1.2. Spatial Misregistration Buffer Protocol Development ...................... 71
5.2. Future Research ................................................................................... 72
REFERENCES ................................................................................................ 73

Table 3.1:

Table 3.2:

Table 3.3:

Table 4.1:

Table 4.2:

Table 4.3:

Table 4.4:

Table 4.5:

LIST OF TABLES

Image Data Sources .................................................................... 16
Land Use /Land Cover Code and Names ........................................ 22
Land Use Class Assignment Comparison ...................................... 24
LULC Classification Comparison .................................................. 38
2001 IFMAP Classification Accuracy (817 AAPS) BEFORE Bufier....45

2001 IFMAP Classification Accuracy (604 AAPS) AFTER Buffer ....... 45
2001 IFMAP / 1998 MRW LULC Map Error Matrix (817 AAPS) ....... 46

2001 IFMAP / 1998 MRW LULC Map Error Matrix (604 AAPS) ....... 47

LIST OF FIGURES

Figure 3.1: 1998 Muskegon River Watershed MIRIS LULC

Classification Map ....................................................................................... 18
Figure 3.2: 2001 IFMAP LULC Classification Map ......................................... 19
Figure 3.3: Research Methods Diagram ....................................................... 26
Figure 3.4: Distribution of 817 Accuracy Assessment Points Across

the MRW ................................................................................................... 34
Figure 3.5: Distribution of 604 Accuracy Assessment Points After Spatial
Buffering .................................................................................................... 35
Figure 4.1: LULC Classification Comparison ................................................. 39
Figure 4.2: IFMAP v. 1998 MRW Urban Classification Comparison ................. 41
Figure 4.3: Distribution of 102 Misclassified IFMAP AAPS ............................. 50

Figure 4.4: Misclassification of Urban (98MRW) as Agriculture Class
(IFMAP) .................................................................................................... 5 1

Figure 4.5: Misclassification of Urban (98MRW) as Deciduous Forest
Class (IFMAP) ............................................................................................ 53

Figure 4.6: Misclassification of Grass (98MRW) as Agriculture Class
(IFMAP) ..................................................................................................... 56

Figure 4.7: Misclassification of Grass (98MRW) as Decid Forest and Conif
Forest as Grass (IFMAP) ............................................................................. 57

Figure 4.8: Misclassification of Conif Forest (98MRW) as Forested
Wetlands (IFMAP) ....................................................................................... 60

Figure 4.9: Misclassification of Forested Wetlands (98MRW)
as Non—Forested Wetlands (IFMAP) .............................................................. 63

Figure 4.10: Misclassification of Non-Forested Wetlands (98MRW)
as Forested Wetlands (IFMAP) ..................................................................... 64

KEY TO ABBREVIATIONS

1998 MRW - 1998/ 1999 Muskegon River Watershed

AAPS — Accuracy Assessment Points

IFMAP - Integrated Forest Monitoring and Assessment and Prescription Project
Landsat ETM+ - Landsat Enhanced Thematic Mapper +

Landsat TM — Landsat Thematic Mapper

LULC — Land Use / Land Cover

MIRIS — Michigan Resource Information Systems

MRW - Muskegon River Watershed

RS 81» GIS —Remote Sensing and Geographic Information Science Research and
Outreach Services

USGS DOQQS — United States Geological Survey Digital Orthophoto Quarter
Quadrangles

CHAPTER I
INTRODUCTION

1. 1. Remote Sensing of Land Use and Land Cover

Humans have been the major agents of global land use and land cover
change Since the middle of the 19th century (Houghton and Woodwell, 1989).
Humans in the latter part of the 20th century have been dramatically
transforming the earth. For example, forested areas have been clearcut for
planting crops and agricultural lands have been cleared to build housing
developments. This change iS widespread though the magnitude of Changes iS
Challenging to quantify.

Satellite remote sensing provide the “ability to View the land use and land
cover patterns and conditions in a regional context, [which can] lead to
understanding the spatial relationships between different uses of the land”
(Loveland and DeFrieS, 2004). The United States’ Landsat satellite system of
sensors was launched in 1972 and many researchers have utilized archived
decadal images from the 1970S, 19805 and 19908 called the North American
Landscape Characterization (NALC) data series in order to research land cover
change over a thirty year period (Castelli et al., 1998; Nelson et al., 2002).
Landsat 7 Enhanced Thematic Mapper (ETM+) images Since 1999 have often
been used to map land cover and aid in environmental management (Aplin,

2004)

1.2. Michigan’s Land Use and Land Cover

Land Use / Land Cover (LULC) Classification can be used to Characterize
land cover at one point in time and a series of temporal LULC maps would Show
the LULC changes in a dynamic region. The Lower Peninsula of Michigan has a
tremendous diversity of land cover types, ranging from urban / built up land to
upland forests to agricultural regions to forested wetlands.

There was an historical effort in the mid-1970s to map all of Michigan’s
land use/land cover. This culminated in a 1978 statewide Land Use/Land
Cover map (State of Michigan, 2004). Recently, several Michigan counties such
as Wexford, Montcalm and Mecosta counties have updated their 1978 LULC
maps to 1998 by mapping out the changes in LULC (Wexford County Extension
Office, 2000; Montcalm County Extension Office, 2000; Mecosta County
Extension Office, 2000).

In addition, within the Muskegon River Watershed (MRW) in the west-
central region of Lower Michigan, a comprehensive LULC update has been
completed as part of an environmental monitoring project for this watershed
(Annis Water Resources Institute, 2004). This 1998 MRW LULC map spans
diagonally across 7,057km2 of the west-central Lower Michigan has a published
thematic accuracy of 95% based on a field verification of 5% of all polygons
within each township (RS 81. GIS, 2004).

Recently, the Michigan Department of Natural Resources (Michigan DNR)
completed a statewide land use / land cover map of Michigan called the
Integrated Forest Monitoring and Assessment and Prescription Project (IFMAP)
(Michigan DNR, 2004). This LULC map was for the 2000 / 2001 period.

Numerous agencies including the Michigan DNR rely on this dataset for current

land use / cover information, though the thematic accuracy of this map product

has not been independently verified.

1.3. Classification Accuracy and Error

When one classifies land cover, classification accuracy is a continual
concern because of issues of Spatial accuracy, and Spectral and phenological
differences between comparison images. When comparing LULC map products,
there is much uncertainty regarding the various reasons for misclassifications.
Congalton et al. (1983) utilized multivariate statistics to assess the
classification accuracy of maps derived from Landsat images. He has set a
standard by which remotely sensed Classification and accuracy assessment
should be used and has developed a solid research method design for limiting
classification error and increasing the accuracy of classified phenomena
(Congalton, 1991; Congalton and Green, 1999).

The spatial accuracy of vector and raster datasets is also important in
LULC map accuracy assessment. More than four decades ago, Perkal (1966)
suggested that an ‘epsilon’ distance should be defined around a cartographic
[polygon] line as a means of generalizing it objectively. Because of this concept
of the ‘epsilon’ zone, understanding the spatial accuracy or uncertainty of a
boundary and how it can affect the accuracy assessment of LULC maps is
important.

More recently, Stow (1999) stated that “[a]ccurate Spatial registration is
the most critical image processing requirement for reliable assessment of land

cover Changes that occur at Spatial scales that are Close to the characteristic

dimensions of the ground resolution element of the imagery used to assess the
changes (i.e., ‘pixel-Ievel’ change).” In addition, land cover “Change may be
overestimated due to positional errors in multi-temporal images, even with sub-

pixel co-registration [root mean square] rms errors” (Verbyla and Boles, 2000).

1.4. Research Questions

Various land use / land cover maps have been generated in recent years
to quantify the LULC and their changes within Michigan’s landscape, although
to what extent are these LULC classification map products accurate and
represent the true nature of Michigan’s LULC and its Changes remain unclear.

The following two research questions were addressed by this study:

1) What are the spatial patterns, magnitudes, and cover-dependencies of
classification uncertainty associated with west-central Lower Michigan’s land
use and land cover products in the past decade?

2) What methods can be developed to minimize the adverse effects of
Spatial misregistration on accuracy assessment of these LULC products?

The goals of this study were to understand whether specific land cover
types were more prone to classification uncertainty and also to what extent
spatial misregistration of LULC maps influences the nature of the uncertainty

and how this uncertainty may be dependent on particular land cover classes.

1.5. Significance of the Study

Researchers such as Verbyla and Boles (2000) estimated the extent to
which positional errors affect land cover change estimates and Salas et al.
(2003) demonstrated through a Perimeter/ Area ratio, that sliver polygons are
the result of spatial misregisration differences between LULC maps. More
recently, Wang 8:: Ellis (2005) computed false Changes by Shifting high
resolution ecological maps a series of intervals ranging from 5 m to 400 m and
quantifying the effects of positional error on change detection.

However, no research has taken Michigan LULC products and attempted
to develop a method to reduce the adverse effects of Spatial misregistration on
classification accuracy assessment. Also, no methods have been proposed or
tested that minimize the adverse effects of spatial misregistration on accuracy
assessment in Michigan LULC products. This study will fill in that gap by
providing a simple solution to minimize the effects of Spatial mis-registration on

the accuracy assessment of LULC maps.

CHAPTER II
LITERATURE REVIEW

2.1. Remote Sensing of Land Cover

Satellite remote sensing can observe and monitor large regions of the
earth because of the continual orbiting of various satellite systems, such as
SPOT (Satellite Pour l’Obseruation de la Terra), Moderate Resolution Imaging
Spectroradiometer (MODIS) aboard Terra, and Landsat. These satellites
provides the “ability to view the land use and land cover patterns and
conditions in a regional context, [which can] lead to understanding the spatial
relationships between different uses of the land” (Loveland and DeFries, 2004).
Through the use of sensors that are sensitive to specific wavelength bands of
the electromagnetic Spectrum, they also have the ability to detect reflected
infrared energy which aids in the study of vegetation condition, and short-wave
energy which aids in the understanding of vegetation moisture differences.

For example, one specific satellite system, the United States’ Landsat
system of sensors was first launched in 1972. Now, many researchers have
utilized the archived decadal images from the 19708, 1980s and 19908 called
the North American Landscape Characterization (NALC) data series to analyze
land cover change over a thirty year period (Castelli et al., 1998; Nelson et al.,
2002). During a Similar period, Since the 19808, remote sensing has helped
with land cover Classification and provided land cover inventories for improving
land management practices. Land cover information has also increased
environmental understanding of the earth by investigating the function of land

cover in environmental processes (Aplin, 2004). Since 1999, land cover

classification has utilized the latest medium scale Landsat 7 Enhanced

Thematic Mapper (ETM+) images with a 30 meters pixel resolution.

2.2. Classification Methods and Analysis

For land cover Classification, numerous algorithms have been developed
to attempt to decipher patterns in the landscape. Jensen (2005) reviewed the
main classification methods such as the supervised classification method. The
supervised classification method is advantageous when the researcher is
familiar with the research area. The researcher selects a representative group of
pixels (training pixels) that are identifiable as a specific land cover type, and
then choose a group of pixels that represents each of the assigned land cover
types. The Classification algorithm will then utilize the reflectance signatures of
the training pixels and then searches the entire remotely sensed image to
identify and find other pixels that match those of the land cover training pixels
(Jensen, 2005). On the other hand, the unsupervised ISODATA (Iterative Self-
Organizing Data Analysis Technique) Classification algorithm is advantageous
for the researcher who is not familiar with the landscape of the remotely sensed
image. This algorithm requires little human input. An ISODATA algorithm
classifies pixels and recalculates cluster mean vectors and assigns them to the
desired number of clusters. These Clusters are then assigned to the specific
LULC type within a specific LULC Classification scheme (Jensen, 2005).

A series of classification methods and algorithms have been developed to
improve classification accuracy. Bruzzone and Cossu (2002) had combined both

“multiple classification algorithms and multitemporal imagery in a new

cascading classification structure to enable the regular updating of land cover
maps” (Aplin, 2004).

For another Classification algorithm, Sohn et al.’s (1999) cosine of the
angle concept (CTAC) classifier showed the power of this method to classify
different shapes in the landscape. When classifying the secondary succession
forest and agricultural land types of the north-central Yucatan, the CTAC
method utilized the spectral angle of the different successional forest lands to
differentiate the slight variations between the forest and agricultural land cover
types. In a subsequent study, this CTAC method was then evaluated against
other unsupervised and supervised land cover Classifiers. Sohn 85 Rebello’s
(2002) research results demonstrated that the CTAC was very effective in
delineating the Spectral Shape patterns of different land-cover/land—use types.

The above Classification algorithms assume a binary type of land cover
classification. However, in a diverse landscape, the boundary between the two
land cover class types, such as the edge between comfields and forested land, is
difficult to classify accurately. Classification depends on the Spatial resolution,
and the image pixel that has spectral reflectance from the corn and forest land
cover types would produce a “mixed” pixel of reflectance. It is neither 100%
com, nor 100% forest. Therefore, there has been a trend in classification to
expand beyond an either/ or dichotomy of land cover Classes, but to utilize a
“fuzzy Classification” in which the land cover can be classified a specific
percentage of a particular land cover class, such as 30% com (Foody, 1999).

Moreover, when evaluating several classification algorithms, the costs of
accuracy assessment play a large role in the type of algorithm utilized and the

investment in field checking of reference data points was important. They

proposed a protocol for “assessing the quality of Classification results, related to
the economical aspects of a specific remote-sensing project” (Smits et al., 1999).

When classification algorithms are evaluated and compared, the kappa

A

coefficient, k statistic, iS the most common and effective statistic utilized to

compare classification accuracy as show in the equaltion below (Lillesand and
Kiefer, 2000; Congalton and Green, 1999; Smits et al., 1999).

observed __ accuracy — chance _ agreement

 

,;=

1 — chance __ agreement

2.3. Classification Accuracy and Uncertainty

When one classifies land cover, classification accuracy is a continual
concern because of issues of spatial accuracy, and spectral and phenological
differences between comparison images. A representative number of ground
truth points are needed within the various land cover types in order to produce
an accurate classification map from remotely sensed data. Congalton et al.
(1983) utilized multivariate statistics in order to evaluate classification accuracy
of Landsat imagery. Congalton (1991) set the standard by which remotely
sensed classification and accuracy assessment should be undertaken and has
developed a solid research method designed for limiting Classification error and
increasing the accuracy of classified phenomena (Congalton, 1991; Congalton
and Green, 1999).

Classification accuracy has often been adversely affected by the problem
of “mixed” pixels that are on the edges of two different land cover types. This

problem of the “mixed” pixel will not be easily resolved, though by

understanding how this problem is linked to pixel size and resolution, will help
the interpretation of classification results.

In addition, Foody (2002a) systematically reviewed the current status of
classification accuracy assessment. He cautioned that hard binary Classification
systems “would be expected to underestimate the area of land undergoing a
change and, where a Change is detected, overestimate the magnitude of change
as it is a simple binary technique” (Foody, 2002a). The traditional approach to
accuracy assessment involves three primary components, the sample design
that determines which subregions or points will be sampled, the response or
measurement design to obtain the true or “reference” attribute for each sampled
unit or point, and the analysis of the data obtained (Stehman and Czaplewski,

2003I

2.3.1. Spatial Accuracy Assessment Protocol

When evaluating map accuracy, Biging et al. (1998) reviewed various
sampling method protocols for effective accuracy assessment. They concluded
that a stratified random sampling protocol should be used to increase the
precision over a simple random sampling protocol technique (Biging et al.,
1998). Nusser and Klaas (2003) demonstrated the method of allocating a set of
stratified random “reference” points across their Iowan research fields. They
randomly chose ground truth points from their 29 land cover type categories,
though in their accuracy analysis they took the “important step of calculating
standard errors for the [accuracy] estimates” (Stehman and Czaplewski, 2003).

Patil and Taillie (2003) studied the differences between two classified

maps having the same spatial extent. Overlaying the maps yielded a

10

dissimilarity matrix. They developed a “parametric model, called latent truth
model, by Specifying the dissimilarity matrix in terms of the true proportions for
the mapping categories as well as the unknown error rates for the two maps”
(Patil and Taillie, 2003). They then generated accuracy assessment
characteristics for the two maps from Wicomico County, Maryland.

However, Stehman (2001) had demonstrated that there needs to be a
balance of statistical rigor and practical utility in thematic map accuracy
assessment, the reporting of standard errors is necessary to evaluate precision
and the sampling protocols of assessment points must be Clearly documented.
Practical utility was framed in terms of tradeoffs between cost versus quality in
which four quality criteria must be evaluated while considering the ‘resources’
available to the researcher: “precision, population representation, assumptions,
and accuracy of the reference data” (Stehman, 2001).

When comparing LULC map products, there iS much uncertainty
regarding the various reasons for misclassifications. Foody and Atkinson (2002)
stated that, “we are a long way from having an ‘uncertainty button’ in our
geographic information systems [software]. Indeed, we are a long way from
appreciating fully uncertainty and its impacts... [There are also] “fundamental

spatial relations, such as neamess” that have not been fully explored (Foody

and Atkinson, 2002).

2.3.2. Spatial Accuracy and Resolution
Accuracy assessment can also be influenced by pixel resolution. The
Spatial scale or pixel resolution can make a Significant impact on the

phenomena being observed. Openshaw (1984) first discussed the modifiable

11

areal unit problem (MAUP) and showed that if the patch Size of the
phenomenon is small, the pixel resolution needs to match this Size in order to
be effective in measuring it. Cao and Lam (1997) stated that “the scale problem
arises because of the uncertainty about the number of zones most appropriate
for a particular study, and the aggregation problem arises because of the
uncertainty about how the data are to be aggregated to form a given number of
zones” (Cao and Lam, 1997).

In addition to spatial resolution, the Spatial accuracy of vector and raster
datasets are also important in LULC map accuracy assessment. Perkal (1966)
suggested that an ‘epsilon’ distance should be defined around a cartographic
[polygon] line as a means of generalizing it objectively. Because of this concept
of the ‘epsilon’ zone, understanding the spatial accuracy or uncertainty of a
boundary and how it can affect the accuracy assessment of LULC maps is
important.

More recently, Stow (1999) stated that “[a]ccurate spatial registration is
the most critical image processing requirement for reliable assessment of land
cover changes that occur at Spatial scales that are close to the characteristic
dimensions of the ground resolution element of the imagery used to assess the
changes (i.e., ‘pixel-level’ Change)” More specifically, Dai and Khorram (1997)
showed “that highly accurate Change detection based on multi-temporal
Landsat Thematic Mapper images requires that the magnitude of mis-
registration be less than 0.2 pixel.”

Verbyla and Boles (2000) found that land cover “Change may be
overestimated due to positional error in multitemporal images, even with sub-

pixel co—registration rmS errors” (Verbyla and Boles, 2000). Moreover, Salas et

12

al. (2003) tried to improve land cover change estimates by employing a
perimeter/ area ratio as an index of misregistration bias. They found that
“almost 10% of the total change from cropland to vegetated areas [in
Guangdong, China] could be due to Single pixel misregistration” (Salas et al.,
2003)

Wang and Ellis (2005) computed false changes by Shifting high
resolution ecological maps a series of intervals ranging from 5 m to 400 m and
quantifying the effects of positional error on change detection. Bruzzone and
Cossu (2003) proposed an adaptive approach aimed at reducing the effects of
registration noise in unsupervised change detection. They developed an
adaptive and nonparametric estimation of the distribution of registration noise
in a magnitude-direction space through a complex, mathematical formulation
involving the change vector analysis (CVA) technique (Bruzzone and Cossu,
2003)

None of these studies proposed an effective method of reducing the

effects of misregistration noise on thematic accuracy assessment.

2.4. Michigan LULC Products

There was an historical effort in the mid-19708 to map all of Michigan’s
land use/land cover. This culminated in a 1978 statewide Land Use/Land
Cover map (State of Michigan, 2004). Recently, several Michigan counties such
aS Wexford, Montcalm and Mecosta counties have updated their 1978 LULC

maps to 1998 by mapping out the changes in LULC (Wexford County Extension

13

Office, 2000; Montcalm County Extension Office, 2000; Mecosta County
Extension Office, 2000).

In addition, within the Muskegon River Watershed (MRW) in the west-
central region of Lower Michigan, a comprehensive LULC update has been
completed as part of an environmental monitoring project for this watershed.
This 1998 MRW LULC map has a published thematic accuracy of 95% based on
a field verification of 5% of all polygons within each township (RS 85 GIS, 2004).

Recently, the Michigan Department of Natural Resources completed a
statewide land use / land cover map of Michigan called the Integrated Forest
Monitoring and Assessment and Prescription Project (IFMAP) (Michigan DNR,
2004). This LULC map was for the 2000 /2001 period. The thematic accuracy of
this map product has not been independently verified.

After reviewing the literature, it became Clear that a comprehensive study
of the Spatial patterns, magnitudes, and cover-dependencies of classification
uncertainty was lacking. The literature review also revealed that all LULC map
products exhibit some degree of spatial inaccuracy and this spatial mis—

registration noise confounds traditional accuracy assessment procedures.

14

CHAPTER III
RESEARCH DESIGN AND METHODOLOGY

3.1. Study Area

The Muskegon River Watershed extends diagonally across the west-
central portion of Lower Michigan encompassing parts of eleven counties. There
is a tremendous diversity of land use / land cover types within the 7,057 km2 of
the watershed, ranging from urban to upland forests to agricultural regions
(Figure 3.1). This area was chosen because the LULC of this diverse landscape
was mapped by high spatial resolution imagery and its 95% thematic accuracy
makes it a good candidate for a reference map to assess the accuracy of the
2001 Integrated Forest Monitoring and Assessment Project (IFMAP) LULC

product referenced below in section 3.2.2.

3.2. Data Sources

For Michigan’s Lower Peninsula, several LULC map products have been
produced over the last decade. Two of the available LULC map products were
chosen, 1) the 1998 Muskegon River Watershed LULC map, and 2) the 2001
IFMAP LULC map (see Table 3.1). These maps and corresponding legends Show
land use / land cover types and are displayed in a variety of colors. Many

images in this dissertation are presented in COLOR to differentiate the LULC

types.

15

Table 3.1: lmgge Data Sources

 

 

 

 

 

 

 

 

 

 

 

 

Data Source Image Source Date of Imagery Spatial
Resolution

selected
USGS DOQQs CIR Orthophotography April 1998 1 m
1998/1999 MRW CIR Orthophotography April 1998, April 1999 1 m
2001 IFMAP Landsat TM and ETM+ Multiple Dates (see below) 30 m

Path22, Row30 25-Mar-2000 Leaf Off (ETM+)

Path22, Row30 13-Jul-1999 MidSeason (ETM+)

Path22, Row30 11-Oct-00,19-Oct-00 (TM, ETM+)

Path22, Row29 26-Apr-2000 Leaf Off (ETM+)

Path22, Row29 2-Jul-2001 MidSeason (ETM+)

Path22, Row29 19-Oct-00 (ETM+)

Path21, Row 29 8-Mar-2001 Leaf Off (ETM+)

Path21, Row 29 25-Jun-2001 MidSeason (ETM+)

Path21LRow 29 29-Sep-2001 (ETM+)

 

3.2.1. 1998 Muskegon River Watershed LULC Map

The 1998 MRW LULC map was derived from 1 meter pixel resolution
color infrared aerial imagery acquired in April 1998 and April 1999 by the
National Aerial Photography Program (NAPP), and also known as the US.
Geological Survey’s Digital Orthophoto Quarter Quadrangle images (RS 8:, GIS,
2004). This LULC classification was performed by the Remote Sensing and
Geographic Information Science Research and Outreach Services (RS 81. GIS),
Michigan State University for the Annis Water Resources Institute (AWRI),
Grand Valley State University. The interpretation of land use / land cover classes
to Level III of the Michigan LULC classification system was performed by two
separate interpreters (RS 8; GIS, 2002). “The [1998 MRW] polygons were at a
scale of 1:24,000 and equates to a Spatial accuracy of 21:12. 19 meters and there
was a 95% overall classification accuracy” (RS 85 GIS, 2004). Also, “5% of

polygons within each township were field verified, especially unknown polygons,

16

ambiguous polygons and random polygons” (RS 8f. GIS, 2004). Field verification
was performed by the staff of the RS 8:. GIS, and the AWRI. This vector polygon
dataset was in a Michigan Georef projection (Figure 3.1). In addition, the
“minimum mapping unit (MMU) was 2.5 acres or 1 ha and the minimum
mapping distance (MMD) was 100 feet (30.48 meters)” (RS 8f. GIS, 2004).

This LULC dataset is the most comprehensive, high-resolution LULC
dataset available for west—central Lower Michigan. Because this “wall to wall”
LULC dataset encompassed more than 7,000 km? of diverse landscape and was
systematically field Checked, it can be considered a ground “truth” map for
verifying and evaluating coarser resolution datasets of the same area.
Congalton and Green (1999) noted that photo interpreted images can be used in
place of true ground reference data points, if a subset of the land area has also
been field verified. The 1998 MRW LULC map meets this criterion and was used

to evaluate the accuracy of the 2001 IFMAP LULC product.

17

 

Figure 3.1: 1998 Muskegon River Watershed

Co L, . .
19 MletnlglessMap MIRIS LULC CIaSSIfication Map

Urban

- rass 0 "‘
Decid. Forest

  

 

- Conif. Forest ‘
- Mixed Forest
Clearcut Forest

- Water
- Forest Wetlands

 

! Barren

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

A
30 0 30 60 Kilometers
F L
1 1/19/2006
Kin Ma
W fl I l

 

- Nonfor. Wetlands ——

 

 

18

 

 

Figure 3.2: 2001 IFMAP LULC Classification Map

County Lines
2 1 IFMAP LULC

- Urban

2 Agriculture

F1 Grass

- Shrubs

- Decid. Forest
Conif. Forest

[ '| Mixed Forest

- Water
Forest Wetlands

:. ‘; Nonfor. Weflands

- Barren

11/1612006
Kin Ma

v

0

 

 

 

 

 

 

 

 

 

 

3O 0 30

 

 

 

 

 

Kilometers N

 

 

 

19

 

3.2.2. 2001 IFMAP LULC Map

The Integrated Forest Monitoring and Assessment Project (IFMAP) 2001
LULC dataset was acquired through the Michigan Department of Natural
Resources (Michigan DNR), Resource Mapping and Aerial Photography Section
and provided in Geotiff format, with a 30m cell Size in a Michigan Georef
projection. This LULC map was derived from a series of Landsat TM and ETM+
images for both the early spring, mid-summer, and early fall periods (Table 3.1).
This classification emphasized forest inventory and classification and, according
to the 2004 IFMAP report, had a Classification accuracy of 87.7% at Level II
utilizing the 2003 Forest Inventory and Assessment (USFS, FIA) data points for
classification verification (Michigan DNR, 2004). The 2001 IFMAP LULC map of
the entire state of Michigan was completed in 2004. Numerous agencies
including the Michigan DNR rely on this dataset for current land use / cover
information, but there has not been an independent verification of its
classification accuracy.

The IFMAP classification was generated in the following manner and the
entire classification procedure paragraph is quoted from the 2004 IFMAP
report:

The IFMAP “primary classifier was a cluster analysis method developed
by Chuveico and Congalton (1988), which matched Clusters from an
unsupervised Classification with the training Site data collected by the field
crews. In this process, several outcomes for a given Cluster were possible. The
cluster analysis method was an iterative process, whereby clusters were labeled
with a land cover class based on their spectral distance from a given set of
training statistics. In the first iteration, the summer image mosaic was
classified using the ISODATA unsupervised method. The resulting classes were
grouped with the supervised classes using an agglomerative hierarchical

procedure (Mathsoft, 2000). This procedure began by considering each
signature as a separate group, then it combined and divided groups based on

20

spectral similarity until all signatures were in a single group, displayed in a
hierarchical structure according to the order in which the groups were merged
or divided. The resulting clustering tree was then examined for matches of
supervised and unsupervised signatures, and in the case of a close one-to-one
relationship, the unsupervised signature in question was labeled with the land
cover label of the supervised signature. After the tree was fully examined for
clusters of this type, the labeled classes were subset fi:om the imagery, and the
procedure was run again. This procedure continued until it was no longer
advantageous to do so. The Cluster analysis method was used initially on the
entire image to achieve an Anderson Level 1 classification, and subsequently on
a subset of the image for each individual Level 1 cover type” (MDNR, 2004, p.
19).
3.2.3. US. Geological Survey Digital Orthophoto Quarter Quadrangles

Selected US. Geological Survey Digital Orthophoto Quarter Quadrangles
DOQQS) for Michigan were obtained from the Center for Geographic
Information, Geographic Data Library (Michigan CGI, 2006). For the 1998 date
series, DOQQS are available for the entire state of Michigan, have a 1 meter
pixel size and are in the Michigan GeoRef projection format. Several of these
images were Chosen to georectify the 2001 IFMAP LULC map. Four DOQQS were
selected to cover the four areas within the expanse of the Muskegon River
Watershed. For example, the DOQQS named Houghton Lake SE, Cadillac North
SE, Twin Lake SE, and Dalton NE, covered the following four regions: 1) the
northeast region near Houghton Lake, 2) the western region near Cadillac, MI in

Wexford County, 3) the central region of Mecosta County near Rodney, MI and

4) the southwestern region north of the city of Muskegon.

3.3. Research Methods

There has been increasing evidence that Spatial misregistration can
negatively affect the overall classification accuracy assessment of LULC map

comparisons (Stow, 1999; Verbyla and Boles, 2000; Salas et al., 2003). A

21

method was developed to eliminate the adverse effect that spatial misregis-
tration has on classification accuracy assessment. Within the following image
processing protocol below, a spatial accuracy and misregistration buffer zone
method was developed to prevent accuracy assessment points (AAPS) from being
positioned in the zone of uncertainty due to spatial misregistration between the

two LULC maps.

3.3.1. LULC Codes, Image Processing and Buffer Zone Method Protocol

3.3.1.1. LULC Code System and Re-Coding Procedure

Since the 1998 Muskegon River Watershed (MRW) LULC map had 1 m
resolution and also an overall accuracy of 95%, this map became the reference
map for evaluating the classification accuracy of the 2001 IFMAP LULC
product. The land use categories of the 1998 MRW LULC map had been
classified to Level III of the Michigan system (RS 86 GIS, 2002). In this
dissertation, these LULC classes were re-coded to conform to the generalized
Level I / 11 combined LULC codes shown in Table 3.2. Land Use/ Land Cover
types from the 2001 IFMAP LULC product were also re-coded to the same

modified Michigan Level I /II LULC Classification system (Table 3.2).

Table 3.2: Land Use / Land Cover Code and Names

 

 

 

 

 

 

 

 

 

 

 

LU Code LU Class Name LU Code LU Class Name
1 Urban 7 Mixed Forest
2 Agriculture 8 Clearcut Forest
3 Grass 9 Water
4 Shrubs 10 Forested Wetlands
5 Deciduous Forest 11 Nonforested Wetlands
6 Coniferous Forest 12 Barren Land

 

 

22

The Specific re-code relationships for both the MRW and IFMAP LULC
maps are shown in Table 3.3. Note that the category of Clearcut forest, MRW
class #44 (a classification of forest that has been Cleared to below 16.7%
stocking density) was not mapped by IFMAP. Because of this discrepancy
between the two classification systems, the clearcut forest category was not

utilized in any of the LULC comparisons.

23

Table 3.3: Land Use Class Assignment Comparison

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

LU LU Class 2001 IFMAP LU_Names 1998 MRW MIRIS Classes
1 Urban Low Intensity Urban Residential Housing
1 Urban Airports Commercial
1 Urban Roads / Paved Industrial
1 Urban High Intensity Urban Transport/Utilities
1 Urban Extractive
1 Urban Parks I Golf Courses Open Land
2 Agriculture Non-vggetated Farmland Cropland
2 Agriculture Row Crops
2 Agriculture Forage Crops

_2_ Agriculture Orchards/ Vineyards Orchards
2 Agriculture Confined Feeding
2 Aflulture Permanent Pasture
2 Agriculture Other Agic. Land
3 Grass Herbaceous Openland Grasses and Forbs
4 Shrub Upland Shrub Shrub Open Field
5 Decid. For. Northern Hardwoods Northern Hardwoods (#411)
5 Decid. For. Oak Association Central Hardwoods/Oak (#412)
5 Decid. For. Aspen Association Aspen/Birch (#413)
5 Decid. For. Mixed Upland Deciduous
6 Conif. For. Pines Pines (#421)
6 Conif. For. Other Upland Conifers Other Upland Conifers (#422)
6 Conif. For. Mixed Upland Conifers Christmas Tree Plantation (#429)
7 Mixed For. Upland Mixed Forest Mixed Forest (#43)
8 Clearcut Forest NOT in classification Clearcut Forest (#44)
9 Water Water StreamsNVaterways
9 Water Lakes/Ponds
9 Water Reservoir

 

 

 

 

 

 

 

 

 

 

 

 

 

 

10 For. Wetlands Lowland Decid. Forest Lowland Decid. Forest (#414)
10 For. Wetlands Lowland Conif. Forest Lowland Conif. Forest (#423)
10 For. Wetlands Lowland Mixed Forest Forested Wetlands (#61)

10 For. Wetlands Lowland Shrub

11 Nonfor. Wetlands Floating Aquatic Nonfor. Wetlands (#62)

11 Nonfor. Wetlands Emergent Wetland Aquatic Bed (#621)

11 Nonfor. Wetlands Mixed NonForest Wetland

12 Barren Sand / Soil Beach/Riverbank (#72)

12 Barren Bare/Sparsely Veg Sand Dune (#73)

12 Barren Exposed Rock Exposed Rock (#74)

 

LU = Land Use Code

LU Class = Land Use Class Names
2001 IFMAP LU_Names = 2001 IFMAP Land Use Class Names

1998 MRW MIRIS Classes = 1998 MRW MIRIS Land Use Class Names

24

 

3.3.1.2. Image Processing of the 1998 MRW LULC map

The 1998 MRW LULC file was originally in an ArCView Shapefile format.
Level I / 11 land cover codes were assigned as shown in Table 3.3 and an ArCInfo
GRID command (shapegrid) converted the shapefile to a 30m cell size raster file
to match the Spatial resolution of the 2001 IFMAP LULC map (Figure 3.3). It was
then converted from a grid file to a 30m pixel Size image file for processing in
ERDAS Imagine software.

This image (.irng file) was subset into three equal parts, 1) top, 2) middle,
and 3) bottom, because the ArcGIS vectOr buffering function had difficulty
processing the large number of polygons within the 1998 MRW vector coverage
(a geoprocessing necessity due to computational limitations). After subsetting,
the three subsection images were each converted back to a grid format and

subsequently transformed into three vector polygon coverages (see Figure 3.3).

3.3. 1.2a. 1998 MRW Buffer Zone method

The spatial accuracy of the 1998 MRW dataset, as provided by the
image’s metadata, was 1:12.19m. The spatial accuracy of the IFMAP image (see
section 3.3.1.3, below) was calculated as a root mean square error (rmse)
distance of :t0.6m relative to the 1998 MRW LULC map. Therefore, the total
combined rmse distance was i12.20m. This distance would be the minimum
buffer zone of spatial misregistration between these two datasets. As a further
safeguard, the Spatial uncertainty zone was increased to 20m, at least 1.5 times
the rmse distance, in order to be certain that the AAPS would fall within the

polygon core areas that were unaffected by spatial misregistration noise.

25

 

Eamon!

 

 

, 35338 0:08 3.95 :58
muonoowhsw Eofiom
.33.: .89 E 88

 

 

 

 

 
 

. spa—Co down—oh SB
- - 0 3H8 Hun—b
— an: manna: 55:5 5... ”WM—ode". ”9681

 

 

nan. coed: spoon u was 5 coma
:32: 3 32 .3 «E 816
sea a mo sneak—ea 3g. Eat noon

\.

Q

 

$8 coma—ca
Deuces. m A. meta So»... 35 20:03
tgeoo "£592 Season .2828 as.

conned .833 m 35 Roman: «saw ”on? «Eu cued:

.Eaenoatfiaoo

 

 

 

_ Leakage... U144 New.“ .32.... In...“
Banana 32502 gonad 5.6 can—arm

 

 

26

 

2: .3: game .24 83.3% 2.23 a.
£53m >§Sw33§3ha§aqsfia
{24% 3.2483 33843389330233:
_ ,_ . 0.333“

 

 

 

 

  

Econ beam 22...:
55.? 8oz 35 S 63.5

 

 

     
        

, £5. 585m
8535 Sm 838qu

 

888 823 oz ,
.093: E

 

888 beam :5 can:

 

 

 

 

    

 

     
    

 

. 83: e335 ,
D82 .8302 owes: «new
Eon 2.2a. ucbsm E - . ., .
l swan: 0282 2
Oman 9539a Eob owin— 8 3.5230 8 swan @5255 Est
895 3 annex etc a
8.509% 0:08 Leann EON ..
enouoom- ham 883m
.0233 88.323 88 ....,

 

8. :30. 32:02 3308M m. a. 3:333

27

The ArcGIS buffer function was used to calculate a 20m buffer zone
inside each LULC polygon. After the buffer function calculated this 20m zone
within every LULC polygon, a shapefile of areas labeled ‘—20’ m around each
polygon edge boundary was generated and converted to a 30m cell size grid
using the ArcInfo ‘shapegrid’ function. Unfortunately, the ‘-20’m buffer grid had
‘floating point’ values that needed to be rounded to integers. This conversion
was accomplished using an ArCView extension, named “Grid Round-Floating
Point to Integer,” that was downloaded from the http: / /www.esri.com website
(Corradini, 2006). Each of the 1998 MRW’S three sub-section grids were
rounded to a ‘-20’ value integer and then multiplied by ‘-10’ so that the
resulting grid would have a positive value of ‘200’. This conversion was
necessary since grids having negative values cannot be imported by the ERDAS
Imagine software program. After each of the three ‘200’ value sub-section buffer
grids were imported into ERDAS Imagine, the three image files for the top,
middle and bottom sections of the MRW, were mosaicked together to form a one
seamless image that had 40-meter-wide epsilon zones centered along the
boundaries of each LULC polygon (i.e., group of pixels) (Figure 3.3, 2nd page).

The 1998 MRW ‘buffer zone’ image described above was used to mask
out the uncertainty zone within the 1998 MRW LULC map. After these
processing steps, a 1998 MRW ‘NO buffer zones’ image was created and ready
for error matrix evaluation with the 2001 IFMAP ‘NO buffer zones’ image (see

Figure 3.3, 2nd page).

28

3.3.1.3. Image Processing of the 2001 IFMAP LULC map

The 2001 IFMAP LULC map was co-registered to the above referenced
USGS DOQQS. Within each of the four selected DOQQS, a Ground Control Point
(GCP) was chosen at the corner of a coniferous forest“ stand that was viewed in
the selected DOQQ images, the 1998 MRW and 2001 IFMAP images. The 2001
IFMAP image was rectified to the 1998 MRW image using the four Chosen GCPs
within the DOQQ images, the rmse distance for the 2001 IFMAP image was
$0.6m. Therefore, the total combined rmse distance for the 2001 IFMAP image
relative to the 1998 MRW image was 1:12.20 m.

After geometric registration, the 2001 IFMAP was subset into three equal
parts, 1) top, 2) middle, and 3) bottom by utilizing the same area of interest
(AOI) boundaries that were used to subset the 1998 MRW image. This
subsetting was necessary because the ArcGIS vector buffering function had
difl'rculty processing the very, large number of polygons (810,000+) generated by
the entire IFMAP image of the MRW. After subsetting, each of the three IFMAP
subsection images were converted to grid files and subsequently converted to
vector polygon coverages (see Figure 3.3). Even after splitting the IFMAP image
into 3 subsections, each of the subsections still had too many polygons to
effectively run the ArcGIS buffering function (e.g., the IFMAP bottom subsection
coverage had 241,000+ polygons).

Since the original IFMAP LULC map was derived from Landat TM and
ETM+ images, the classification generated numerous 30m single-cell polygons.
Using Arclnfo, these small polygons were eliminated to match the minimum
mapping unit Classification of the 1998 MRW LULC map, (1 ha or 10,000 m2).

The equivalent of 3 x 3 30m cell size group of polygons, or 8,100 m2, were

29

eliminated from each of the three IFMAP section vector coverages. Nearly 90%
of the polygons within each of the section coverage files were equal to or smaller
than the 8,100 m2 minimum mapping unit. The LULC labels of the small
polygons were re—assigned to the LULC label of the closest, adjacent polygon
that was larger than 8,100 m2, and the boundary lines of the small polygon
shapes were dissolved (Figure 3.3). The resulting spatially, filtered coverages

had fewer than 27,000 polygons.

3.3. 1.3a. IFMAP Buffer Zone method

The rrnse distance for the IFMAP image (1:12.20m) would be the
minimum buffer zone of Spatial uncertainty for this dataset. The uncertainty
zone due to spatial misregistration was increased to 20m, at least 1.5 times the
rrnse distance, in order to be certain that the AAPS would fall within the polygon
core areas that were unaffected by spatial misregistration noise.

The ArcGIS 9.0 buffer function was used to calculate a 20m buffer zone
inside each LULC polygon in the IFMAP product and a shapefile of areas labeled
‘-20’ m around each polygon edge boundary was generated. The ‘-20’m buffer
shapefile was converted to a 30m cell Size grid file by an Arclnfo ‘shapegrid’
function. The buffer zone generation protocol for the three IFMAP section
coverages matched that of the 1998 MRW image bufiering protocol above and
generated an IFMAP buffer zone image that had 40-meter-wide epsilon zones
centered along the boundaries of each LULC polygon (i.e., group of pixels) (see
Figure 3.3, 2nd page).

Before proceeding to the next step, the IFMAP ‘eliminated’ polygon

coverages that had the minimum mapping unit (MMU) polygons (8,100 m2)

30

removed, were converted back to Imagine (.img) files using Arclnfo grid. These
three “MMU-eliminated” IFMAP images of the top, middle, and bottom sections
of the Muskegon River Watershed were then mosaicked back together to form a
seamless 2001 IFMAP ‘MMU-eliminated’ image.

The IFMAP ‘buffer zone’ image from above was used to mask out the
spatial uncertainty zone within the IFMAP ‘MMU-eliminated’ image. After these
processing steps, the IFMAP “NO buffer zones” image was ready for error matrix
evaluation against the 1998 MRW “NO bufi'er zones” image (see Figure 3.3, page
2). This complete process (Figure 3.3) of generating a buffer zone around LULC
polygon edges to limit the adverse effects of spatial misregistration on accuracy
assessment, is called the spatial misregistration buffer method protocol or

SPATIALM__BUFFER protocol.

31

3.3.2. Choice of Accuracy Assessment Points

Since the 1998 MRW LULC map had a high overall classification
accuracy (95%), the 40-meter-wide epsilon zone buffer-removed version of this
dataset was utilized as the base Classification for comparison to the 2001
IFMAP LULC map product. Congalton and Green (1999) have suggested that in
a heterogenous landscape (like the MRW), effective classification accuracy
assessment requires generating a group of stratified random accuracy
assessment points (AAPS) within each of the land cover categories. Utilizing
ERDAS Imagine Accuracy Assessment module, the following Choices under
‘Class Value Assignment’ options were selected: Window Size = 3; Window =
Majority Rule; and Majority Threshold = 9. These choices force the Accuracy
Assessment module to generate stratified random AAPS only within the “center
of mass” of each LULC category (i.e., the point must fall within a homogeneous
3 x 3 cell or larger window). Stratified random points were generated in order to
have at least 25 AAPS for each of the 12 land cover types listed in Table 3.2.

For improved classification accuracy analysis by eliminating the
influence of spatial misregistration on the comparison, the uncertainty zones in
both maps were avoided when selecting potential AAP locations. Overall, 817
stratified random AAPS were generated utilizing the 1998 MRW “NO buffer
zones” map as the “truth” map. These AAPS were well distributed throughout
the Muskegon River Watershed region (see Figure 3.4). A classification accuracy
and kappa statistic was generated for these 817 AAPS.

However, when these 817 AAPS were overlaid on the 2001 IFMAP

Classification map, 213 of them (26.1% of the total) landed within the IFMAP

32

spatial uncertainty zones. Including these AAPS would confound the error
matrix analysis, since the influence of the spatial misregistration, in addition to
thematic misclassification, cannot be discounted in these uncertainty zones.
Therefore, these 213 AAPS were eliminated from the total AAP selection, and
thus, 604 stratified random AAPS remained, none of which were located within
either of the spatial misregistration buffer zones. Classification accuracy and
kappa statistics were generated for both the 817 AAPS comparison and the 604
AAPS comparison, in order to demonstrate the efficacy of the spatial
misregistration buffer method protocol and its effect on thematic accuracy

assessment (Figure 3.5).

3.3.3. LULC Map Comparisons and Error Matrix Evaluation

Producer’s and user’s accuracies were analyzed to Characterize the
misclassifications between the 1998 MRW “truth” map and the 2001 IFMAP
LULC map. The error matrix comparison and analysis helped to “provide an
effective way to accurately represent map accuracy in that the individual
accuracies of each category are plainly described along with both errors of
inclusion (commission errors) and errors of exclusion (omission errors) present
in the classification” (Congalton and Green, 1999). There was also a goal to
better understand the specific Circumstances where specific AAP locations were

misclassified.

33

 

 

Figure 3.4: Distribution of 817 Accuracy

/V County Lines Assessment Points Across the MRW
. 817 StratifRandom AAPS
LULC Classif. w/‘20m Buffer

  

- _
- Agriculture
G

 

  
   
 
   
  

rass
- Shrubs

Decid. Forest
I“. Conif. Forest
- Mixed Forest

Water

Forest Wetlands
m Nonfor. Wetlands
- Barren

AAP inside 20m Buffer
(middle of Black Circle)

 

 

 

 

 

 

 

 

 

 

12/4/2006 20 O 20 I 40 IGlometers
lfin Ma ,, * :

 

34

 

 

 

Figure 3.5: Distribution of 604 Accuracy
Assessment Points After Spatial Buffering

/\/ County Lines

. 604 Strat. Random AAPS
LULC Classif. w/20m Buffer

- Forest Wetlands
r! Nonfor. Wetlands
I: Barren

r—UU

 

 

 

 

Kin Ma
12/20/06

 

 

 

 

 

20 0 20 40 Kilometers
SE

 

35

 

3.3 .4. Classification Uncertainty

Land Cover Dependency Analysis

Spatial accuracy was not the only concern for Classification accuracy
analysis. Misclassified AAPS within each of the two error matrices between the
1998 MRW “truth” map and the 2001 IFMAP LULC map were analyzed and
interpreted via spatial, temporal, spectral, and phenological considerations of
the difierent LULC classes. Error matrix cells that had values of 5 AAPS or
greater were selected for further review and analysis. This number of AAPS was
chosen because it represents about 5.0% of the total number of misclassified
AAPS from the 2001 IFMAP product. Error matrix cells with values less than 5
may have been anomalies or outliers and were de-emphasized in the individual

error matrix cell analyses.

36

CHAPTER IV
RESULTS AND ANALYSIS

4. 1. LULC Classification Comparison

4.1.1. 2001 IFMAP LULC Map

The LULC area comparison between the 1998 MRW “truth” map and the
2001 IFMAP product for each of the twelve land cover types is presented in
Table 4.1 and graphically in Figure 4.1. The IFMAP classification was derived
from a combination of Landsat TM and ETM+ scenes and utilized a clustering
analysis that incorporated both an unsupervised ISODATA algorithm for the
summer time images and also a supervised algorithm approach on some
Classes (Table 3.1). The difference between the total area Classified as urban
land cover by the 2001 IFMAP (222.4 km?) compared to the 1998 MRW (522.0
km?) is notable (Table 4.1). Since the IFMAP classification was based on spectral
reflectance patterns, pixels associated with urban land uses (rather than cover)
would tend to be underestimated. In areas of low—density residential housing,
for instance, large backyards, front lawns and wooded land between
neighboring houses would spectrally reflect a grass or forest land cover signal
and these pixels would likely be classified as one of the vegetative classes and
not the urban class.

There were also some large areal differences between the 2001 IFMAP
and 1998 MRW LULC maps, in which the 2001 IFMAP overestimated the areal
extent of the agriculture, grassland, and non-forested wetland LULC categories.
IFMAP classified 1,486.6 km2 of the MRW in agricultural land, while the 1998
MRW classified 1,253.9 km2 in the agricultural land Class. IFMAP Classified

nearly 700 km2 in grassland while the 1998 MRW classified 542.2 km2 in

37

grassland. For the non-forested wetland class, the IFMAP product Classified
166.8 km2 in this category, compared to only 90.4 km2 in the 1998 MRW LULC

map (Table 4. 1).

Table 4.1: LULC Classification Comparison

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

LULC Class 1998 MRW 2001 IFMAP

Urban 522.0 224.4
_A_griculture 1253.9 1486.6
Grass 542.2 699.9
Shrubs 142.5 131.1
Decid. Forest 2403.8 2196.9
Conif. Forest 635.6 454.7
Mixed Forest 1.7 301.3
Clearcut Forest 76.0 0.0
Water 288.1 307.1
Forested Wetlands 1099.9 1069.9
Nonforested Wetlands 90.4 166.8
Barren 1.0 11.4
Total Area (km2) 7057.0 7057.0

 

 

The 2001 IFMAP also underestimated the areal extent of the deciduous and
coniferous forest LULC categories (Table 4.1). IFMAP classified 2,196.9 km2 of
the study area as deciduous forest, while the 1998 MRW had 2,403.8 km2 in
this class. IFMAP had 454.7 km2 in the coniferous forest class, while the 1998
MRW had over 40% more area in this category (635.6 km2). Finally, the
remaining LULC categories of Shrubland, water, and forested wetlands LULC

categories had relatively equal areal estimates (Table 4.1).

38

 

 

 

 

 

 

 

 

 

 

 

 

 

2.2“: some. .
l; . >>m:2 mam—.I all lillli II... I l l I oom N

- oood

 

coming—=00 no.2
8:85.338 33 ”E. 2:2“.

I, uni) 2er

39

4.1.2. 1998 MRW LULC map

The 1998 MRW LULC map classified 522.0 km2 as urban land -- more
than twice the urban area as classified by the 2001 IFMAP product (222.4 km?)
The 1998 MRW classification relied on the definitions of the Michigan LULC
Classification system (RS 8f. GIS, 2002) and visual interpretation of color
infrared aerial imagery. When closed canopy forest and grassland occurred in
the urban built areas, the Michigan system assigns those regions to the “urban”

class. Therefore, within the urban land use setting, the 1998 MRW map

 

prioritizes the classification of land m over land cover, because residential
grassy fields, and backyard forests, would be Classified as ‘urban’ land use.
Because of the definition of this classification system, it tended to classify more
areas in the urban category. On the other hand, the 2001 IFMAP classification
system classified by multi-spectral reflectance patterns. The land cover of a low-
density residential area, for example, is dominated by the vegetative reflectance
from the trees and grass, not the small reflective signal from the roofs of the
houses or the driveways.

One example of this land cover vs. land use problem can be observed
near the town of Croton along the Muskegon River (Figure 4.2). The 1998 MRW
LULC map near AAP #724 classified the shoreline area as ‘urban’ while the

IFMAP LULC map classified the entire area as deciduous forest.

40

 

 

Figure 4.2: IFMAP v. 1998 MRW Urban

was... Citifilfififélm Companson

n as c

Grass classified as Agric (IFMAP) g’ahz's‘dfgtgflggg’zom

Agric. classified as Grass (IFMAP) ’ '

Shnrb classified as Grass (IFMAP)

ConifFor classified as Grass (IFMAP)

Urban classified as DecIdFor (IFMAP)

Grass classified as DecidFor (IFMAP)

Shrub classified as DecidFor (IFMAP)

DecidFor classified as ForWet (IFMAP)

ConifFor classified as ForWet (IFMAP)

NonForwet classified as ForWet (IFMAP)

ForWet classified as NonForWet (IFMAP)

County Lines

LU C w/20m Edge Buffer

(T: 20 in Buffer

a Urban

:7; Agriculture

’ " Grass

3 Shnrbs

- Decid. Forest
Conif. Forest

E‘ 1 Mixed Forest

    

OIOOIO‘I'OOODb

b) 2001 IFMAP with 20m
White Buffer

c) 1998 MRW with 20m
Water White Buffer

- Forest Wetlands
' Nonfor. Wetlands

E] Barren

Kin Ma
12/14/2006

 

 

 

41

 

4.2. 2001 IFMAP/1998 MRW Error Matrix Analysis and Buffer Zone
Method Development

4.2.1. Error Matrix Analysis

Classification accuracy with the entire set of 817 stratified random AAPS
was performed when the buffer zone of spatial misregistration between the two
LULC maps was not taken into account. The statistics in Table 4.2 Shows the
comparison of classification accuracy between the 1998 MRW and the 2001
IFMAP by utilizing 817 stratified random AAPS. The overall classification
accuracy is 68.79% (kappa = 0.627). When the 817 AAPS were overlaid onto the
2001 IFMAP LULC map, 213 of these AAPS landed within the 20m spatial
misregistration buffer'zone (Figure 3.4). After eliminating these 213 AAPS from
the 20m buffer zone, 604 stratified random AAPS remained (Figure 3.5). When
accuracy assessment was run on this subset of 604 points, the Classification
accuracy percentage increased by 9.85% to 78.64% (Table 4.3). These 604
stratified random AAPS did not reside in the 20m uncertainty zone due to
spatial misregistration between the two LULC maps. Hence, the thematic
accuracy of the interior core areas of the LULC polygons on the IFMAP product
is nearly 79% and the impact of spatial misregistration between the IFMAP and
1998 MRW maps could add nearly 10% apparent misclassification.

Without accounting for spatial misregistration, the 817 AAPS classification
evaluation had only two LULC categories, water and barren classes, with
conditional kappa coefficients equal to or greater than the industry standard of
0.85. Within this accuracy assessment (which includes the effects of
misregistration), the classification of the grass, shrubs, and non-forested

wetland classes performed very poorly with conditional kappa coefficients below

42

0.40 (Table 4.2). The grass and Shrubland categories generated conditional
kappa coefficients less than 0.25. The mixed forest category only had seven
AAPS, so its near-zero conditional kappa score may not be meaningful.

When the effects of spatial rrrisregistration were minimized, such as in the
604 AAPS accuracy assessment, two additional LULC classes, Urban and
Coniferous Forest generated conditional kappa coefficients that exceeded 0.85
(Table 4.3). Nearly every Conditional Kappa for each LULC type improved in the
604 AAPS comparison, except the “Mixed Forest” class. The polygon sizes of
Mixed Forest stands were very small, so that 4 out of 7 AAPS from the 817 AAPS
Classification accuracy assessment were eliminated Since they had landed in
the spatial misregistration buffer zone. As explained above, this small sample
size renders the near-zero kappa score for Mixed Forest meaningless.

For the 604 AAPS Classification accuracy comparison, the role of spatial
misregistration between LULC maps was eliminated by utilizing the
SPATIALM_BUFFER protocol. The overall classification accuracy increased to
78.64% (kappa = 0.742), which more clearly reflects the core thematic accuracy
of the IFMAP product. Nonetheless, this accuracy (at Level UN) is still 9.1%
lower than Michigan DNR’S reported 87.7% classification accuracy for IFMAP at
the Level II classification (Michigan DNR, 2004). Despite the elimination of the
role that spatial misregistration may have on accuracy assessment, the IFMAP
overall classification accuracy remains below the 85% industry standard.

Within the 817 AAPS classification accuracy comparison, there were 255
AAPS within the 2001 IFMAP LULC map that were misclassified (31.2% of the
total AAPS). 213 of these 255 AAPS (83.5% of the total misclassified AAPS) were

located within the 20m spatial misregistration uncertainty zone. In a

43

heterogeneous landscape like the Muskegon River Watershed, these data
indicate that the deleterious impact of Spatial misregistration during accuracy
assessment can be substantial and must be addressed to gain a better
understanding of core thematic accuracy.

Table 4.4 displays the error matrix for the 2001 IFMAP LULC map product
using all 817 AAPS (i.e., misregistration noise is included). A total of 255 of the
817 AAPS were rnisclassified. The largest category of misclassified AAPS were
the 20 Non—forested wetland AAPS from the 1998 MRW “truth” map, that were
misclassified as forested wetlands in the 2001 IFMAP. The second largest
number of errors of omission were 19 agriculture AAPS classified as grassland
by the 2001 IFMAP. The largest urban class errors of omission (l4 AAPS) were
misclassified as agriculture. More detailed discussions for each of the highly
misclassified matrix cells will occur in the following pages of the error matrix
analysis.

With misregistration noise removed from consideration, the 604 AAPS error
matrix (Table 4.5) displays some similarities in the categories of misclassified
AAPS from the accuracy assessment of the original 817 AAPS. The largest
number of misclassified AAPS are the 12 grassland errors of omission that were
misclassified as agriculture by the 2001 IFMAP. The second largest number of
misclassified AAPS are the 10 non-forested wetland errors of omission that
IFMAP classified as forested wetlands. The third largest number of misclassified
AAPS are the 9 agriculture errors of omission that IFMAP classified as grass.
More detailed discussions for each of the highly misclassified matrix cells will

occur in the following pages of the error matrix analysis.

44

Table 4.2: 2001 IFMAP Classification Accuracy (817 AAPS) BEFORE Buffer

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Overall Classif. Accuracy =68.79%

Overall Kappa = 0.627

Class Refer. Classified Number Producers Users

Name Totals Totals Correct Accuracy Accuracy Conditional Kappa
Urban 58 26 17 29.31% 65.38% Urban 0.6274
Agric 141 159 120 85.1 1% 75.47% _A_gric 0.7036
Grass 57 86 25 43.86% 29.07% Grass 0.2375
Shrubs 31 13 2 6.45% 15.38% Shrubs 0.1205
Decid For 239 238 197 82.43% 82.77% Decid 0.7565
Conif For 70 52 41 58.57% 78.85% Conif 0.7686
MixedFor 7 19 1 14.29% 5.26% MixedFor 0.0444
Water 53 55 51 96.23% 92.73% Water 0.9222
For Wet 1 16 125 85 73.28% 68.00% For Wet 0.6270
NFor Wet 35 33 13 37.14% 39.39% NFor Wet 0.3668
Barren Land 10 1 1 10 1 00.00% 90.91% Barren 0.9080

Totals 817 817 562

Urban = Urban/Built up, Agric = Agriculture, Grass = Grasslands, Shrubs = Shrubland
Decid For = Deciduous Forest, Conif For= Coniferous Forest, MixedFor = Mixed Forest,
For Wet = Forested Wetlands, NFor Wet = Non-forested Wetlands, Barren = Barren Land

Table 4.3:.2001 IFMAP Classification Accuracy (604 AAPS) AFTER Buffer

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Overall Classif. Accuracy =78.64% Overall Kappa = 0.742

45

Class Refer. Classified Number Producers Users

Name Totals Totals Correct Accuracy Accuracy Conditional Kappa
Urban 37 14 12 32.43% 85.71% Urban 0.8478
_A_gric 111 131 102 91.89% 77.86% _A_gric 0.7288
Grass 37 48 18 48.65% 37.50% Grass 0.3342
Shrubs 17 5 2 1 1.76% 40.00% Shrubs 0.3826
Decid For 175 184 162 92.57% 88.04% Decid 0.8317
Conif For 51 35 33 64.71% 94.29% Conif 0.9376
MixedFor 3 5 0 0.00% 0.00% MixedFor -0.005
Water 53 54 51 96.23% 94.44% Water 0.9391
For Wet 88 95 74 84.09% 77.89% For Wet 0.7412
NFor Wet 22 22 1 1 50.00% 50.00% NFor Wet 0.4811
Barren Land 10 1 1 10 100.00% 90.91% Barren 0.9076

Totals 604 604 475

 

 

pawn cot—mm u 5:3 .355; 3898.52 u .o>>n.z access; .0235“. u .035“.
.205“. 855. u Loaves. cacao“. anchor—Eco n :50 .220“. 303.com. n 2000
953sz n 32.5 8966me u 890 83325.4 u 9.9.. d: 5.6.595 u :35

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

h 5 0.. mm m 5. mm s 2. mnN en N... 5.. mm .50...
E. S. o o o o o o o o o _. :0..me
an o n .. 2 N o .. v .. o o o «0.5".2
mu.. 9 ON mm o o m o_. m N o o «0210".
mm o o o 5 o o o o o o v .325
me o o N o v m n .. o o m cents—2
um o o o o N S m o F o N 2:00
RN 0 N m o N w haw N .. w o m Bean.
2. o o o o F o v N m o .. chasm
mm o o N o o .2. o.. m mN m_. m «.890
mm: o o N o F N N m m.. oNr 3 o_._ml<r
mN o o N o o N N o .. N h .. cash:
.50... serum EZEZ 3.5.5". 33>) ou.x_s. 2:00 .3000 2.3.—cm mug—0 o_._d< can—.5 n.<.z...._ SON

 

 

8: 3.. 3:883 .25. 333.8: 82
3...? :3 5822 3am 2: 0.5.. 3m: 82 .> 2.2.... Son "3. can»

46

can. amtmm u :mtmm .3523; 00328.52 u .m>>...z .mucm_.m>> 00520“. n 53.6“.
.620“. 005.2 u .325. .690“. 990.750 u n.Eoo .690”. 2.3280 u 280
ucmfiacw n 32.5 .mccmfimfio u $90 6.3395... n 0:9. .3 5.5235 n :35

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

#00 0.. NN mm mm n 2... mt. hr hm .. —. .. um .30...
S. 0.. o o o o o o o o o .. :0...5m
NN o E. N N o .. o .. o o o u0>>n.z
mm o 0.. tn 0 o m m .. o o 0 “0.5.0”.
vm o o o 5 o o o o o o m .302,
m o o .. o o o F o o o m .0....x.E
mm o o o o 0 an N o o o o 2:00
em .. o .. m o N 0 New 0 m o m 0.02..
m o o o o o o N N _. o o 00:25
3 o o o o o o.. m m 2. m m 009.0
..m.. o o N o .. N o N N .. No .. 0.. 0%
3. o o .. o o o o o .. o N.. :mn...
.30... catam .052 «03.0". .325 Louis. 2:00 0.000 mmbEm 0020 0.594 :09: n.<Eu._ rcoN

 

 

ans. 3 3:22.25 35E 59.32 82
an: 33 x3“: Btu an: 0.5.. 3:: 82 . “.52... 38 ”3. 29¢

47

When the SPATIALM_BUFFER protocol was followed, AAPS were excluded
from the 20m spatial misregistration uncertainty zone and only 102 AAPS were
misclassified or 16.9% of the 604 AAPS (Tables 4.3 and 4.5). This group of 102
misclassified AAPS demonstrated that even after factoring out the influence that
spatial misregistration has on accuracy assessment, other factors such as
spectral differences, temporal and phenological issues, and classification
system schemes, may also contribute to classification inaccuracy between these
two LULC maps (Figure 4.3). These other factors will be explored in the pages

below.

4.2.1.1. Misclassifications of Urban class AAPS

4.2.1.1a. Urban misclassified as Agriculture
There were ten AAPS within this error matrix category, and they
occurred mostly in the central and southern regions of the MRW (Figure 4.3). In
one example, north of Big Rapids along 19 Mile Road, the AAP #490 misclassi-
fication has been influenced by the likely spectral confusion between the
reflectance of urban concrete and weathered asphalt and dry mineral soil within
agricultural fields. During the late summer or early autumn period, agricultural
fields that have been harvested may display considerable areas of bare soil.
Bare mineral soil reflectance signatures often mimic the reflectance of urban
asphalt/ concrete features. As such, agricultural bare soil regions are often
spectrally confused with true urban land cover regions. As shown in Figure 4.4,
the 2001 IFMAP misclassified large tracts of the neighboring urban areas,

particularly to the south of AAP #490 as agriculture. IFMAP’s unsupervised

48

classification algorithm performed on October, 2000 images clearly had
difficulty distinguishing between true urban land use pixels and dry harvested

agricultural plots.

49

 

 

Figure 4.3: Distribution of 102 Misclassified IFMAP AAPs

Misclassified AAPS

000?}

.fl.$El©

Urban classified as Agric (IFMAP)

Grass classified as Agric (IFMAP)

Agric. classified as Grass (IFMAP)

Shrub classified as Grass (IFMAP)

ConifFor classified as Grass (IFMAP)

Urban classified as DecidFor (IFMAP) Background: Landsat 7 ETM+, 10/19/2000
Grass classified as DecidFor (IFMAP) Path 22, R29 and 30 mosaic (Bands 4.3.2)
Shrub classified as DecidFor (IFMAP) but P21, Row 29 Partial Image Not included
DecidFor classified as FofWet (IFMAP)

ConifFor classified as ForWet (IFMAP)

NonForwet classified as ForWet (IFMAP)

ForWet classified as NonForWet (IFMAP)

County Lines

 

40 lGlometers

 

50

 

 

 

Figure 4.4: Misclassification of Urban (98MRW)

Mimed Mp5 as Agnculture Class (IFMAP)

A Urban classified as Agric (IFMAP) a) Landsat 7. 10/19/2000

A Grass classified as Agric (IFMAP)

0 Agric. classified as Grass (IFMAP)

0 Shrub classified as Grass (lFMAP)

o ConifFor classified as Grass (IFMAP)
Urban classified as DecidFor l IFMAP)
Grass classified as DecidFor llFMAP)
Shrub classified as DecidFor (IFMAP)
DecidFor classified as ForWei (lFMAP)
ConifFor classified as ForWet (IFMAP)
NonForwet classified as ForWet (IFMAP)
ForWet classified as NonForWet (IFMAP)
County Lines

U C w/20m Edge Buffer
L: 20 m Buffer
- Urban
: Agriculture
i‘“ "'3 Grass

 

 

" Conif Forest b) 2001 IFMAPwith 20m

White Buffer
- Water .. _
- Forest Wetlands "(5:2 .-
" " Nonfor. Wetlands

'3 Ban'en

c)1998 MRWwith20m

    

 

 

51

 

4.2.1.1b. Urban misclassified as Deciduous Forest

Five of these misclassified AAPS occur near urban communities in the
southem half of the MRW (Figure 4.3). For the first example, AAP #439, is
located in central Mecosta County, west of the city of Mecosta. The point is very
close to the edge of a golf course (Figure 4.5), which IFMAP correctly classified.
Golf courses are members of the ‘Outdoor Recreation” sub-category of the Level
I (Urban/ Built up) category. As shown on the 10 / 19 / 2000 Landsat image
(Figure 4.5a), AAP #439 is surrounded by woody vegetation which is why IFMAP
classified that area as deciduous forest (Figure 4.5c). This misclassification is
undoubtedly the result of the difference between land use (1998 MRW) and land
cover (IFMAP) and the inability of 30m Landsat imagery to resolve the street
patterns of the woods (Figure 4.5d). The thin, white, horizontal lines crossing
the datasets (Figures 4.5b, c) are artifacts of having divided the datasets into

three sections to produce the 20m spatial buffer files.

52

 

 

Figure 4. 5: Misclassif cation of Urban (98MRW)
as Deciduous Forest Class (IFMAP)

Misclassified AAPs
A Urban classified as Agric (lFMAP) a) Landsat 7. 10/19/2000
Grass classified as Agric (IFMAP) Bands 4,3,2 image

   
  

      
 
   

e Agric. classified as Grass (IFMAP)

e Shrub classified as Grass (IFMAP)

e ConifFor classified as Grass (IFMAP)
Urban classified as DecidFor (IFMAP)

0 Grass classified as DecidFor (IFMAP)

I Shrub classified as DecidFor (IFMAP)

e DecidFor classified as ForWet (IFMAP)

e ConifFor classified as ForWet (IFMAP)

I NonForwet classified as ForWet (IFMAP)
ForWet dassified as NonForWet IFMAP

fl Shrubs ’

- Decid. Forest b) 2001 IFMAP with 20m
t c°""' Fm“ White Buffer White Buffer

['1 Mixed Forest ~

- Water

7 Nonfor. Wetlandsg," . ’42.»
m Barren 9M!"

Kin Ma
12l14l2006

 

 

53

 

4.2.1.2. Agriculture misclassified as Grass

Of the nine AAPS within this category, most were attributable to the land
cover vs. land use problem. Grasslands can be spectrally similar to late-season
agricultural fields because the brown biomass reflectance of the grassland is
very similar to mature agricultural row crops. The 30m spatial resolution of the
Landsat imagery use for IFMAP could not resolve the row structure to
differentiate grassland (continuous brown biomass) from mature agricultural

fields.

4.2.1.3. Misclassifications of Grass Class AAPS

4.2.1.3a. Grass misclassified as Agriculture

Twelve grass AAPS were misclassified as agriculture by the 2001 IFMAP
(Table 4.3). The spectral similarity of grasslands and agricultural fields
discussed above has also adversely affected the proper classification of the AAPS
within this error matrix cell.

AAP #762 within Figure 4.6 is an example where the IFMAP classified most
of the area west of Fremont Lake as agriculture. The 1998 MRW “truth” map
was derived from 1m resolution color infrared (CIR) aerial photography which A
allowed small swaths of grass to be mapped within the sea of agriculture
around it (Figure 4.6d). Such areas are not well resolved by the 30m Landsat

imagery used for IFMAP.

4.2.1.3b. Grass misclassified as Deciduous Forest
Five Grass AAPS within this category were misclassified as Deciduous Forest

within the IFMAP LULC map. AAP #549 presents an interesting circumstance

54

because it is situated close to a grass /deciduous forest boundary. According to
the IFMAP classification, AAP #549 sits squarely within a deciduous forest plot
(Figure 4.7b), though when the position of the AAP #549 is plotted on the 1998
MRW map, it is located about 5m from the boundary between grass and
deciduous forest.

Normally, after running the spatial misregistration buffer protocol, the
boundary of each LULC polygon would be buffered 20 meters inside the polygon
edge. However, in this case the comer of the deciduous forest polygon was an
irregular polygon that had folded in on itself about a distance of 60 m on the
right side and 120 m on the left side (see Figure 4.7(d) below). Since the polygon
had inside edges, the transformation from a vector shapefile to this 20m inside
buffer grid, on the “pink” grass side, produced some additional error. The buffer
was only 20m wide, so that when the 20m buffer polygon shapefile was
transformed into a 30m grid, the inside corner where AAP #549 was located, did
not produce a 30 m pixel of deciduous forest but ONE isolated 30m pink pixel
of ‘grass’ (Figure 4.7d). Because of the technical transformation from vector
shapefiles to grids and then to Imagine images, AAP #549 fell on an isolated
‘grass’ pixel and appeared to be misclassified. This is one instance that the
spatial buffering technique has not minimized the spatial misregistration of
LULC map comparison. At this time, it is unclear why the spatial constraints
selected within the ERDAS Accuracy Assessment module (section 3.3.2.) did not
reject this one—pixel location. Further testing of this module is definitely

warranted.

55

 

 

Figure 4.6: Misclassification of Grass (98MRW)

 
   
   

   
 
  
 

Misdassified AAPS as Agriculture Class (IFMAP)
A Urban classified as Agric (IFMAP) a) Landsat 7,10/19/2000
A Grass classified as Agric (IFMAP) Bands 4, 3, 2 image
0 Agric. classified as Grass (IFMAP) ;—
e Shrub classified as Grass (lFMAP)
e ConifFor classified as Grass (IFMAP)

Urban classified as DecidFor (lFMAP)

Grass classified as DecidFor (IFMAP)

Shrub classified as DecidFor (IFMAP)

DecidFor classified as ForWet (IFMAP)

ConifFor classified as ForWet (IFMAP)

NonFonIvet classified as ForWet (IFMAP)

ForWet classified as NonForWet IFMAP)

County Lines .-_ ’ 4

U C w/20m Edge Buffer

: m Buffer

- Urban d) DOQQ

= Agriculture image

*' Grass . 7 ‘. ..

- Shmbs “ -

n Decid. Forest b) zoowmfim‘ 2°“ c) 1998 MRW with 20m

Conif. Forest . White Buff“
[:3 Mixed Forest '
- Water
Fm Wetlands

“ Nonfor. Wetlands

HBarren

N

A

Kin Ma , .
12I14I2006 3000

 

56

 

 

 

Figure 4.7: Misclassification of Grass (98MRW)
as Decid Forest and Grass as Conif Forest (IFMAP)

Mlsciassified AAPs

Urban classified as Agric (IFMAP)

Grass classified as Agric (IFMAP)

Agric. classified as Grass (lFMAP)
Shrub classified as Grass (IFMAP)
ConifFor classified as Grass (lFMAP)
Urban classified as DecidFor (IFMAP)
Grass classified as DecidFor (IFMAP)
Shrub classified as DecidFor (IFMAP)
DecidFor classified as ForWet (IFMAP)
ConifFor classified as ForWet (IFMAP)
NonForwet classified as ForWet (IFMAP)
ForWet classified as NonForWat(iFMAP) »

County Lines
u Cw/20m Edge Buffer d) AAP #5491998 MRW
Close-up View

at}:

0.0... 000')

      
  

b) 2001 iFMAP with 20m
White Buffer

a“ ' ' align
tail": "‘

‘ Nonfor. Wetlands
E| Barren

   

Kin Ma .-
12/14/2006 1
I «

we're: a...
fi

. -

 

 

a) Landsat 7, 10/19/2000
Bands 4,3,2 Image

  

c) 1998 MRW with 20m
White Buffer

J

file

il. -. . .i
’1)”J->: R5111
eters

 

 

57

 

4.2.1.4. Shrubland misclassified as Deciduous Forest

Six AAPS in this category were classified as deciduous forest in the 2001
IFMAP image. Four of these AAPs were positioned within the southern region of
Newaygo County. Within the 1998 MRW map, there is large area in south
central Newaygo County that spans an area of 4.2km x 4.3km in size that was
labeled ‘Shrub’ (see Large Black circle in Figure 3.1). IFMAP classified most of
this area, deciduous forest, and within the 10 / 19 / 00 Landsat ETM+
10/ 19 / 2000 image, the reflectance values are very similar to the deciduous
forest classified area to the northwest of the “shrub” area.

With the reflectance evidence from the Landsat imagery, this is likely one of
the few locations in the 1998 MRW LULC map that is incorrect. In order to
verify whether this is a true error, the original field check visits must be

reviewed and perhaps new field visits may need to be conducted.

4.2.1.5. Deciduous Forest misclassified as Forested Wetlands

Five AAPS in this category were misclassified, two of these AAPs occurred in
central Newaygo, and three of these AAPs were in the northern part of the MRW
and found in Eastern Missaukee and western Roscommon County. For the two
examples from eastern Missaukee County, AAPs #96 and #242, they were both
classified as deciduous forest within the 1998 MRW map, though these AAPS
were located close to the edge borders of an adjacent forested wetlands polygon,
about 62m and 80m, respectively. The proportion of the three most prominent
land cover types in this specific region was: 1) deciduous forest, 30%; 2)

forested wetlands, 50%; and 3) coniferous forest, 20%.

58

When the AAPS were overlaid onto the 2001 IFMAP classification, it was
clear that the IFMAP classification had overwhelmingly classified about 70% of
this region’s land as the Forested Wetlands category, and 20% in the deciduous
forest and then 10% of this land as the non-forested wetland category. The
Forested Wetlands category surrounded the areas around AAPs #96 and #242.

Nevertheless, the differences between the 1998 MRW and 2001 IFMAP, may
be in the definition of forested wetlands. The Wetland category is a complex
land cover type to define. Wetlands are “those areas there the water table is at,
near, or above the land surface for the significant part of many year” (RS 85 GIS,
2002). Also, for the 1998 MRW classification, “two separate boundaries are
important with respect to wetland discrimination: the upper wet boundary
above which practically any category of land cover may exist” (RS 8; GIS, 2002).

The IFMAP LULC classification system defined “lowland forest classes as
areas with evidence of flooding in the past five years and supporting lowland
indicator species” (Michigan DNR, 2004). The IFMAP codes had divided their
lowland forest classes into Level III types: 1) Lowland Deciduous, 2) Lowland
Coniferous, and 3) Lowland Mixed Forest categories. However, at level II
classification, these classes were re-assigned to Level II, “Forested Wetland”
category (see Table 3.3). IFMAP LULC’s category re-coding has likely caused

these AAPS to be misclassified in eastern Missaukee County.

59

 

 

Figure 4.8: Misclassification of Conif. Forest

(98MRW) as Forested Wetlands (IFMAP)
MisclassifiedAAPs

    
 

 

A Urban classified as Agric (IFMAP) a) Landsat 7» 10/19’2000
A Grass classified as Agric (lFMAP) Bands 4:32 'mage

e Agric. classified as Grass (IFMAP) .

e Shrub classified as Grass (IFMAP)

e ConifFor classified as Grass (lFMAP)

Urban classified as DecidFor (IFMAP)

0 Grass classified as DecidFor (IFMAP)
I Shrub classified as DecidFor (lFMAP)
e DecidFor classified as ForWet (IFMAP)
0 ConifFor classified as ForWet (lFMAP)
- NonForwet classified as ForWet (IFMAP)
e ForWet classified as NonForWet (IFMAP)
County Lines
LU C wiZOm Edge Buffer
f 20 m Buffer
Urban
,2: Agriculture
s .. G
- Shrubs
III Decid. Forest
[1,! m gum” b) 2001 IFMAP with 20m c) 1998 MRW with 20m

hite Buffer

0
I

   

 

 

60

 

4.2.1.6. Misclassifications of Coniferous Forest Class AAPs

4.2. 1.6a. Coniferous Forest misclassified as Grass

Ten AAPS were generated within this category and seven of these AAPS were
located within forested regions in the northern half of the MRW. When viewing
two examples near Blue Lake in central Mecosta County, AAP #458, and AAP
#725 were located at the edge of a coniferous forest patch on the Landsat ETM+
10/ 19/ 2000 Bands 4,3,2 composite image (see Figure 4.7a). The 1998 MRW
classification indicated that the coniferous forest patch was wider than the
2001 IFMAP and had encompassed the region east of these two AAPS (Figure
4.7c), though within the 2001 IFMAP classification map, the spatial extent of
the coniferous forest patch had decreased in size (i.e., forest cutting) and shifted
west of AAPs #458 and #725 (Figure 4.7b). With this visual evidence and the
different image dates of the 1998 MRW and 2001 IFMAP maps were derived
from, 1998/1999 v. 2000/2001, respectively, true land cover change in this

area explains these two misclassified AAPS.

4.2.1.6b. Coniferous Forest misclassified as Forested Wetlands

Five AAPs in this category were located in the northern section of the MRW
just west of Houghton Lake (Figure 4.8). The misclassified AAPS #647 and #331
can be influenced by the definitional difl'erences between the 2001 IFMAP and
the 1998 MRW forested wetland classification schemes. As described above, the
forested wetland category is a complex land cover type to define.

When AAP #331 was overlaid onto the 10 / 19 / 2000 Landsat ETM+ image, it
was located within a coniferous forest patch as shown by the magenta colored

reflectance from the Bands 4,3,2 composite image (Figure 4.8). The 1998 MRW

61

classification followed the similar classification scheme and that small patch
was indeed labeled coniferous forest. However, within the 2001 IFMAP
classification, the “lowland coniferous forest” (Level III) category was re-coded to

the “Forested Wetlands” Level II generalization (Table 3.3).

4.2.1.7. Forested Wetland misclassified as Nonforested Wetlands

Seven of these AAPS were misclassified in the IFMAP as Non-forested
wetlands. Four of these AAPS were located in the Muskegon State Game Area
northeast of Muskegon. Within the IFMAP LULC classification index of values,
they define Non-forested wetlands as wetlands with “proportion of trees is less
than or equal to 25% of land area,” (MDNR, 2004), and therefore include a
Lowland shrub category (IFMAP, Code #622, see Table 3.3). In the re-coding of
the IFMAP LULC map and to maintain compatibility, the Lowland shrub
category was re-coded to the Forested Wetland category (Table 3.3).

AAP #34, the furthest AAP to the west, is close to an adjacent Non-forested
wetland category within the 2001 IFMAP, though the IFMAP LULC classified the
entire region as Non-forested wetlands (Figure 4.9b), while the 1998 MRW map
classified the same region as forested wetlands (Figure 4.9c). In the 10 / 19 / 00
image of the area (Figure 4.9a), there is evidence of low lying vegetation.

The 1998 MRW LULC map was derived from April 1998 Color infrared
photography for early Spring, but the IFMAP Landsat images were dated in mid-
autumn (10/19/00, 9 / 29 / 01). The temporal and phenologic differences of
seasonal variation are some of the potential reasons for these four AAPs to be

misclassified.

62

 

 

Figure 4.9: Misclassification of Forested Wetlands
(98 MRW) as NonForested Wetlands (IFMAP)

Misdassified AAPS a) Landsat 7, 1011912000
Urban classified as Agric (IFMAP) Bands 4,3,2 Image

Grass classified as Agric (lFMAP)

Agric. classified as Grass (IFMAP)
Shrub classified as Grass (IFMAP)
ConilFor classified as Grass (IFMAP)
Urban classified as DecidFor ( IFMAP)
Grass classified as DecidFor (IFMAP)
Shrub classified as DecidFor (IFMAP)
DecldFor classlfied as ForWe1 (IFMAP)
ConifFor classified as ForWet (IFMAP)
NonForwet classified as ForWet (IFMAP)
ForWet classified as NonForWet (IFMAP)
County Lines

    

00.))

 

. Conif. Forest b) 200‘ 'EMAP "it“ 20'“ c) 1998 MRW with 20m
[-3 Mixed Forest WW Buffer Whlte Buffer

 

 

 

 

63

 

 

 

Figure 4.10: Misclassification of NonForested Wetlands
(98 MRW) as Forested Wetlands (I F MAP)
mammal MP8 a) Landsat 7 10/19/2000
uma classified ' IFMAP '

Grassn classified 2: 23:: (IFMAP; Band” 4'32. “‘9‘?
Agric. classified as Grass (IFMAP)
Shrub classified as Grass (IFMAP)
ConifFor classified as Grass (IFMAP)
Urban classified as DecidFor (IFMAP)

0...?

 

DecidFor classified as ForWet (IFMAP)
ConifFor classified as ForWet (IFMAP)
NonForwet classified as ForWet (IFMAP)
ForWet Llclassified as NonForWet (IFMAP)

(\{cCounty wl20m Edge Buffer
(: 20 in Buffer

 

. orest
‘“ Conif, Forest b) 2001 IFMABP with 20m c) 1998 MRW with 20m
[TI Mixed Forest ,, White . _ ‘ _ White Buffer

   

- :5-
- Fm ww‘nds .§:1:%‘ \‘VM
‘ Nonfor. Wetlands , ,.
El Barren

Kin Ma
12I14l2006

 

 

64

 

4.2.1.8. Non-forested Wetlands misclassified as Forested Wetlands

Ten AAPS within this category were distributed across the MRW though
five of these AAPs occurred in the northern third of the MRW. The two
examples, AAPS #126 and #792, were located in eastern Mecosta County
northeast of Big Rapids, in close proximity to Youngs Lake near the Haymarsh
Lake State Game Area (Figure 4.10).

For AAP #126, the shape of the polygon that surrounds this point is very
similar between the 1998 MRW and 2001 IFMAP classification maps. There was
agreement regarding the spatial extent of this feature but relative heights of the
vegetation may have influenced the non-forest v. forested wetland designation.
Spring April 1998 imagery, vegetation may only begin budding though in the
summer of early autumn period (10/19/00 Landsat image), vegetation in the
same area would be full bodied and potentially influence the IFMAP interpreters

to label this region forested wetlands.

65

4.3. Issues Affecting Classification Accuracy Assessment

The effects of spatial mis-registration have been largely reduced by the
SPATIALM_BUFFER protocol, though a series of other variables have affected
the overall classification accuracy percentage of 78.64% for the 604 AAPS
comparison (Table 4.3). The summary of other variables that affected the
classification accuracy of the 2001 IFMAP versus 1998 MRW MIRIS map
comparison include differences in classification systems, scale issues,

radiometric issues, phenological and temporal issues.

4.3.1. Classification System Issues

Researchers produce land cover maps and classify LULC according to
their own purposes or to those of their funding sources. The 1998 MRW LULC
map was based on the Michigan LULC Classification system and is the
standard for Michigan LULC classification. On a given subject, the definition of
specific concept or object must be clear to the reader and the audience in order
to enhance mutual understanding and communication. If classification systems
are not in agreement on the details of specific classification categories,
confusion between classification systems may cause misclassifications of AAPS.

IFMAP had created a modified LULC classification system that were
similar to the Michigan LULC system, though they had also assigned some
categories that did not exactly match those of the 1998 MRW LULC (Table 3.3).
For example, the IFMAP’s Lowland Forest/ Forested Wetland and Non-forested
Wetland did not match up with the 1998 MRW categories. A total of 15
misclassified AAPS were due to classification definition differences. The 15 AAPS

were misclassified in the following error matrix categories: Urban classified as

66

Deciduous Forest, Deciduous Forest classified as forested wetlands, and
coniferous forest classified as forested wetlands. Thus, these errors totaled,

14.7% of the 102 misclassified AAPs for this study.

4.3.2. Scale Issues

The 1998 MRW MIRIS base map was initially a vector shape file with a l
m resolution, and had a minimum mapping unit of 2.5 acres or 1 hectare
(10,000 m2). This vector file was aggregated and transformed into an Arclnfo
grid file that had a 30 m resolution to match the resolution of the 2001 IFMAP
LULC map. Spatial aggregation of data scales will lower the variance within the
image and affect ratio variables such as elevation and biomass (Bian, 1997),
though land cover classification type is a nominal variable. Therefore, the
effects of aggregation of the image resolution from 1 m to 30 m would mostly
affect the edges of land cover types, especially within the “urban” land cover
category. At least one AAP, grassland misclassified as deciduous forest, was

affected by this scale issue.

4.3.3. Radiometric Issues

Radiometric issues adversely influenced the misclassifications of regions
where harvested agricultural fields exposed the bare mineral soil or when
forests were clearcut and exposed the bare mineral soil of the forest floor. As
stated earlier, the reflectance of bare mineral soil land cover type is very similar
to concrete / asphalt features within the urban landscape. This issue primarily
affected the grass AAPs misclassified as agriculture (IFMAP) and Agriculture

(1998 MRW) AAPS misclassified as Grass (IFMAP). A total of 22 AAPS were

67

adversely affected by these classification accuracy errors (21.6% of the 102
misclassified AAPS).

When evaluating misclassifications within these land cover categories,
extra caution must be employed to determine whether the AAP is located within
an agricultural row field crop or adjacent to an urban area. This issue affected
primarily the agriculture AAPS misclassified as Urban, the Grass AAPS
misclassified as agriculture in the IFMAP, and the grass AAPS misclassified as
Agriculture. When evaluating misclassifications within these land cover
categories, extra caution must be employed to determine whether an urban
residential settlement was really built on the former forested plot or the location
has bare mineral soil agricultural fields. Additional field research should be
conducted to clearly ascertain the current true status of the land cover

classification category.

4.3.4. Temporal and Phenological Issues

The images, from which the two LULC maps were based, play a large role
in the misclassifications within the comparison error matrices. Table 3.2
displays the image inventory, the 1998 MRW LULC map was based on April
1998/ April 1999 aerial photo imagery, while the 2001 IFMAP LULC map was
based on a series of Landsat TM and ETM+ scenes ranging from 10/ 19/ 2000
and 9 / 29 / 2001 (for fall vegetation senescence) and July 1999 and 2001 mid-
summer images. There were a selection of points and small sections along the
edge boundaries of agriculture and urban land use classes that showed true

changes between the 1998 and 2001 image dates.

68

Since there was a difference of 2-3 years between the 1998/1999 MRW
DOQQ images and the 2001 IFMAP Landsat TM and ETM+ images, the
following land cover classes have the greatest probability for true land cover
changes, e.g., agricultural fields converted to subdivision housing, forested
lands converted to agriculture, grasslands growing back in an area where
forested trees have been clearcut. There were sixteen misclassified AAPS that
were affected by differences in phenology, and they consist of about 15.7% of
102 misclassified points.

In addition, these phenological differences mostly affected deciduous
forest / forested wetland, and agriculture / grass land cover categories because of
the growth rates of these vegetative classes during different times of the year.
The different temporal images captured the extent of vegetative growth at only
one point in time, and thereby causing the misclassification of these vegetative

classes.

69

CHAPTER V
CONCLUSIONS AND
FUTURE RESEARCH

5. 1. Conclusions

Land Use /Land Cover classification accuracy assessment is a complex
process. Misclassification or classification uncertainty can be affected by a large
number of issues, such as the spatial mis-registration between the images
being compared, phenological differences between the image dates and seasons,
and potential spectral confusion of similar vegetated classes or other landscape

characteristics within the satellite images.

5.1.1. Spatial Patterns, Magnitudes, and Land-Cover Dependencies of West-
Central Lower Michigan LULC products

When initially comparing the 2001 IFMAP LULC map to the 1998 MRW
map product, 817 stratified random AAPS were distributed across the 12 LULC
types, and the result was overall classification accuracy of 68.79%
(kappa=0.627). However, when the 817 AAPS were overlaid onto the 2001
IFMAP classification, 213 AAPS resided within a 20m spatial uncertainty buffer
zone (26.1%). These 213 AAPS were removed, and the remaining 604 AAPS were
used to assess the map accuracy and results showed that overall classification
accuracy was 78.64% (kappa=0.742).

However, even after removing misregistration noise from consideration,
102 out of 604 AAPs were still misclassified (16.9%). The following land cover
classes had the largest number of misclassified AAPS: grassland (12),

urban/built up (10), coniferous forest (10), non-forested wetlands (10), and

70

agriculture (9). Therefore, thematic misclassifications were still land-cover
dependent and can be influenced by classification system, radiometric,
temporal and phenological issues. These issues accounted up to a total of 55
misclassified AAPS, or 56% of the total 102 misclassified AAPs within the 1998
MRW and 2001 IFMAP error matrix evaluation for the Muskegon River

Watershed in the west-central Lower Michigan region (Tables 4.3 and 4.5).

5.1.2. Spatial Misregistration Bufier Protocol Development

Classification accuracy assessment of different LULC maps is dependent
on the sampling and distribution of accuracy assessment points (AAPS) across
the spatial extent of the LULC map and stratified according to the requisite
number of land cover types. In order to reduce the potential LULC
misclassification of AAPs, due to them locating in spatial uncertainty zones
within both the 1998 MRW and 2001 IFMAP LULC maps, a method of
systematically buffering from the edge of LULC polygons was developed.

My study developed a spatial misregistration buffer protocol (SPATIALM_
BUFFER), to directly address this issue that has plagued many LULC
comparisons up to this point in time. With the development of this relatively
simple SPATIALM_BUFFER protocol, classification accuracy assessment issues
related to spatial misregistration have been reduced to a minimum because the
potential AAPS were excluded from being located within the zone of spatial
uncertainty.

The development of this method protocol to create buffers equal to the
combined spatial accuracy and spatial misregistration between the 1998 MRW

and 2001 IFMAP LULC maps demonstrated that for the 2001 IFMAP dataset,

71

9.85% of overall classification accuracy, 7 8.64% minus 68.79%, was influenced
by the spatial misregistration between the two LULC map products. Hopefully
with the SPATIALM_BUFFER protocol, this method will increase the
understanding of the classification uncertainty of LULC map comparisons and
serve to move one step closer to developing an “uncertainty button” for our

geographic information systems [software] (Foody and Atkinson, 2002).

5.2. Future Research

Within the 817 AAPS comparison, an alarming proportion of stratified
random selected AAPS (83.5% of 255 misclassified AAPS, see Table 4.2) fell
within the 40-meter-wide epsilon zone of spatial uncertainty. Presumably, the
heterogeneous nature of the land cover/ use in this test watershed contributed
to this high proportion of near-edge samples. Further analysis of landscapes
with different fragmentation regimes will be necessary to appropriately evaluate

these stratified random sampling relationships.

72

REFERENCES

Annis Water Resources Institute. (2004). Developing Sustainable Futures for
the Muskegon River Watershed: A Decentralized Approach. URL:
http: / / gvsu.edu /wri/ isc/ index.cfm?id=4DADFC69-O746-A593-
33F3A69457DD556C, Annis Water Resources Institute, Muskegon,
Michigan (last date Accessed: 10 October 2006).

Anderson, J. F., Hardy, E. E., Roach, J. T., and Witmer, RE. (1976). A land use
and land cover classification system for use with remote sensor data,
US. Geological Survey Professional Paper 964, US. Geological Survey,
Washington, DC, 28 p.

Aplin, P. (2004). Remote Sensing: land cover. Progress in Physical Geography,
28 (2), 283-293.

Bian, L. (1997). Multi-scale nature of spatial data in scaling up environmental
models. In Scale in Remote Sensing and GIS, Eds., D. A. Quattrochi 85 M.
F. Goodchild, Boca Raton, FL: Lewis Publishers, pp. 13-26.

Biging, G. S., Colby, D. R., and Congalton, R. G. (1998). Sampling systems for
change detection accuracy assessment. In Remote Sensing Change
Detection: Environmental monitoring methods and applications, Eds., R.
S. Lunetta and C. D. Elvidge, Chelsea, MI: Ann Arbor Press, pp. 281-
308.

Bruzzone, L. and Cossu, R. (2002). A multiple cascade—classifier system for a
robust and partially unsupervised updating of land-cover maps. IEEE
Transactions on Geoscience and Remote Sensing, 40, 1984—1996.

Bruzzone, L. and Cossu, R. (2003). An adaptive approach to reducing
registration noise effects in unsupervised change detection. IEEE
Transactions on Geoscience and Remote Sensing, 41 (11), 2455- 2465.

Cao, C. and Lam, N. S. (1997). Understanding the scale and resolution effects
in remote sensing and GIS. In Scale in Remote Sensing and GIS, Eds., D.
A. Quattrochi and M. F. Goodchild, Boca Raton, FL: Lewis Publishers,
pp. 57-72.

Castelli, V., Elvidge, C. D., Li, C.-S., and Turek, J. J. (1998). Classification-
based change detection: Theory and applications to the NALC data set.
In Remote Sensing Change Detection: Environmental monitoring methods
and applications, Eds, R. S. Lunetta and C. D. Elvidge, Chelsea, MI: Ann
Arbor Press, pp. 53-7 3.

Chuvieco, E. and Congalton, R. G. (1988). Using cluster analysis to improve the
selection of training statistics in classifying remotely sensed data.
Photogrammetric Engineering and Remote Sensing, 54 (9), 1275- 128 1.

73

Congalton, R. G., Oderwald, R. G., and Mead, R. A. (1983). Assessing Landsat
classification accuracy using discrete multivariate statistical techniques.
Photogrammetric Engineering and Remote Sensing, 49 (12), 167 1- 1678.

Congalton, R. G. (1991). A review of assessing the accuracy of classifications of
remotely sensed data. Remote Sensing of Environment, 37, 35-46.

Congalton, R. G. and Green, K. (1999). Assessing the accuracy of remotely
sensed data: Principles and practices. Boca Raton, Florida: Lewis
Publishers, 137 p.

Corradini, K. (2006). "Grid Round - Floating Point to Integer" ArcScript.
ArcScripts Details - ESRI Support.
<http://arcscripts.esri.com/details.asp?dbid=14459>, Last Accessed: 15
November 2006.

Dai, X., and Khorram, S. (1997). Quantification of the impact of misregistration
on the accuracy of remotely sensed change detection. In Proceedings of
the IEEE Geosciences and Remote Sensing Symposium (Texas: IEEE), pp.
17 63- 1765.

Foody, G. M. (1999). The continuum of classification fuzziness in thematic
mapping. Photogrammetric Engineering and Remote Sensing, 65 (4), 443-
45 1.

Foody, G. M. (2001). GIS: the accuracy of spatial data revisited. Progress in
Physical Geography, 25 (3), 389-398.

Foody, G. M. (2002a). Status of land cover classification accuracy assessment.
Remote Sensing of Environment, 80, 185-201.

Foody, G. M. (2002b). The role of soft classification techniques in the refinement
of estimates of Ground Control Point location. Photogrammetric
Engineering and Remote Sensing, 68 (9), 897-903.

Foody, G. M. and Atkinson, P. M. (2002). Current status of uncertainty issues
in remote sensing and GIS. In G. M. Foody and P. M. Atkinson, Eds.,
Uncertainty in Remote Sensing and GIS, West Sussex, England: John
Wiley 81. Sons, Ltd., 287-302.

Jensen, J. R. (2005). Introductory digital image processing: A remote sensing
perspective, 3rd Ed. Upper Saddle River, NJ: Pearson Prentice-Hall, pp.
337-388, pp. 467-487.

Lillesand, T. M. and Kiefer, R. W. (2000). Remote sensing and image

interpretation, 4th Ed., New York: John Wiley 85 Sons, pp. 373-416, 470-
615.

74

Loveland, T. R. and DeFries, R. S. (2004). Observing and monitoring land use
and land cover change. In R. S. DeFries, G. P. Asner, and R. A.
Houghton, Eds., Ecosystems and land use change, Geophysical
Monograph Series 153, Washington, DC: American Geophysical Union,
23 1 -246.

Mecosta County Extension Office. (2000). Michigan Resource Information
System (MIRIS) 1998 County wide Update of 1978 Land use land cover
map, ArcView shape files and tables digital data.

Michigan Center for Geographic Information. (2006). MDNR Spatial Data
Library - 1998 Digital Orthophoto Quadrangles, URL:
http: / /www.mcgi.state.mi.us/mgdl/doqs_zip/ DOQQ98_Map.htm, Center
for Geographic Information, State of Michigan, Lansing, MI (last date
accessed: 20 October 2006).

Michigan Department of Natural Resources. (2004). Integrated Forest
Monitoring and Assessment and Prescription Project (IFMAP): Review of
Remote Sensing Technologies used in the IFMAP Project, Final Report.
Prepared by Space Imaging for Michigan Department of Natural
Resources, State of Michigan, Lansing, MI, 47 p.

Montcalm County Extension Office. (2000). Michigan Resource Information
System (MIRIS) 1998 County wide Update of 1978 Land use land cover
map, ArcView shape files and tables digital data.

Nelson, S. A. C., Soranno, P. A., and Qi, J. (2002). Land-cover change in Upper
Barataria Basin Estuary, Louisiana, 1972-1992: Increases in wetland
area. Environmental Management, 29 (5), 716-727.

Nusser, S. M. and Klaas, E. E. (2003). Survey methods for assessing land cover
map accuracy. Environment and Ecological Statistics, 10, 333-356.

Openshaw, S. (1984). The modifiable areal unit problem. Concepts and
Techniques in Modern Geography, No. 38, Norwich: Geo Books.

Patil, G. P. and Taillie, C. (2003). Modeling and interpreting the accuracy
assessment error matrix for a doubly classified map. Environment and
Ecological Statistics, 10, 357-373.

Perkal, J. (1966). On the length of empirical curves. Discussion paper 10, Ann
Arbor Inter-University Community of Mathematical Geographers.

Remote Sensing 81. Geographic Information Science Research and Outreach
Services. (2002). Michigan Land Cover/ Use Classification System.
Introduction to Aerial Photo Interpretation. Michigan State University,
East Lansing, MI, 52 p.

75

Remote Sensing 81. Geographic Information Science Research and Outreach
Services. (2004). Muskegon River Watershed Project Metadata Readme
file. Michigan State University, East Lansing, MI, 3 p.

Salas, W. A., Boles, S. H., Frolking, S., Xiao, X., and Li, C. (2003). The
perimeter / area as an index of misregistration bias in land cover change
estimates. International Journal of Remote Sensing, 24(5), 1165- 1170.

Smits, P. C., Dellepiane, S. G., and Schowengerdt, R. A. (1999). Quality
assessment of image classification algorithms for land-cover mapping: a
review and a proposal for a cost-based approach. International Journal of
Remote Sensing, 20(8), 146 l- 1486.

Sohn, Y., Moran, E., and Gurri, F. (1999). Deforestation in North-Central
Yucatan (1985-1995): Mapping secondary succession of forest and
agricultural land use in Sotuta using the cosine of the angle concept.
Photogrammetric Engineering and Remote Sensing, 65 (8), 947-958.

Sohn, Y. and Rebello, NS. (2002). Supervised and Unsupervised Spectral Angle
Classifiers. Photogrammetric Engineering and Remote Sensing, 68 (12),
127 1- 1280.

State of Michigan. (2004). 1978 Michigan Resource Information System, Center
for Geographic Information, Geographic Data Library. Found at
http: / / www.mcgigate .mi.us / mgdl / ?rel=thext&.action=
thmname&cid=5&cat=Land+Cover°/o2FUse+MIRIS+ 1978

Stehman, S. V. (2001). Statistical rigor and practical utility in thematic map
accuracy assessment. Photogrammetric Engineering and Remote Sensing,
67 (6), 727-734.

Stehman, S. V. and Czaplewski, R. L. (2003). Introduction to special issue on
map accuracy. Environment and Ecological Statistics, 10, 301—308.

Stow, D. A. (1999). Reducing the effects of misregistration on pixel-level change
detection. International Journal of Remote Sensing, 20(12), 2477-2483.

Verbyla, D. L., and Boles, S. H. (2000). Bias in land cover change estimates due
to misregistration. International Journal of Remote Sensing, 21 (18),
3553-3560.

Wang, H. Q. and Ellis, E. C. (2005). Image misregistration error in change
measurements. Photogrammetric Engineering and Remote Sensing, 71 (9),
1037- 1044.

Wexford County Extension Office. (2000). Michigan Resource Information

System (MIRIS) 1998 County wide Update of 1978 Land use land cover
map, ArcView shape files and tables digital data.

76

   

liiiiliiiiiiiiiiiiiiiiiiil