I.“

a II.

39....
n.

E ‘1‘. Will

.. .i l...
. , ....I

x. .
7 .c

:3

L .01...

.u

 

via: .5. 1..
. 1...... ,

a. .6...

.l...$..$
3.... 1.2))
. 74.3.1...

.3 an.

.A:
4.

fig???

. , Ewiwvn.
smifl..h..x .
$.....4§..mw?

 

a. .
ivfinuxmrs

1.5.1... .

..r 3...}...

 

I)...

 

a... 1... D

1...... t.

a L. 5):, 120'}
1.9.2.4.: i... i... i...

1..

523;:

:35

. u

u.
.35..

...§.fi.w..». .13 3.13...

w,

.. mad.
I. V
. .1.
G

h

, "Mask“.

3|. h.

2 an. xs
£-

3. .

A".

z...

.8... 0|...» .1. .r!
133.4“- '7
.k

s ‘4

..... ..
lul‘

”1:34-: .1: .. p
3.24:)..." is \..r

 

.r’...L.... WHWJWQ

LIBRARY
Michigan State
University

AI/TN

é/éjgai" .’ 9’

This is to certify that the
thesis entitled

COMPARATIVE ANALYSIS OF FOREST CLASSIFICATION IN
FOREST MANAGEMENT INFORMATION DATABASES IN
MICHIGAN

presented by

Nirmal Subedi

has been accepted towards fulfillment
of the requirements for the

Master of Science degree in Forestg

 

 

pig/W1 a. flag“.

MajSiProfessor’s Signature
Auqufl" I8, 2005'
d .

 

Date

MSU is an Affinnative Action/Equal Opportunity Institution

..-.—.--.—--- no--o.—---u---o-q-I-o-u--u.-a-o--u-.-------o-a-n-n--n-c-a-1-.-.-o-n-n-n-o-I-c-o-o-o-n-n-a-n-I---

 

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

0 ES 5 Z 5 ‘E‘m

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

2/05 c:/CIRC/DateDus.lndd-p.15

—_‘_.

 

CONIPARATIVE ANALYSIS OF FOREST CLASSIFICATION IN FOREST
MANAGEMENT INFORMATION DATABASES IN MICHIGAN

By

Nirmal Subedi

A THESIS

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

MASTER OF SCIENCE

Department of Forestry

2005

ABSTRACT
COMPARATIVE ANALYSIS OF FOREST CLASSIFICATION IN FOREST
MANAGEMENT INFORMATION DATABASES IN MICHIGAN
By

Nirmal Subedi

In Michigan, there are four primary sources of forest management information for public
forest lands, namely a raster land-cover map (IFMAP), Forest Inventory and Analysis
(FIA) plot-level information, Natural Resource Information System Field Sampled
Vegetation (NRIS-FSVeg) for national forest lands, and Operations Inventory (01) for
state-owned forest lands. The objective of this study is to compare forest classifications
between and among the forest management databases with FIA data as the reference
location for comparison. Difference matrices were created between and among forest
classifications and descriptive accuracy assessments for overall accuracy, producer’s
accuracy and user’s accuracy were computed. The overall accuracy of IFMAP with FIA
as reference was 63.6% for state forest lands and 64.8% for national forest lands. Overall
accuracy of IFMAP with 01 as a reference was 60.3% and IFMAP with NRIS-FSVeg as
a reference was 68.3%. Overall accuracy of OI with FIA was 84.5% and NRIS-FSVeg
with FIA was 82.2%. Overall accuracy of three-way forest classification was 54.8% and
58.5% for state and national forests lands, respectively. Kappa statistic, calculated from
three approaches, ranged from 0.568 to 0.628 for state forest lands and 0.555 to 0.612 for

national forest lands. This finding is consistent with a previous study of IFMAP.

Copy right by
NIRMAL SUBEDI
2005

ACKNOWLEDGEMENTS

This study is a collaborative effort Of the Department Of Forestry, MSU, the Michigan
Department of Natural Resources, the national forests of Michigan, and FIA unit, NCRS.
Many people working in different organizations contributed significantly for the
successful completion of this thesis. Committee members, Dr. David MacFarlane and Dr.
Rique Campa provided valuable contributions in improving the quality of work. Dr.
Larry Pedersen, another committee member, was very instrumental for creating amicable
environment for inter-agency collaboration and providing empirical understanding of the
problem. In particular, my advisor, Dr. Larry A. Leefers, guided me in the MS program,
provided generous support for better understanding the research problem, resolving the
issues during data analysis and improving the quality of research work. Similarly, Mr.
Geoff Holden, NCRS, played a very crucial and cooperative role to accomplish this study
on time. Including my committee members, Dr. Donald Dickmann, Dr. Mark Hansen,
Mr. Jason Stephens and Mr. Joseph Gates contributed significantly to improve the
crosswalk tables for forest classifications. MSU Faculty Dr. Steven Friedman, Dr. Kelly
F. Millenbah and Dr. Murari Suvedi were generous in providing valuable inputs for my
research as well. The fellow colleagues at MSU particularly Eric White, Anita Morzillo,
Qing Xiang and Murari Regmi were very helpful to me while learning new skills required
for this research.

I would like to thank Fulbright Scholarship for providing an excellent Opportunity to
study MS degree at Michigan State University. Special thanks are due to my spouse
Manjana, my daughter, Swastika for their patience and understanding and support.

Without their cooperation, it would have been impossible to accomplish my goal.

iv

TABLES OF CONTENTS

LIST OF FIGURES ................................................................................... v

CHAPTER 1

INTRODUCTION ..................................................................................... 1
Michigan Statewide Raster Map or IFMAP ................................................ 3
Operational Inventory .......................................................................... 5
Field Sampled Vegetation of Natural Resource Information System
(NRISFSVeg) ................................................................................... 6
Forest Inventory Analysis (FIA) ............................................................ 10
Problem Statement ............................................................................ 12
Study Purpose .................................................................................. 13
Organization of Thesis ....................................................................... 14

CHAPTER 2

LITERATURE REVIEW .......................................................................... 16

Land Use

Classification ......................................................................................... 16
Michigan Resource Inventory System ...................................................... 21
National Land Cover Database (NLCD) for United States .............................. 22
IFMAP Classification ......................................................................... 23
Forest Classification of FIA/OI/NRIS-FSVeg ............................................ 25
FIA Sampling ................................................................................. 26
Digital Imagery Classification ............................................. . ................. 29
Remote Sensing Based Cover Type Mapping Studies .................................. 35
Accuracy Measurement Methods Used for Remote Sensing Imagery
Classification. . . ......................................................................................................... 41

CHAPTER 3

DATA SOURCES, STUDY AREA, CLASSIFICATION APPROACHES

AND ANALYSIS METHODS ..................................................................... 51
Data Sources ................................................................................... 51
Study Area ..................................................................................... 53
Classification Approaches ................................................................... 54
Analysis Methods ............................................................................. 55
Chapter Summary .............................................................................. 63

CHAPTER 4

RESULTS AND DISCUSSION .................................................................. 64
Using FIA Forest Types as Reference Classification .................................... 68
Accuracy Assessment for State Forest Lands ............................................. 72
Accuracy Assessment for National Forest Lands .......................................... 77
Three-way accuracy of FIA/OI/IFMAP Assessment in the State Forest Lands ..... 85

Three-way accuracy of FIA/NRIS-FSVeg/IFMAP Assessment for the National

Forest Lands ................................................................................... 88
Comparison of Classification Agreement Between IFMAP and FIA for State
Forest Lands ................................................................................... 91
Comparison of Classification Agreement Between IFMAP and FIA for National
Forest Lands ....................................................... . ........................... 98
Explanation of Differences in Classification of IFMAP, 01 with FIA ............... 104
CHAPTER 5
SUMMARY AND CONCLUSION ............................................................. 110
Conclusions ................................................................................... 1 12
Limitations ................................................................................... 1 13
Policy Implications .......................................................................... 115
Additional Research. . . ............................................................................................ l 16
REFERENCES CITED .......................................................................... 117
APPENDIX A
A1. NRIS—FSVeg FIA Crosswalk ....................................................... 125
A2. IFMAP-FIA Crosswalk ............................................................. 126
A3. OI-FIA Crosswalk ................................................................... 127
A4. IFMAP-OI Crosswalk ............................................................... 128
A5. IFMAP-NRIS-Veg Crosswalk ..................................................... 129

vi

LIST OF TABLES

Table 1. Forest Classification of Michigan Statewide Raster Map (IFMAP) ................. 5
Table 2. Forest Type Classification in Operational Inventory .................................. 7
Table 3. Forest Types Used in Michigan National Forest ....................................... 9
Table 4. FIA Forest Type Classification ......................................................... 12
Table 5. Anderson et a]. (1976) Land Use Classification ...................................... 17
Table 6. Comparison of Forest Management Information Database in Michigan ......... 65
Table 7. Difference Matrix of IFMAP and FIA for State Forest Lands ..................... 73
Table 8. Difference Matrix of OI and FIA for State Forest Lands ........................... 74
Table 9. Difference Matrix of IFMAP and 01 for State Forest Lands ....................... 75
Table 10. Difference Matrix of IFMAP and FIA for National Forest Lands ............... 78
Table 11. Difference Matrix of NRIS-FSVeg and FIA for National Forest Lands. . . .80

Table 12. Difference Matrix of IFMAP and NRIS-FSVeg for National Forest Lands. . ..82

Table 13. Three-way accuracy comparison of FIA, 01 and IFMAP for State

Forest Land .......................................................................................... 86
Table 14. Three-way accuracy comparison of FIA, NRIS-FSVeg and IFMAP for

National Forest Land ................................................................................ 89
Table 15. Error Matrix of IFMAP-FIA for State Forest Land (UA constant) ............... 93
Table 16. Error Matrix of IFMAP-FIA for State Forest Land (PA constant) ............... 94
Table 17. Error Matrix of IFMAP-FIA for State Forest Land (Aggregated) ................ 95
Table 18. Individual Error Matrix Kappa Analysis Results for State Forest Land ......... 96
Table 19. Kappa Analysis Results for Pairwise comparison of Error Matrices ............ 98
Table 20. Error Matrix of IFMAP-FIA for National Forest Land (UA constant) .......... 99

vii

Table 21. Error Matrix of IFMAP-FIA for National Forest Land (PA constant) ......... 100
Table 22. Error Matrix of IFMAP-FIA for National Forest Land (Aggregated) .......... 101
Table 23. Individual Error Matrix Kappa Analysis Results for National Forest Land... 102

Table 24. Kappa Analysis Results for Pairwise comparison of Error Matrices ........... 102

viii

LIST OF FIGURES
FIGURE 1. National Forest and Digitized State-Owned Forest in Michigan .............. 52

FIGURE 2. Illustration of an Error Matrix or a Confusion Matrix ........................... 57

ix

CHAPTER 1

INTRODUCTION

Forest resources are increasingly important in providing ecosystem stability, human
recreation and other goods and services. Information about forest resources is needed to
meet the demand for goods and services from forests, whether in Michigan or elsewhere.
“Michigan’s forests are making important contributions to the quality of life by providing
a wide array of benefits including wildlife habitat, biological diversity, outdoor
recreation, and improved air and water quality. Economic contributions are also
significant—an estimated $12 billion of value added and 200,000 jobs annually are
supported by forest based industries/ tourism/ recreation” (Schmidt et al. 1997). Michigan
forestland area totals 19.3 million acres, and timberland totals 18.7 milliOn acres (MDNR,
2002, Leatherberry et al. 2005). The state timberland acreage is the sixth largest in United
States exceeded only by the states of Oregon, Georgia, Alabama, Montana and North

Carolina (Smith et al. 2004).

Management of forest resources requires information about the resources, such as their
location, forest type and other characteristics. In addition, demographic, social and
economic information are important to assess demand. Nowadays, publics are demanding
a variety of goods and services from forestlands, and multiple use management is the
main principle for providing such demands. Economic multiple-use forestry planning is a

relatively challenging task and it requires information about resources and their demand.

In this way, multiple use forestry planning has enhanced the importance of spatial
information of forest stand characteristics, as it compares the economic benefits from
joint production and specialized production. This thesis focuseson spatial information

regarding forest type as defined by various information databases.

Forest type information is one of the most valuable pieces of information for forest
managers and planners. It can reveal the types of vegetation that are growing which
relates to information about the potential production of goods and services at a particular
location. In Michigan, four primary sources provide information for forest types available
on public lands. However, they sometimes provide conflicting information. The first is a
raster map based on remote sensing satellite imagery information, locally named
Integrated Forest Monitoring Assessment and Prescription (IFMAP); it provides .
information statewide (PMR, 2001a). The second is Operations Inventory (OI) developed
by the Forest Minerals and Fire Management Division of the Michigan Department of
Natural Resources (MDNR) which provides information for state-owned forestlands. The
third is the Field Sampled Vegetation of the Natural Resource Information System
(NRIS-FSVeg), developed by the United States Department of Agriculture-Forest Service
(USDA-PS) which has information for national forests. The other information source is
the sampled plot information inventoried by the USDA-F8 Forest Inventory and Analysis
(FIA) staff. For every FIA plot on state and national forests, there is a corresponding
IFMAP pixel (raster) and either an 01 (MDNR) or NRIS-FSVeg (USDA-FS)

classification.

Michigan Statewide Raster Map or IFMAP

IFMAP is a statewide map of land cover over all ownerships within the state of
Michigan. It also provides a means for the MDNR to develop management prescriptions
for their lands. Only the former is used in this research. Basically, this map is derived
from Landsat Thematic Mapper 5 and 7 Enhanced Thematic Mapper plus (ETM +) with
some ground verification. There are two separate maps, one for the Upper Peninsula and
the other for the Lower Peninsula. This is a coarse resolution map for producing a land
cover map and dataset, which can serve multiple functions (MDNR 2003b, MDNR
2003c). It is anticipated that this map will provide a resource for ecosystem—scale
management and a statewide planning tool for wildlife habitat (PMR, 2001a). In addition,
the pixel-level information of land cover can be integrated to assess the resources on a
unit or county-wide scale. Each pixel has spatial resolution of 30 meters X 30 meters

(PMR, 2001a).

The inventory module of IFMAP records information on present land cover, and the
activity tracking module records information on treatments and disturbances (PMR,
2001b). These two types of information are kept in separate GIS map layers. In IFMAP, a
proposed treatment polygon can be a subset for a land cover stand or even from several

adjacent stands.

The percentage of canopy occupied by a given species was used for rules to derive land

cover and association categories (PMR, 2001a). The percentage of canopy cover is

determined by a sensing mechanism, either satellite-home or airborne. Because it is
overhead, there is an extremely limited capability for canopy penetration; below-canopy
data is not used in this classification. The classification schemehas been examined by
representatives of all regions of the state, and resulting changes/additions made help
account for the variety of species, species associations, and land cover types found in

Michigan.

Land use in IFMAP is classified into eight land-use classes: Urban, Agriculture, Upland
Open land, Upland Forest, Lowland Forest, Non-forested Wetland and Bare/Sparsely
Vegetated. In the Upland classifications lands not periodically flooded or not on hydric
soils are included, and Lowlands refer to lands that are periodically flooded and/or on
hydric soils. Land is classified as forest when the canopy cover exceeds the 25% of the
ground. The algorithm for the forest type classification can be assessed via the Michigan

Geographical Data Library website (MDNR 2003a).

IFMAP has classified forestland into 12 different forest types in the Lower Peninsula and
into 11 different forest types (excludes other upland deciduous) in the Upper Peninsula
(Table 1). There are 20 other non-forestry land use classifications to cover the entire
State. Thus, IFMAP facilitates statewide identification of the land-use within 32 land-use

categories (for details see, MDNR, 2003e).

Table 1. Forest Classification of Michigan Statewide Raster Map (IFMAP) a

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Forest Land Use Class Forest Type Grid Value
Upland Deciduous Northern Hardwood Association 14
Forest Oak Association . 15
Aspen Association 16
Other Upland Deciduous 17
Mixed Upland Deciduous 18
Upland Conifer Forest Pines 19
Other Upland Conifers 20
Mixed Upland Conifers 21
Upland Mixed Forest 22
Lowland Forest Lowland Deciduous Forest 24
Lowland Coniferous Forest 25
Lowland Mixed Forest 26

 

Source: Michigan Geographic Data Library, 2003 (http://www.mcgi.state.mi.us/mgdl/)
a Non-forested land with lower canopy class (less than 25%) include Herbaceous
openland, Upland and lowland shrub / Low density trees, Mixed non-forested wetland
and Other bare/ sparsely vegetated.

Operations Inventory

Operations Inventory (01) is an inventory system developed by the Forest, Minerals and
Fire Management Division (FMFMD) of the MDNR. OI locates and identifies physical,
biological, economic, and social information on each unit of land (MDNR, n. d.). OI is
expected to provide operational level information related to resource management issues
regarding timber, wildlife, forest recreation, water quality, reforestation and land use. It
provides the descriptive information at the stand level (the smallest record keeping unit)
and a plan of operation for the stand after a multidisciplinary review of a preliminary

prescription.

Within OI, stands are grouped into compartments (group of stands with average area of

1500 to 3000 acres) based on proximity, common access, landform and soil properties

and uniformity on distribution of major cover type acreages (MDNR, n. d.). The 01
definition of forest is very similar to FIA, it identifies a stand as a forest when it has at
least 16.7% stocked and land capable of producing 20 cubic feet per year (Pedersen, pers.
comm.). Compartments are grouped by year-of-entry (YOE). The 01 analysis proceeds
by compartment within a given YOE. Each YOE contains approximately 10 percent of
the compartments in a forest area (MDNR, n. d.). By the end of Fiscal Year 2004, the
inventory of compartments with YOE 2003, 2004, 2005 and 2006 was completed, and the
boundaries of compartments and stands were digitized using a geographic information
system (GIS). The inventory of whole forest area owned by MDNR will be completed in
six more years. This research is limited to the compartments within these four YOE. The

01 has classified the forest type (referred to cover types) into 16 categories (Table 2).

Field Sampled Vegetation of Natural Resource Information System (NRIS-

FSVeg)

The USDA-FS’s Field Sampled Vegetation of Natural Resource Information System
(NRIS-FSVeg) combines a standard corporate database and computer applications
designed to support field-level users. NRIS databases contain basic natural resources data
in standard formats built to run within the Forest Service computing environment (NRIS
FSVeg, 2005), and it provides agency personnel with the information needed to respond
to public concerns and to address complex issues. Basically, it provides a diverse range of
basic and calculated information in standard formats that can be shared across

administrative and ecological boundaries.

Table 2. Forest Type Classification in Operations Inventory 8

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Cover
Type Description Short Description
A ASPEN (UPLAND) Aspen
B PAPER BIRCH Paper Birch
C CEDAR Cedar
E SWAMP HARDWOODS Swamp Hrdwds
SPRUCE-FIR (UPLANDS-INCLUDING UPLAND BLACK
F SPRUCE) Spruce Fir
H HEMLOCK Hemlock
I LOCAL USE Local Name
J JACK PINE Jack Pine
M NORTHERN HARDWOOD Upland Hdwds
O OAK Oak
BALSAM POPLAR & SWAMP ASPEN and SWAMP WHITE
P BIRCH Lowlnd Poplr
Q MIXED SWAMP CONIFER Mx Swmp Cnfr
R RED PINE Red Pine
S BLACK SPRUCE-SWAMP Black Spruce
T TAMARACK Tamarack
W WHITE PINE White Pine

 

Source: Operations Inventory Manual, MDNR (Unpublished)

“ Non-forest classes include Tree bog, Grass, Rock, Lowland brush, Marsh, Upland brush,
Bog or Muskeg, Other non-stocked or non-forest or non-productive land, Sand dunes,
and Water.

Field Sampled Vegetation (FSVeg), one of the resource modules of NRIS, provides
guidelines for National Forest management planning. This component includes point and
plot vegetation data from field surveys. Data on trees, surface cover, understory
vegetation and down woody material are included in this component. The USDA-FS
defines Forest land as “[l]and at least 10 percent stocked by forest trees of any size,
including land that formerly had such trees cover and that will be naturally or artificially
regenerated. Forest land includes transition zones, such as areas between heavily forested
and nonforested lands that area at least 10 percent stocked with forest trees and forest

areas adjacent to urban and built-up lands. The minimum area of classification for forest

land is 1 acre. Roadside, streamside and shelterbelt strips of trees must have a crown

width of at least 120 ft to qualify as forest land. Unimproved roads and trails, streams,
and clearings in forest areas are classified as forest if less than 120 ft wide” (Smith et al.

2004).

NRIS conducts ongoing strategic inventories, tactical inventories and stand examinations.
The strategic inventory scale is national, the tactical inventory scale is for the Region or
Forest, and the stand examination scale is for a project area, stand or vegetation
condition. Basically, the stand examinations are one "type" of inventory conducted by the
USDA-PS. Its purpose is to obtain the site and setting characteristics required to identify
stand conditions and capabilities. The information may be collected by simple
observations, or by formal, intensive examinations. The appropriate method to be chosen
depends on factors such as stand complexity, the decisions to be made, and the purpose
of the exam. Each examination method has varying data and accuracy requirements.
However, stand examination data provide information for a wide variety of uses ranging
from determining silvicultural treatments to evaluating wildlife habitat and modeling

water yields.

Integration of the stand level information, based on inventory design, produces national
forest planning information. Without knowing the compatibility of scope, scale, and
objective, there is a risk of introducing bias into a combined data set (NRIS FSVeg
2004). However, the design considerations and data acquisition guidelines prepared by

NRIS are conceptual for national forest management planning rather than directional.

The Ottawa, Hiawatha and Huron-Manistee national forests are located in Michigan. The
forest cover classification used in the NRIS-FSVeg database of the national forest
includes 60 forest cover types and has separate codes for each class (Table 3).

Table 3. Forest Types Used by Michigan National Forests

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Code Forest type Code Forest type
1 Jack Pine 55 Northern Red Oak
2 Red Pine 57 Scarlet Oak
3 White Pine 59 Mixed Oak
4 White Pine-Hemlock 60 Oak-Hardwoods
5 Hemlock 63 Northern pin Oak
6 Scotch Pine 70 Sugar Maple-Black cheery
7 Norway Spruce 71 BI Ash-Elm-R Maple
8 White Spruce 74 White Ash
9 Conifers 76 Red Maple(Wet)
10 Spruce 79 Mixed Lowland Hdwd
ll Balsam Fir-Asp-PB 80 Sugar-Maple-northern red oak
12 Black Spruce 81 S Maple-Beech -YB
14 Northern Wh Cedar 82 S Maple-Basswood
15 Tamarack 83 BI CH-W Ash-Y Pop
16 Wh-Sp-BF 84 Red Maple(Dry)
17 Upland BI Spruce 85 Sfiar Maple
18 Mix Swamp conifer 86 Beech -
l9 Cedar-Aspen-PB 87 Sugar maple-beech/yellow
20 Northern Hdwd-Heml birch/red spruce
21 Mixed Northern Hdw 88 Black Locust
22 N. White Cedar- UP 89 Mixed Upland Hdwd
23 W. Spruce-BF-Aspen 90 Sugar maple-beech! basswood
24 Balsam Fir 91 Quaking Aspen
41 Wh Pine NRO-W Ash 92 Paper Birch
42 E. Red cedar Hardwood 93 Bigtpoth Aspen
43 Oak-Eastern white pine 94 Balsam Poplar
47 Oak-Aspen 95 Asp-W Spruce BF
48 Jack Pine-Oak 97 Lowland Brush
49 Red Pine-Oak 98 Upland Brush
53 Black Oak 99 Open
54 White Oak

 

Source: NRIS FSVeg Data Dictionary Version 1.7 January 2005,
Appendix (pp E6 and E-8)

Forest Inventory Analysis (FIA)

The primary objective of FIA is to determine the extent, condition, volume growth, and
depletions of timber on the Nation’s forestland (Miles et al. 2001). The FIA’s continuing
endeavor is mandated by Congress in the McSweeney-McNary Forest Research Act of
1928 and the Forest and Rangeland Renewable Resources Planning Act of 1974. Further,
the 1998 Farm Bill required FIA to collect data on 20 percent of plots annually within

each State (Miles et al. 2001).

FIA defines forest areas as “[l]and at least 16.7 percent stocked by forest trees of any
size, or formerly having had such tree cover, and not currently developed for nonforest
use. The minimum area for classification of forest land is one acre. Roadside, streamside,
and shelterbelt strips of timber must have a crown width of at least 120 feet to qualify as
forest land. Unimproved roads and trails, streams or other bodies of water or clearings in
forest areas shall be classified as forest if less than 120 feet wide” (Hahn and Hansen

1985).

“The FIA inventory is based on aerial photo and/or remote sensing activity used to
characterize the acreage of forest and non-forest land in the US. These classes are based
on land-use. For forested land, more detailed classes are sometimes defined based on
criteria such as forest type, volume per acre, stand size, stand density, ownership and/or
stand age” (Miles et al. 2001). Then, ground plots are measured to adjust the remote
sensing sample for changes since its data acquisition and to correct any rrrisclassification.

"The remote sensing classification of these ground plots, together with the area estimates

10

from the remote sensing sample is used to assign area expansion factors to all ground
plots” (Miles et al. 2001). These area expansion factors are used to weigh plot-level

estimates when computing estimates for selected strata of the population.

FIA plots are designed to cover a 1-acre sample area; however, not all trees on the area
are measured. Recent inventories use a national standard, fixed-radius plot layout for
sample tree selection. Ground plots may be new plots that have never been measured
during a previous inventory. For all plots several observations are recorded for each
sample tree, including its diameter, species and other measurements that enable the
prediction of the tree’s volume, growth rate, quality, and forest health data. These tree
measurements form the basis of the data on the tree records in the FIA database.
According to the sixth FIA inventory online information, there were 45 forest types

reported as classified by field crew for the public forest land in Michigan (Table 4).

A variety of tools like maps, aerial photographs/imagery, and Global Positioning System
(GPS) units are utilized to properly install the ground plots (Burkman 2005). Once a
ground plot location has been selected on an aerial photograph, it is established and
measured in the field. On all forested field plots, quantitative and qualitative
measurements are made for conditions such as tree diameter, length, damage, amount of
rotten or missing wood and tree quality, tree regeneration, site quality information,
stocking, and general land use. And the general stand characteristics are gathered for
forest type, stand age and disturbance, change in land use, general stand characteristics

and estimates of growth, mortalities, and removals are gathered (Burkman 2005).

ll

Table 4. FIA Forest Type Classification

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Code Forest Type Code Forest Type
101 Jack pine 515 Chestnut oak / black oak / scarlet oak
102 Red pine 519 Red mafle / oak .
103 Eastern white pine 520 Mixed upland hardwoods
104 White pine / hemlock 700 Elm/ Ash / Cottonwood Group
105 Eastern hemlock 701 Black ash / American elm / red maple
121 Balsam fir 703 Cottonwood
122 White spruce 704 Willow
125 Black spruce 706 Sugarberry / hackberry / elm / green ash
126 Tamarack 708 Red maple / lowland
127 Northern white-cedar 709 Cottonwood / willow
380 Exotic Softwoods Group 800 Maple / Beech / Birch Group
381 Scotch pine 801 Sugar maple / beech / yellow birch
383 Other exotic softwoods 802 Black cherry
400 Oak Pine Group 803 Cherry / ash / yellow-poplar
401 White pine / Red oak / white ash 805 Hard maple / basswood
409 Other pine / hardwood 807 Elm / ash / locust
500 Oak Hickory Group 809 Red maple / upland
503 White oak / red oak / hickory 900 Aspen / Birch Group
504 White oak 901 Aspen
505 Northern red oak 902 Paper birch
507 Sassafras / persimmon 904 Balsam poplar
509 Bur oak 999 Non-Stocked
513 Black locust

 

 

 

 

 

 

Problem Statement

When there are multiple sources of forest resource information many users of these
information sources try to make comparisons to meet their policy, planning and
management needs. However, the adoption of different definitions for forest land and
forest types makes it difficult to infer similar conclusions in some cases. For example,
FIA and 01 define forest area as land at least 16.7 percent stocked by forest trees of any
size or formerly having had such tree cover, and not currently developed for non forest
use. IFMAP defines forest as land with the proportion of crown cover exceeding 25 % of

the land area (MDNR 2003d). The difference has resulted in classification of a number of

12

non-forest types that are not delineated in FIA (Pedersen 2002). The 01 and IFMAP have
non-forest types like T reed bog, Grass, Rock, Lowland/Upland Brush and Marsh. The
other distinction between the IFMAP and NRIS-FSVeg/OI is that the former will identify
a recent clear cut as the non-forest while the latter will consider it as forest. This is
problematic when only a single date of imagery from all medium- to coarse-resolution

sensors in a landscape where agriculture and forestry are interwoven (Wynne et al. 2000).

IFMAP has classified forest land into the least number of forest types followed by 01
which has 16 categories. NRIS-FSVeg is classified into 60 types, and FIA has 45 forest
types in public forest land in Michigan. FIA plot data can be viewed as the most precise
classification scheme given rigorous data collection procedures at the plot level. Other
schemes are used, but their relationship to FIA classifications is largely unknown. There
are no studies which compare the IFMAP, OI and NRIS—FSVeg classifications with FIA
classifications in Michigan. Taking FIA plot location as the point of reference for
comparisons, I will provide information on the sirrrilarity and dissimilarity between forest
classification approaches. The comparison will be helpful in explaining discrepancies in
total public land acreage of different forest types computed from these data bases. Results
from a comparative forest classification study may be useful developing

recommendations to harmonize the available databases.

Study purpose

The purpose of this research is to compare classifications of forest types between forest

resource databases for public lands in Michigan. The general research questions are:

13

“What are the classifications made by the four data sets? And how consistent are the four
classification systems?”
This study addresses the following objectives:
0 To compare and contrast the four forest resources inventory/database designs in
forest type classification (FIA, IFMAP, 01 and NRIS-FSVeg),
0 To examine consistency among FIA, NRIS-FSVeg, and IFMAP classifications on
national forest lands,
0 To examine consistency among FIA, 01, and IFMAP classifications on state
forest lands,
0 To compare the agreement of IFMAP and FIA classifications on national and
state forest lands, and
0 To explain differences in classifications using plot and stand level characteristics

for the state forest lands.

Organization of Thesis

The thesis has five chapters. A synopsis of each chapter is presented in this section.
Chapter 2, Literature Review, exarrrines previous studies on forest classification
compared against ground truth plots. Previous studies comparing the accuracy of
classification of satellite imagery and forest classifications are reviewed. Studies

particularly relevant to Michigan land use are also reviewed.

14

Chapter 3, Classification Approaches, Data Sources, Study Area and Analysis Methods,
presents background information on forest classifications for FIA, NRIS-FSVeg, 01, and
IFMAP. Details on data sources regarding spatial locations and data attributes are
summarized. Methods used to assess accuracy of the various information sources with
reference to FIA sample plot information are described. Specifically, construction of
error matrices, errors of omission, errors of commission, and Kappa statistics are the

major tools used to classify the accuracy of the databases.

Chapter 4, Results and Discussion, presents analysis of the error matrices and IFMAP, 01
and NRIS-FSVeg accuracy in classifying forest types on the publicly-owned forestland.
Major findings, strengths and weaknesses of the comparisons and data limitations are

discussed.

In Chapter 5, Summary and Conclusions, policy, planning and management implications
and several additional issues on forest classification are discussed. Among them are the
definition of the forest area and inclusion of the non-forested land in classification
systems. Also, policy recommendations to reconcile the forest classification in public

forest land in Michigan are discussed. Additional research ideas are presented.

15

CHAPTER 2

LITERATURE REVIEW

This study compares the forest type classification among the forest management
planning databases of national forest, state-owned forest and the coarse-resolution
state map taking the exact FIA plot location as the reference for comparison. The
previous chapter introduced major public forest management databases in Michigan.
This chapter reviews the land use classification systems broadly, the Michigan
Resource Inventory System (MIRIS), the National Land Cover Database (NLCD) for
the United States, forest classification of FIA/OI/ NRIS-FSVeg, FIA sampling, forest
classification of IFMAP, digital imagery classification, IFMAP imagery classification,
Remote Sensing (RS) based forest cover type studies, and studies related to accuracy

measurement of RS imagery classification.

Land Use Classification

Existing use of land is the main characteristic for defining land use. Anderson et al.
(1976) pioneered land use classification from remote sensing information and outlined
criteria to effectively utilize the remote sensing information for land use and land
cover classification. The land use classification developed by Anderson et al. (1976)
has level I and level H classifications (Table 5). Information at levels I and H are a
basis for national level or statewide aggregation. In addition, more detailed land use
and land cover data, categorized at level III and IV, will be used more frequently by

those who need and generate location information at the intrastate, regional, county, or

16

municipality level (Anderson et al. 1976). Further, land use and land cover

classification level V can be added if a finer level of classification is desired.

Table 5. Anderson et al. (1976) Land Use Classification

 

Level I Level H
1 Urban or Built-up Land 11 Residential.
12 Commercial and Services.
13 Industrial.
14 Transportation, Communications, and Utilities.
15 Industrial and Commercial Complexes.
16 Mixed Urban or Built-up Land.
17 Other Urban or Built-up Land.
2 Agricultural Land 21 Cropland and Pasture.
22 Orchards, Groves, Vineyards, Nurseries, and
Ornamental Horticultural Areas.
23 Confined Feeding Operations.
24 Other Agricultural Land.
3 Rangeland 31 Herbaceous Rangeland.
32 Shrub and Brush Rangeland.
33 Mixed Rangeland.
4 Forest Land 41 Deciduous Forest Land.
42 Evergreen Forest Land.
43 Mixed Forest Land.
5 Water 51 Streams and Canals.
52 Lakes.
53 Reservoirs.
54 Bays and Estuaries.
6 Wetland 61 Forested Wetland.
62 Nonforested Wetland.
7 Barren Land 71 Dry Salt Flats.
72 Beaches.
73 Sandy Areas other than Beaches.
74 Bare Exposed Rock.
75 Strip Mines, Quarries, and Grave Pits.
76 Transitional Areas.
77 Mixed Barren Land.
8 Tundra 81 Shrub and Brush Tundra.
82 Herbaceous Tundra.
83 Bare Ground Tundra.
84 Wet Tundra.
85 Mixed Tundra.
9 Perennial Snow or Ice 91 Perennial Snowfields.
92 Glaciers.
Source: Anderson et al. (1976), Public Database Warehouse, USGS
(http://www.wsdot.wa.gov/environment/envinfo/docs/RSPrj_USGS__lulcclass.pdf)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

l7

However “[t]he Level H is the fulcrum of the classification system as Level H can be
created by aggregating the similar Level IH categories classification system”

(Anderson et al. 1976).

The U. S. Geological Survey (USGS) classification system provides flexibility in
developing categorization at the more detailed levels and provides freedom to the
users to develop categories that meet their particular needs. To retain the compatibility
of the information, whatever categories are used at the various classification levels,
special attention should be given to providing potential users of the data with
sufficient information so that they may either compile the data into more generalized
levels or aggregate more detailed data into the existing classes. Basically, the system
satisfied the three major attributes process: (1) it gives names to categories by using
accepted terminology, (2) it enables information to be transmitted, and (3) it allows
inductive generalizations to be made (Anderson et al. 1976). This classification system

is capable of further refinement on the basis of more extended and varied use.

Regarding the forest land classification, Anderson et al. (1976) stated that “Forest
lands have a tree-crown areal density (crown closure percentage) of 10 percent or
more, are stocked with trees capable of producing timber or other wood products, and
exert an influence on the climate or water regime”. Further, the authors noted when
the “trees reach the marketable size”... they may be harvested and replanted and
“there will be large areas that have little or no visible forest growth. The pattern can

sometimes be identified by the presence of cutting operations in the midst of a large

18

expanse of forest”. Such areas should be included in the forest land category, “[u]nless
there is evidence of other use”. And, “[l]ands that meet the requirements for forest
land and also for an urban or built-up category should be placed in the latter category.
The only exceptions in classifying forest land are those areas which would otherwise
be classified as wetland if not for the forest cover. Since the wet condition is of much
interest to land managers and planning groups and is so important as an environmental
surrogate and control, such lands are classified as forested wetlands” (Anderson et al.

1976).

At Level H, forest land is divided into three categories: Deciduous, Evergreen, and
Mixed. According to the Anderson et al. (1976) classification, Deciduous Forest Land
includes all forested areas having a predominance of trees that lose their leaves at the
end of the frost-free season or at the beginning of a dry season. In most parts of the
United States, these would be hardwoods such as oak (Quercus), maple (Acer), or
hickory (Carya) and “soft” hardwoods, such as aspen (Populus tremuloides).
Deciduous forest types characteristic of wetland, such as tupelo (Nyssa) or cottonwood
(Populus), also are not included in this category. Evergreen Forest Land included all
afforested areas in which the trees are predominantly those which remain green
throughout the year. Both coniferous and broadleaf trees are included in this category.
The coniferous evergreens are commonly referred to or classified as softwoods. They
include eastern species such as balsam fir (Abies balsamea), various spruces (Picea),
white pine (Pinus strobus), red pine (Pinus resinosa), jack pine (Pinus banksiana), and

hemlock (Tsuga canadensis) (Anderson et al. 1976). Evergreen species commonly

19

associated with wetlands, such as tamarack (Larix laricina) and black spruce (Picea

mariana), are not included in this category.

Similarly, Anderson et al. (1976) identified Mixed Forest Land as all forested areas
where both evergreen and deciduous trees are growing and neither predominates.
When more than one-third interrnixture of either evergreen or deciduous species
occurs in a specific area, it is classified as mixed forest land. Where the intermixed

land use or uses total less than one-third of the specified area, the category appropriate

to the dominant type of forest land is applied, whether deciduous or evergreen. Further

to this, Forested Wetlands are wetlands dominated by woody vegetation. Forested

wetland includes seasonally flooded bottornland hardwoods, mangrove swamps, shrub

swamps, and wooded swamps including those around bogs (Anderson et al. 1976). As
the forested wetlands can be detected and mapped by the use of seasonal,
(winter/summer) imagery, and delineation of Forested Wetlands is needed for many
environmental planning activities. For these reasons, they are separated from other

categories of forest land.

Though several classification approaches are available, the land-use and land-cover
classification system devised for the USGS program, developed by Anderson et al.
(1976), has become one of the most widely used classification systems for land use

maps prepared by interpretation of remotely sensed images (Campbell 2002).

20

Michigan Resource Inventory System

The Michigan Resource Inventory System (MIRIS) has two types of inventories about
the State’s land and water resources: (1) a current use inventory to illustrate land cover
and land use and (2) a land resource inventory which includes resources, unique areas,
and areas hazardous to development (Goodwin et al. 2002). The first current use
inventory was compiled from photo interpretation of color infrared aerial photography
(1224,000 scale to 1 inch to 2,000 ft) obtained in 1978/79. Aerial photography
obtained in 1985 was used for the inventory of Detroit and seven highly urbanized
countries in Southeast Michigan. The second Michigan Land Cover/ Use
Classification System (Division of Land Resource Programs 1981) is similar to the
national system developed by the US. Geological Survey (Anderson et al. 1976). It is
multi-level, hierarchical system that classified Michigan’s land cover/use into
approximately 500 categories (Goodwin et al. 2002). The current use inventory is a
subset (approximately 60 level I and H categories) of the Michigan land cover/use
classification system (MIRIS, 1981). However, some of the category numbers and

category definitions were changed.

The MIRIS was upgraded in 2000 by using the 1998/1999 National Aerial
Photography Program (NAPP) imagery, flown by the USGS for the entire state of
Michigan (RS and GIS, 11. d.). These changes in MIRIS version H were made in order
to correct problem areas that existed with the design and application of the earlier

version. Changes in H version were based on several criteria. These criteria included:

21

(l) “describing the major components of each category group within the confines of a
three-level hierarchy, (2) assigning map codes to categories in the lowest classification
levels within each group, (3) creating separate categories for those items that may
need to be cross-referenced under various user defined aggregation schemes and (4)
maintaining clear distinction between “upland” and “wetland” natural cover types”
(MIRIS 2000). In addition, the current use inventory of 22 sampled counties was
supplemented with a detailed land component (level IV and V data which included

species designation, stand size and stocking classification).

National Land Cover Database (NLCD) for the United States

In the last decade, a major provider of land cover information within the federal
government has been the Multi-Resolution Land Characteristics Consortium GVIRLC)
(Homer et al. 2004). The MRLC was originally formed in 1993, to meet the needs of
several federal agencies: United States Geological Survey (U SGS), Environmental
Protection Agency (EPA), National Oceanic and Atmospheric Administration
(NOAA), and US. Forest Service (U SFS) for Landsat 5 imagery and land cover
information (Loveland and Shaw 1996). For NLCD 2001 land cover classification, a
method that optimally classifies many database layers in a single step, with the ability
to document this relationship in a rule base was highly desirable, and the decision tree
classification method was chosen (Homer et al. 2004). The authors used the
commercial decision tree program C5©. They claimed that this decision tree

classification provided an efficient, robust method for classifying large quantities of

22

information in documentable form, and moreover, it allowed them to export mutually
exclusive rules generated by the classification into generic textual rule sets allowing
users access to classification parameters. NLCD 2001 defined land cover into 29 land
cover classes. And forest cover is defined as “areas dominated by trees generally
greater than 5 meters tall, and greater than 20 percent of total vegetation (Homer et al.
2004). This database is developed using the Mapping Zone approach, with 66 Zones in

the continental United States and 23 Zones in Alaska.

IFMAP Classification

IFMAP has a hierarchical scheme of classification of land use. The hierarchical
scheme contains various levels from very broad categories to more detailed ones.
Pacific Meridian Resources (PMR 2001a) stated that the objectives of the
classification scheme of IFMAP are: (1) to provide useful land cover labels for forest
and wildlife management, (2) to provide suitable strata to support stratified inventory
data, and (3) to generate land cover information for use by managers and researchers.
In addition, the IFMAP classification kept “the number of classes of land use to a
minimum, and ensure that classes agree with the definitions of the other ecoregions”
(PMR 2001a). The rule established for this classification system was that each class
should have a local management objective. The vegetation classification rules were
based on percentage of the ground covered by specific vegetation covers. Vegetation
was broken into woody and non-woody with the definition of woody being a plant that

contains a secondary xylem (dicotyledons and conifers), then into shrub and tree. The

23

cut off point between forest and non-forest, and vegetated and non-vegetated was set

as 25% of ground covered by canopy.

The IFMAP classification is very close to the Gap Analysis natural terrestrial cover
classification system. In fact, it is a modified UNESCO natural terrestrial cover
classification that has six hierarchical levels namely: forests, woodlands, shrublands,
dwarf-shrublands, grasslands, and barren (Jennings 1993). The Gap Analysis
classification defined forest as areas dorrrinated with a total canopy cover of 61% or
more, trees crown usually interlocking; woodlands as areas dominated by trees with a
total canopy cover of 26 to 60%, most of the trees not touching each other; shrublands,
dwarf—shrublands and grassland as areas with less than 26% total canopy of tree; and
barren land as areas with vegetation cover less than 5% (Jennings 1993). At level H,
the Gap land use classification combined the land with morphologically similar main
vegetation into a class. For the classes of forests, woodlands, shrublands and dwarf-
shrublands, the similarities were based on evergreen, deciduous and xeromorphic
characteristics (Jennings 1993). The IFMAP level 11 classification has adopted the
60% rule of canopy cover between coniferous and deciduous so that only those stands
that were neither dominated by coniferous nor deciduous fell into mixed stands (PMR
2001a). The Level IH classification of IFMAP was developed based on the majority of
a particular tree species in a forest stand. And the classification scheme has been
viewed as a series of sequential if-then statements. The detail of the IFMAP
classification can be accessed in the Michigan Geographical Data Library website

(http://www.mcgi.state.mi.us/mgdll)

24

IFMAP uses imagery obtained from the TM 5 and ETM 7+. To understand the
IFMAP imagery classification it is important to review the remote sensing techniques
of imagery classification, after first describing on-the-ground management and

research databases like FIA, 01, and NRIS-FSVeg.

Forest Classification of FIA/OI/NRIS-FSVeg

FIA and the forest management planning operational databases, 01 and NRIS-FSVeg,
provide more detailed forest classification (See Chapter 1). These databases have
Level IVN classification. IFMAP has information up to Level HI land use
classification (SI 2004). According to the FIA National Core Field Guide Version 2.0
(2004), the forest types of the Continental US. and Alaska have been classified into
28 forest groups, and there are 140 types that best describe the plurality of stocking for
all live trees. Similarly, based on online FIA condition table information, the public
forest of Michigan has 14 forest groups and 45 distinct forest types (Table 4). FSVeg
has classified the national forests of Michigan into 60 forest types, including lowland
brush, upland brush and open as a type (Table 3). OI has classified state forests into 16
forest types (Table 2). 01 has defined treed bog, grass, rock, lowland brush, upland
brush, marsh, bog, musk, sand dunes, water and other non-stocked or non-forest or

non-productive land as non-forest classes.

There are differences in the minimum percentage of stocking required to define an

area as forest among various databases. The minimum level of forest tree stocking

25

requirement is 10% for FSVeg (Smith et al. 2004), 16.7% for FIA (Hahn and Hansen
1985) and for 01 (Pedersen, per. com..), 25 % for IFMAP (Michigan Geographical
Data Library, SI, 2004). Thus there are big differences in inclusion and non-inclusion
of areas with lower levels of forest tree stocking. For example, the NRIS-FSVeg has
forest type as lowland brush and upland brush and 01 and IFMAP has defined the

lowland brush, upland brush, treed bog, as non-forest areas.

FIA Sampling

For this study, FIA information is serving as the ground truth for assessing the
accuracy of forest type classifications of the databases. Therefore, it is important to
understand the nature of FIA. With passage of the 1998 Farm Bill, formerly known as
the Agricultural Research, Extension, and Education Research Act of 1998 (PL 105-
185), Congress required that the Forest Service conduct annual forest inventories in all
states (McRoberts 1999). This Bill made some changes in FIA sampling procedures
and intensity. For example, the Farm Bill established requirements that (1) each year,
20 percent of total plots are to be measured in each eastern state and 10 percent of
plots are to be measured in each western state, (2) the annual data are to be made
available each year, and (3) statewide resource reports are to be published every five
years with integration of FIA and the Forest Health Monitoring (FHM) program

(McRoberts 1999, Brand 2005). FI-IM is a national program that uses data from

26

ground plots, aerial surveys, and other sources to produce annual estimates of status

changes and trends in indicators of health.

FIA inventories are commonly designed to meet the specified sampling errors at the
State level at the 67 percent confidence limit (Miles et al. 2001). FIA precision
standards require sampling intensity of one plot for approximately every 6,000 acres in
the North Central region (McRoberts 1999). To satisfy this requirement, the
geographical hexagons established for the FHM programs were divided into 27
smaller FIA hexagons, each of which contains approximately 5,900 acres. A grid of
field plots was established by selecting or establishing a plot in each smaller hexagon:
(1) if an FHM plot fell within a hexagon, it was selected as the grid plot; (2) if no
FHM plot fell within a hexagon, the plot from existing network of permanent FIA
plots that was nearest the hexagon center was selected as the grid plot; and (3) if
neither the FHM nor the existing FIA plot fell within the hexagon a new permanent
FIA plot was established at the hexagon center and selected as the grid plot
(McRoberts 1999, Brand 2005). The grid of plots is called the federal base sample and
is considered as an equal probability sample. In this way, FIA uses grid sampling
(Schreuder et al. 2003) that covers a l-acre sample area (Miles et al. 2001). Recent
inventories use a common sampling design consisting of four 24.0-foot radius subplots
(approximately (1/24th acre) for trees at least 5 inches in diameter and four 68me
radius microplots (1/300th acre) for smaller trees (Miles et al. 2001). Another

characteristic of the new design is the mapping of differing forest conditions. If two or

27

more conditions occur within a plot, the boundary between them is mapped and the

proportion of the plot in each condition is recorded.

Forested plots are installed and measured regardless of intended use or any restrictive
management policy. After the adoption of the national plot design in the mid 1990’s,
all FIA units have implemented a common sampling design consisting of four 24.0-
foot radius subplots (approximately 1/24th acre) for trees at least 5 inches in diameter
and four 6.8-foot radius microplots (approximately 1/300th acre) for smaller trees
(Miles et al. 2001). In this way, tree expansion factors are approximately 6 for trees at
least 5 inches in diameter and approximately 75 for smaller trees. Subplot l is the
center of the cluster with the other three subplots located 120 ft away at azimuths of
360°, 120°, and 240°, respectively (Miles et al. 2001). In addition, the temporal
regularity was incorporated by systematically assigning each hexagon to one of 5
inter-penetrating panels. Plots located in panels 1 to 5 hexagons were to be measured
in first to fifth years, respectively. Once the five- or ten-year cycle is complete, the
sequence starts again. In fact, FIA inventories are extensive inventories that provide
reliable estimates for large sampling areas. As data are subdivided into smaller and
smaller areas, such as geographic unit or a county, the sampling errors increase and
the reliability of the estimates goes down. There are nine tables in the FIADB (Forest
Inventory and Analysis Database) Version 1.0. In this study Condition table and Plot

table information is used.

28

Digital Imagery Classification

Digital image classification is the process of assigning pixels to classes (Campbell
2002). The aim of the classification is to assign each pixel in an image to a distinct
cover class or theme, broadly called an “information class” (Foody 2003). By
comparing pixels to one another, and to pixels of known identity, it is possible to
assemble groups of similar pixels into classes that are associated with the
informational categories of interest to users of remotely sensed data (Campbell 2002).
The informational classes are the categories of interest to the users of the data.
Informational classes are, for example, different kinds of forests or land uses that
convey information to the remotely sensed data users (e. g. policy planners, resource

managers, the scientific community and publics).

Unfortunately, the information classes are not recorded directly on remote sensing
images; they can only be derived indirectly by using evidence contained in the
brightnesses recorded by each image. A group of pixels that are uniform with respect
to the brightnesses in their several spectral channels is called a spectral class
(Campbell 2002). Thus, remote sensing classification proceeds by matching spectral
categories to information categories. In general, each pixel is assigned into the class
that it most nearly resembles or to which it is closest, according to some measure of
distance in a broad sense. In this way, information classes are typically composed of

numerous spectral subclasses.

29

Unsupervised classification is defined as the identification of natural groups, or
structures, within spectral data (Campbell 2002). In unsupervised classification these
natural groups are defined, identified, labeled and mapped. On. the other hand,
Campbell (2002) informally defines supervised classification as the process of using
samples of known identity (i.e., pixels already assigned to informational classes) to
classify pixels of unknown identity (i.e., to assign unclassified pixels to one of several
informational classes). Samples of known identity are obtained from those pixels
located within training areas, or training fields. Pixels located within these areas from
the training samples are used to guide the classification algorithm to assign specific
spectral values to appropriate informational classes (Campbell 2002). However, there
is no unique criterion to use as the basis for classification; different criteria yield
different classifications resulting in a trade-off between accuracy and efforts (Foody

2003).

Chuvieco and Congalton (1988) developed a hybrid method which combines the
training statistics generated from both supervised and unsupervised classification
approaches using two multivariate statistical techniques: cluster and discriminant
analysis. This method was used to classify the pixels for IFMAP. In this method, the
cluster analysis is used to group together training statistics generated from both
classification approaches. The strength of these groupings was then tested using
discriminant analysis. “The cluster analysis is not simply a reduction process for the
unsupervised approach, but rather a way of combining similar groupings from both the

supervised and unsupervised approach and the groupings that contain both supervised

30

and unsupervised statistics provide a powerful match between the informational

categories and the spectral classes” (Chuveico and Congalton 1988).

The cluster analysis proposed by the authors uses the hierarchical method in which the
classes are merged progressively, two in each step, until all classes belong to the same
cluster or group. This grouping can result either in a merging of two single classes, or
in a class being merged with an already formed grouping, or in merging two
groupings. At the end of the process, the user selects the step at which the clustering
stops, usually when two classes very remote in the distance matrix are merged. The
squared Euclidian distance is selected to calculate the distance between the classes.
Once the cluster analysis of the training statistics is completed, discriminant analysis is

used to evaluate the strength of the grouping.

Discriminant analysis is a multivariate technique that attempts to find a new set of
functions which maximize the ratio of the variance between and within groups
(Chuvieco and Congalton 1988). The analysis involves a linear transformation of the
original variables, which are orthogonal, in such a way that the new functions
maximize the separation between the already formed groups. After these functions
have been found, it is possible to regroup each one of the original training classes to
test its membership in the correct grouping. After running a discriminant analysis on
the groups formed by the clustering of the training statistics, each one of the final
groups was defined by the average value of its members. Then these average statistics

were input into the classification algorithm to be used in the assignment phase of the

31

classification process. In this way, by the clustering and discriminant analysis, “an
improvement in the classification results is expected because of the improved

grouping of training statistics” (Chuvieco and Congalton 1988).

In IFMAP, classification schemes were first developed for the Southern Lower
Peninsula in the spring of 2000, and then updated to reflect the differing ecoregions of
the Northern Lower Peninsula and the Upper Peninsula in the winter and spring of
2001 (PMR 2001a). Imagery were acquired for different seasons namely, spring (leaf-
off imagery), summer (growing season) and senescence imagery (fall images). The
images were used to enable separation of species based on their phenological
differences, that is, different cropping cycles and moisture regimes that allowed
differentiation of these categories (e. g., aspen from northern hardwoods and from
oaks). A listing of the classification schemes and the rules developed toderive the
classes can be accessed via the Michigan Spatial Data Library website
(http://www.dnr.state.mi.us/spatialdatalibrary/sdl2/land_use_cover/2001/

IFMAP_lp_landcover.htm).

The acquisition of imagery from Landsat 5 and Landsat 7 sensors for the rest of the
state of Michigan took place in Fall 2001. In 2002, imageries from both sensors were
collected in leaf-off, mid-season, and senescence for the Northern Lower Peninsula
and the Upper Peninsula. A stratified approach was adopted to obtaining the training
data. The training data were collected comprising reflectance samples, taken from the

images to be classified of the cover types to be classified. These reflectances may

32

differ from region to region, because of phenologic differences, soil variability,
hydrologic differences, and elevation changes, though the latter was not a big factor in
the State of Michigan (PMR 2001a). To obtain an adequate stratification of training
samples, a coverage of ecosystems was intersected with a coverage of the TM scene
boundaries, and the resulting polygons were named eco-scenes (PMR 2001a). Eco-
scenes of interest were identified and specific areas were chosen for field visits. Since
the IFMAP project is specifically concerned with gaining information on forested

lands, efforts were made to identify areas of high natural land cover diversity.

Training fields are areas of known identity delineated on the digital image, usually
specifying comer points of a square or rectangular area using line or column numbers
within the coordinate system of the digital image. Specific training areas need to be
identified for each informational class following the standard guideline of image
classification (Campbell 2002). The key characteristics of training areas are number of
pixels, size, location, number, placement, and uniformity (Campbell 2002). According
to PMR (2001a), while performing the training of IFMAP imagery, the raw
information, such as the canopy closure of the three size classes, various shrub,

wetland, and herbaceous species, were recorded from training sites.

Field data were digitized onto the imagery using the air photographs and the manual
delineations (on the mylar overlays) as guides. The supervised training sites were
extracted from this image using a region-growing method. The region growing method

ensured a spectrally pure signature, which was appropriate for the supervised

33

classification. It involved specifying a spectral threshold for inclusion in the area of
interest (AOI), which eliminated spectrally outlying portions of the sites. After
digitizing, the training sites were used to recode a single band. of a TM mosaic, with
unique identifiers for each training sample. This layer was used to extract training
statistics from any of the imagery mosaics and/or derivatives without re-drawing the
training sites. As the atmosphere can substantially impact the signatures derived from
imagery, scenes were selected that had a minimum of haze and clouds and no
atmospheric correction was made. Images of Southern Michigan were divided into
areas with similar date scenes, usually taken on the same date. They were mosaiced,

and areas that represented different conditions on the ground were classified separately

(PMR 2001a).

The 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 the first iteration, the summer
image mosaic is classified using the ISODATA unsupervised method (PMR 2001a).
The resulting classes were exarrrined with the supervised classes using an
agglomerative hierarchical procedure by the application of Mathsoft 2000 (PMR
2001a). This procedure begins by considering each signature, the pixel reflectance
value, as a separate group; then it combines and divides groups based on spectral
similarity until all signatures are 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 is then examined for matches of supervised and unsupervised

34

signatures, and in the case of a close one-to-one relationship, the unsupervised
signature in question can be labeled with the land cover label of the supervised
signature. After the tree is fully examined for clusters of this type, the labeled classes
are subset from the imagery, and the procedure is run again. This procedure continues

until it is no longer advantageous to do so.

The cluster analysis method was used initially on the entire image to achieve an
Anderson Level I classification, and subsequently on the individual level I cover types
to classify the land cover to the desired level. Once the imagery was subset to level I,
the signatures derived from the individual subsets will contain less overall spectral
diversity, and as a result the subtle differences were more evident and therefore easier
to separate. Level H classifications were based on analysis with leaf off imagery and

Level HI classification were derived using different masks and use of texture bands.

Remote Sensing-Based Cover Type Mapping Studies

Ross Nelson and his colleagues (1987) developed a method to assess the continental
forest land cover using Landsat Multispectral Scanner (MSS) data. In their study, the
authors assessed the Anderson et al. (1976) level H classification for conifer and
hardwood forest area in the continental United States. The authors used a stratified
random sampling (SRS) approach to allocate Landsat MSS scenes country-wide so
that the aerial extent of the conifer/hardwood resources of the United States could be

evaluated (Nelson et al. 1987). The authors generalized a priori information in the

35

form of a major forest cover type map (USGS 1980) and used it to stratify the country
into six forest strata, namely northern conifer, northern hardwood, central hardwood,

southern conifer, western conifer, pinyon-juniper and one non-forest strata.

The MSS scenes were allocated to each forest stratum with the constraint that each
contain at least two MSS scenes (Nelson et al. 1987). Within each MSS scene, four
200 X 200 pixel sample blocks were located systematically, and each block’s forest
cover was identified. The authors’ selection of fewer large blocks was derived from
their preliminary study. The sample blocks were first located in the Landsat MSS data
and on 1:250,000 USGS topographic map sheets and to facilitate the acquisition of
National High Altitude Photography (NHAP). NHAP are color infrared photos flown
at a scale of 1:58,000. These photos were used to help identify spectral classes in MSS
data and to refine the digital land cover classifications. The authors used most of the
photos which were acquired leaves off, facilitating conifer/hardwood identification. A
multicluster blocks procedure was used to classify the MSS data. In this method, for a
given scene, small blocks of MSS digital data ranging from 40 X 40 to 60 X 60 pixels
were clustered into spectral classes using an unsupervised classifier using IDIMS
(Interactive Digital Image Manipulation System, ESL 1978). Each spectral class was
identified using ancillary data (NHAP, B&W photo quad, false CHI composites)
(Nelson et al. 1987). Forest was defined as any spectral class which included forested
area >30% canopy closure. Conifer or hardwood forest was defined as a forested area

where > 50% of the canopy was coniferous or hardwood, respectively.

36

The Landsat MSS data products and the statistical estimates of the conifer, hardwood,
and water resources of the continental United States were evaluated to determine their
reliability. The authors used the two different assessments to characterize the accuracy
of this MSS classification. The first assessment compared the MSS classification and
airphotos on a point by point (pixels by pixel) basis. The second assessment compared
aerial estimates of the four cover types derived from the MSS data and from the
corresponding areas on the airphotos. A third assessment was carried out to determine
the accuracy of the MSS-based national estimates of conifer, hardwood, and water
compared with national estimates generated by other US. government agencies. The
findings of this study revealed that the national estimates of conifer and hardwood
derived using this sampling method is within 3% of the total USDA Forest Service
acreage. Comparison of the MSS classification products and airphotos showed that the
conifer cover class was correctly identified 74% of the time and hardwood 80% of the
time (Nelson et al. 1987). The average classification accuracy countrywide for the four
types considered (conifer, hardwood, water and “other”) is 74%, the overall accuracy

is 85%.

Pax-Lenney et al. (2001) developed a generalized classifier method to monitor
temperate conifer forests, ultimately at the global scale with Landsat TM and ETM+
data. Within this context, the generalization refers to a concept in which a classifier is
trained with data from one domain, but applied to data from different domains (e. g.,
different geographic location, time, and /or imaging sensors). The authors mentioned

that analytical methods based on image-by-image interpretation are too time-

37

consuming and labor intensive for studies of large areas to be undertaken with any
degree of frequency. The authors found generalization is well suited for multitemporal
classifications of one Landsat scene using simple dark-object-subtraction (DOS)
atmospheric corrections to produce classifications with comparable accuracies as
classifications from the more complex radiative transfer corrections, based on over
200 classifications. However, the high degree of variability in the classification
accuracies underscores the importance of extensive, in-depth analysis of remote
sensing techniques and applications, and highlights the potential problem for

misleading results based on just a few tests (Pax-Lenney et al. 2001).

Miguel-Ayanz and Bi ging (1997) compared the performance of TM and SPOT data
for cover type mapping on the Central Sierra of Spain. The authors used three single-
stage and one multistage iterative classification in their study. The three-single stage
classifications were: (1) supervised classification with band selection by spectral
analysis, (2) supervised classification with band selection with spectral separability
indices, and (3) supervised classification with prior probabilities and band selection by
spectral separability indices. The multi-stage supervised classification was performed
using an iterative classification method. In this method the class that attains the
highest average accuracy (Congalton 1991) in each iteration is masked and classified
as part of the image. In this way, GIS analysis was used to obtain the prior
probabilities for classes being discriminated and these probabilities were used in the
band selection and classification processes. However, prior probabilities did not

significantly improve the band selection process. In most cases the best band

38

combinations were selected both by the weighted (including the prior probabilities)
and nonweighted separability indices. The best overall accuracy, 66%, was obtained
for TM data with the iterative classification approach. Accuracy of 61% was the best
overall accuracy for SPOT data, which was obtained with the iterative classification
methods. For TM imagery, the five most abundant classes, which account for over

72% of the study area, where classified with 90% overall accuracy.

Similarly, Wang et al. (2003) compared the dry season ETM+ and 1—m panchromatic
sharpened IKON OS imagery classified as tree canopies and open area taking the latter
imagery as valid ground truth to assess the tropical deforestation in the in the
Amazonian state of Mato Grosso, Brazil. The authors found the squared correlation
coefficients (R2) between the canopy cover values derived from ETM+ and H(ONOS
were 0.92 and 0.96 at the 30-m and 90-m scales respectively. Thus, the authors argued
ETM+ imagery can be used to estimate canopy cover across large areas of tropical

forest.

Silbemagel et al. (1997) compared the distribution of the landscape measure among
landtype association groups in historic (1840’s) and present (1990-1992) landscapes in
the Eastern Upper Peninsula of Michigan comprising six counties: Alger, Chippewa,
Delta, Luce, Mackinac and Schoolcraft. In addition, the authors compiled quantitative
information on landscape metrics, to supplement existing qualitative descriptions of
landtype associations (LTA) in the study area. Cover type boundaries between each

section line were interpolated using elevation lines, surface geology maps, and other

39

early vegetation maps. The authors used the cover classes based on expanded MIRIS
land cover codes. Prevalence or dominance of cover classes was based on class area
(CA) and landscape similarity index (LSIM), or class area weighted by total landscape
area. “In the four physiographically based land type association groups studied:
bedrock-controlled, lowland sand lake plain, morainal origin, and outwash-northem
hardwoods and mixed conifer were most prevalent cover types of the 10 studied,
historically and currently” (Silbemagel et al 1997). Northern hardwoods were
especially prevalent in the moraine groups, while the mixed conifer type was more
prevalent in the bedrock group. Wetlands and mixed pines, in addition to northern
hardwoods, were also prominent in the lowland group. In the outwash group, these
types were also present, but mixed and white pine were more prevalent. Largest single
patches (LPI) were found redundant to the LSIM, and therefore were not assessed to
the same extent as other indices. The highest LPI values were found in the northern

hardwoods in the moraine groups.

Skole et al. (2002) developed a model “Forecast Michigan” to forecast the land use or
urban sprawl in the context of spatial decision system support. The authors claimed,
“the Forecast Michigan models are process models using most of the state’s standard
GIS data layers, as well as inputs from economic and demographic models.” They also
utilized network analysis algorithms to include transportation routing and traffic
demand with enough spatial resolution and sensitivity to provide transportation
planners a way to evaluate different corridor alignments and access points in the

context of secondary and cumulative impacts on land use change in a project region.

40

There are a large number of studies about detailed forest classification, see for
example, Saatchi and Rignot (1996), Mayaux and Lambin (1997), Martin et al.(l998).
However, they are based on high spectral resolution remote sensing data like Synthetic
Apertures Radar (SAR), NOAA’s Advanced Very High Resolution Radiometer
(AVHRR) and Airbrone Visible/Infrared Irnagin g Spectrometer (AVHlIS),

respectively.

Accuracy Measurement Methods Used for Remote Sensing Imagery

Classification

Stehman (1999) reviewed several basic probability sampling designs useful for
accuracy assessment. According to the author, the first step in choosing'the
appropriate sampling design is to “define the population for which the accuracy
assessment is needed, and to determine if this population will be partitioned into
pixels, polygons, or some other aerial unit. Then the probability sampling design
forms the statistical foundation of the assessment.” Basically, “[c]hoosing a design
from among a basic probability sampling design options should be guided by the
project objectives and the relative importance of other remaining design criteria”
(Stehman 1999). In addition, the criteria to consider when planning the sampling
design are that “the sample should: (1) satisfy the probability sampling protocol, (2) be

simple to implement and analyze, (3) result in low variance for the key estimates of

41

the assessment, (4) permit adequate variance estimation, (5) be spatially well

distributed, and (6) be cost effective” (Stehman 1999).

Congalton (1991) mentioned “researchers and users of remotely sensed data have a
strong knowledge of both the factors needed to be considered as well as techniques
used in performing any accuracy assessment”. The accuracy assessment task can be
defined as one of comparing two maps, one based upon analysis of remotely sensed
data (the map to be evaluated), and another based upon a different source of
information (Campbell 2002). Basically, the second map is designated the reference
map, assumed to be accurate, that forms the standard for comparison. The reference
data are of obvious significance; if they are in error, the attempt to measure accuracy

will be in error (Campbell 2002).

The simplest method of evaluation is to compare the two maps with respect to the
areas assigned to each category and the result of such comparison is to report the areal
proportions of categories. These values report the extent of the agreement between the
two maps with respect to total areas in each category, but do not take into account
compensating errors in misclassification that cause this kind of accuracy measure to be
itself inaccurate (Campbell 2002, Congalton and Green 1999). In addition, this form of
error assessment is sometimes called non-site specific accuracy, because it does not
consider agreement between the maps at specific locations, but only the overall figures
for the two maps. The second form of accuracy, site-specific accuracy or classification

error, is based upon the detailed assessment of agreement between maps at specific

42

locations (Campbell 2002). This computation is performed by comparing a sample of
locations on the map with the same locations on the reference data and keeping track
of the number of times there is agreement (Congalton and Green 1999). In the
majority of analyses, the units of comparison are simply pixels derived from the
remote sensing data, although if necessary, a pair of matching maps can be compared
using any network of uniform cells (Campbell 2002). After the maps are evaluated on
over all accuracy, the need to evaluate individual categories within the classification
scheme is recognized, and so began the use of the error matrix to represent map

accuracy (Congaltan and Green 1999).

An error matrix is a square array of numbers set out in rows and columns which
express the number of sample units (i.e. pixels, clusters of pixels, polygons) assigned
to a particular category in one classification relative to the number of sample units
assigned to particular category in another classification (Congalton 1991, Congalton
and Green 1999). As noted, one of the classifications is considered to be correct (i.e.
the reference data) and may be generated from aerial photography, airborne video,
ground observation or ground measurement. The columns usually represent this
reference data, while the rows indicate the classification generated from the remotely
sensed data. In this way, the error matrix is a very effective way to represent map
accuracy in that the individual accuracies of each category are plainly described along
with both the errors of inclusion (commission errors) and errors of exclusion
(omission errors) present in classification (Congalton and Green 1999, Congalton

1991). Sometimes “the error matrix is referred to as a confusion matrix because it

43

identifies not only overall errors for each categories but also misclassifications (due to

confusion between categories) by a category” (Campbell 2002).

In addition to clearly showing errors of omission and commission, the error matrix can
be used to compute other accuracy measures, such as overall accuracy, producer’s
accuracy and user’s accuracy (Story and Congalton 1986). Overall accuracy is simply
the sum of the major diagonal (i.e., the correctly classified sample units) divided by
the total number of sample units in the entire matrix and this is the most commonly
reported accuracy assessment statistic and is probably most familiar to the readers
(Congalton and Green 1999). Producer’s and user’s accuracy are ways of representing

individual category accuracies instead of just the overall classification accuracy.

The inspection of the error matrix only reveals the overall nature of errors present;
there is often a need for more objective assessment of classification (Campbell 2002).
For example, if we are interested to know “are the two maps in agreement?” —this is a
question very difficult to answer without the help of just an error matrix. The notion of
agreement is difficult to define and implement. The error matrix is an example of a
more general class of matrices, known as contingency tables. Some of the procedures
that have been developed for analyzing contingency tables can be applied to

examination of the error matrix (Campbell 2002).

Congalton (1981), Congalton et al. (1983) and Congalton and Green (1999) proposed

application of techniques described by Bishop et al. (1975) and Cohen (1960), a

discrete multivariate technique, as a measure of improving interpretation of an error
matrix. A shortcoming of the usual error matrix is that even chance assignments of
pixels to classes can result in surprisingly good results, as measured by percentage
correct (Campbell 2002). Hord and Brooner (1976) and others have noted that the use
of error matrix accuracy measures is highly dependent upon the samples, and therefore

upon the sampling strategy used to derive the observations used in analysis.

Kappa is a discrete multivariate technique used in accuracy assessment for statistically
determining if one error matrix is significantly different than another G3ishop et al.
1975, Congalton and Green 1999). The result of performing a Kappa analysis is a

KHAT statistic

Observed — expected

 

A
k =
1 — expected

There are numerous studies on the assessment of the accuracy of remotely sensed data
using the error matrix that calculate overall accuracy, producer’s accuracy, user’s
accuracy and the Kappa statistic. Several studies that are relevant to this study were

found.

Appropriate sample size requirement for the Error Matrix is one of the very important
aspect, however “[a] balance between what is statistically sound and what is
practically attainable must be found” (Congalton and Green, 1999). Further to this, the

authors mention, a general guideline or good “rule of thumb” is to collect a minimum

45

of 50 samples for each vegetation or land cover category in the error matrix. However,
for the larger area (i.e. more than a million acres) or the classification with a large
number of vegetation or land cover categories (i.e., more than 12 categories), the
minimum number of samples needs to be increased to 75 or 100 samples per category

(Congalton and Green, 1999).

Berlanga-Robels and Ruiz-Luna (2002) studied land use change mapping and change
detection in the costal zone of Northwest Mexico using remote sensing. The authors
used a multitemporal post-classification study with data from Landsat Multispectral
Scanner (MSS) and TM to detect landscape changes. The authors compared four
thematic maps (1973, 1986, 1990 and 1997) and classified the land-use into six classes
as direct indicators of landscape condition. The accuracy of the classification (only in
1997 scene) was calculated from an error matrix, using overall accuracy assessment

and the Kappa coefficients (Berlanga-Robels and Ruiz-Luna 2002).

Lawson (n. d.) used the National Resource Inventory (NRI) data, a compilation of
natural resource information on non-Federal land in the United States, with Landsat
TM scenes in Iowa. The author demonstrated the use of NR1 point data for image
classification and assessment of accuracy of a broad cover/use digital layer (TM
scence) obtained from the Iowa Department of Natural Resources participating in the
GAP program. The author used accuracy assessment tools including an error matrix,

overall accuracy, user’s accuracy and producer’s accuracy.

46

Schreuder et al. (2003) made a number of recommendations for accuracy assessment
of percent canopy cover, cover type and size class. The authors recommended to “well
define vegetation types, stand size and canopy cover percentage to... [e]xplore the use
of ambiguous classes, compute the contingency table and Kappa statistics for each

of. . .. mapped categories, producers and users accuracy”.

Kurvonen and Hallikainen (1999) studied the accuracy assessment of multitemporal
ERS-l and JERS-l synthetic aperture radar (SAR) images in Finland test sites. The
authors used the confusion matrix for land-cover type and forest type classification
accuracy assessment. The authors mentioned that the use of textual parameters

significantly improved the classification of land-cover and forest type classification.

Liu et al. (2003) compared the neural networks and statistical methods in classification
of ecological habitats using FIA data. The authors used two artificial neural networks
(ANN) and three traditional statistical classification methods to classify FIA plots into
six ecological habitats in the US. Northeast and found four variables (overstory,
understory species composition, hardwood basal area percentage, and current FIA
forest type) as the most important discriminating variables for habitat classification. In
this study also the authors used classification accuracy, Kappa statistics, and a

classification success index to compare the classification of ecological habitats.

Wicham et al. (2004) assessed the thematic accuracy of the 1992 National Land Cover

Data (NLCD) for the six western mapping regions of United States. The authors

47

collected the reference data in each region for a probability sample of pixels stratified
by map land-cover class. The authors assessed the thematic accuracy using overall

accuracy percentage and an error matrix for Anderson Level H classification.

Wessels et al. (2004) compared the classification of the Moderate Resolution Irnagin g
Spectroradiometer (MODIS) with the existing Landsat TM land cover maps as
reference data for two major conservation areas (Greater Yellowstone Ecosystem-
GYE, USA and the Para State, Brazil). In this study, the Landsat TM land cover was
processed to their fractional composition at the MODIS resolution (250 m and 500 m).
The authors used the error matrix, overall accuracy, producer’s accuracy and user’s
accuracy to assess the accuracy of MODIS thematic maps. The findings of this study
suggested, in GYE, the MODIS land cover was very successful at mapping extensive
cover types (e. g. coniferous forest and grasslands and far less successful at mapping
smaller habitats (e. g. wetlands, deciduous tree cover) that typically occur in patches
that were smaller than the MODIS pixels. For Para State it was successful at
producing a regional forest/non-forest product (Wessels et al. 2004). However, a
single 500 m MODIS forest/non-forest product cannot be expected to reflect all the
complex human impacts on biodiversity such as secondary regrowth, local land-use

matrix dynamics and low intensity logging.
Similarly, Powell et al. (2004) studied the sources of error in an accuracy assessment

of thematic land-cover maps in the Brazilian Amazon. The authors tried to quantify

the subjectivity in reference data labeling and compared reference data produced by

48

five trained interpreters. In addition, the authors identified the impact of other error
sources, including geolocational errors between the map and reference data, land-
cover changes between the dates of data collection, heterogeneous reference samples,
and edge pixels. By the findings of this study the author suggested “ (1) labels of
continuous land-cover types are more subjective and variable than the commonly
assumed, especially for the transitional classes]; (2) validation data sets that include
only non-mixed, non-edges samples are likely to result in overly optimistic accuracy

estimates, not representative of the map as a whole” (Powell et al. 2004).

Katila et al. (2000) introduced a statistical calibration method aimed at reducing the
effects of map errors on multisource forest resource estimates. The authors developed
a correction method based on the confusion matrix between land use classes of the
field sample plots and the corresponding map information with the empirical example

from the ninth National Forest Inventory of Finland.

There is a key concern that “the land cover maps derived are often judged to be of
insufficient quality for operational applications” (Foody, 2003). For this reason, there
is a need to compare the forest management databases available in Michigan.
However, comparing the different forest management databases/maps, generated for
specific management purposes, is a challenging task because of differences in
objectives, spatial resolution and in the definitions of forest type classification.

Different government entities like the MDNR, national forests and Forest Service

 

' However, using multiple interpreters to produce the reference data classification increases reference
data accuracy.

49

research are using their own forest type classifications based on their management
objectives. In this study, IFMAP is a thematic map and the other databases like FIA,
01 and NRIS-FSVeg are developed based on field observation. In other words, the
FIA, 01 and NRIS-FSVeg are the databases developed from on—the-ground
observation and IFMAP is generated from space observation. In addition, the spatial
resolution of the each database is different. For example the 01 spatial resolution is
900 m2 and FIA is 4 X 1/24th acre. The NRIS-FSVeg and 01 spatial resolution are
based on timber stand sizes, the smallest forest management unit. Objectives of the
NRIS-FSVeg and 01 database are more or less the same; however, they are not
spatially overlapping because of different ownerships. In addition, the IFMAP is a
coarse-resolution map, derived from Landsat TM with cluster analysis (Chuvieco and

Congalton 1988) using FIA data as the ground truth.

A number of authors (see for example, Katila et al. 2000, Wessels et al. 2004, Wicham
et al. 2004, Kurvonen and Hallikainen 1999, Lawson, n. d.) have compared the
accuracy of thematic maps with the other thematic maps, topographic maps, or
ground truth data collected for different objectives. The review of previous studies on
the accuracy of thematic maps suggests that, regardless of differences in various
databases, the comparison of the accuracy using the standard accuracy assessment
statistics like overall accuracy, producer’s accuracy, user’s accuracy and the Kappa
coefficient of agreement can be used for this study. The rigorous field data collection
of FIA and its exact plot location provide a sound basis for comparative analyses for a

large number of forest types representing the majority of the forest area of Michigan.

50

CHAPTER 3

DATA SOURCES, STUDY AREA, CLASSIFICATION APPROACHES

AND ANALYSIS METHODS

This research was a collaborative project among the Department of Forestry at Michigan
State University, the Michigan Department of Natural Resources, national forests of
Michigan and USDA-FS Forest Inventory Analysis unit at the North Central Research
Station. Details of the data sources, study area, analysis methods are presented in the

following sections.

Data Sources

Forest resource management and planning databases were obtained from the USDA
Forest Service and the Michigan Department of Natural Resources. The NRIS-FSVeg
information was obtained from the Huron-Manistee National Forest, Hiawatha National
Forest and Ottawa National Forest. 01 data for state-owned forests, which included the
compartments with YOE 2003 to 2006 available in digitized GIS environment, were
obtained from the Forest, Minerals and Fire Management Division of the MDNR. The
national forest data were in vector GIS and were re-projected into Michigan Georef State
Plane Coordinate System 1983 using the Projection function of ArcToolbox of ArcGIS
Desktop (8.2), (Environmental Systems Research Institute, Redlands, California, USA).

The OI data were already in Michigan Georef State Plane Coordinate System 1983.

51

Similarly. IFMAP, land cover 2001 which was in image format, was downloaded from
the Michigan Spatial Data Library (SDL). The image file of IFMAP was transferred to
the grid file using the Irnport-Export function of Imagine 8.7 (Leica Geosystems GIS &
Mapping, LLC, St. Gallen, Switzerland ERDAS). IFMAP is in Michigan Georef, so the

raster data was not re-projected. IFMAP land cover 2001 satellite imageries were taken

during 1997 to 2001.

The data fields of the NRIS-FSVeg, national forest database, and 01, state-owned forest
were carefully reviewed. Initially, a limited number of the stand parameters were chosen
which may be useful in explaining differences/similarities of the forest type among the
databases. Four key attributes of the stand were selected for each database. From NRIS-
FSVeg the selected stand attributes were size density, stand dbh, stand age and survey
year. From OI, the stand attributes of size density, stand age class, total basal area and
understory type were selected. All the selected attributes were aggregated (generalized) to
help maintain confidentiality of FIADB. By law, the USDA-PS must protect the
confidentiality of FIA plot locations. Spatial location of forest for both NRIS-FSVeg and

01 were aggregated to broader categories.

Similarly from FIADB, two plot-level information variables were selected, namely: (1)
measurement year, and (2) kind code, and three attributes of stands were selected from
the condition table, namely: (1) code for forest type of the condition (assigned by field
crew), (2) aggregated stand age-the average total age, and (3) growing stock stocking

code. The Ecological subsection code, forest types derived from an FIA algorithm and

52

attributes information about present level of stocking similar to stand dbh from NRIS-
FSVeg and total basal area from OI were initially selected. These attributes were dropped
later from the analysis due to confidentiality concerns. The revision of the attributes
useful in explaining differences in classification results was based on the potential
problems in maintaining the confidentiality of FIADB. In the end, only the equivalent of

forest type, from three databases, was selected.

Study Area

Out of the 19.28 million acres of forestland in Michigan, 7.14 million acres of forestland
is owned publicly (Smith et al. 2004). Public ownership of forestland is distributed
among the Federal government (National Forest, Bureau of Land Management and
other), State government, counties and municipalities. In Michigan, there are three
national forests, namely the Hiawatha, Huron-Manistee and Ottawa, covering about 2.68
million acres (Smith et al. 2004) (Figure 1). Forestland under State ownership is about
3.95 million acres (Smith et al. 2004). This study covers all national forests and the state

forest land owned by the MDNR for which 01 information has been digitized.

53

   

Ottawa N atio nal Forest

 

33 Natlon al Forests
- State-owned Forest, MDN R a, ‘ z; . ,
” ‘ ,,.,.____ County boundary .2, . W» M" My 7. . "9‘“ .. 3 _ .1. ,J

50 0 50 100 150 Miles
E

Figurel. National Forests and Digitized State-Owned Forests in Michigan

Classification Approaches

When every information source or map or database has its own classification system,
making comparisons between and among them is not straightforward. The classification

of the reference database or map needs to be regrouped by making it more compatible

54

with other databases. In order to make forest classifications more compatible to each
other, all the available forest types in each database were enumerated and a separate list
of forest classification of each database was made (See Tables 1 to 4, Chapter I). Then
pair-wise crosswalk tables were made by the author, and they were reviewed and revised
based of expert judgment by experienced professionals working at Michigan State
University (MSU, Drs. Donald Dickmann and Larry Leefers), the MDNR (Dr. Larry
Pedersen and Mr. Jason Stephens) , and USDA-FS (Dr. Mark Hansen, NCRS and Mr.
Joseph Gates, HMNF). Feedback and comments were incorporated to create the final

crosswalk table (Appendix A).

Analysis Methods

After compiling agency data, the spatial and stand attributes were aggregated at the
Forest Social Science and Economics Lab at Michigan State University (MSU). Methods
to pick up data field from the grid and vector data were developed in the ESRI ArcView
3.2 environment. There were two types of data in this research. The NRIS-FSVeg and 01
were in vector data shape files, and IFMAP data was in image or raster data. To compile
the forest classification of these two types of data set for a point location (e. g an FIA plot)
two approaches were proposed. For vector data the Mila Grid extension was downloaded
from the ESRI website, which can be used to pick up the grid value from underlying
raster data; this picks up the forest type values from the raster layer (IFMAP). Similarly
for vector data sets, a script was downloaded from the ESRI website, which can be used

to pick up the data field from the vector data. In this way, two methods were proposed to

55

append the stand level data field like forest type, and the other variables selected to
explain the difference in the forest types at the sample points. By overlaying the FIA plot
vector data with the plot level and stand level information requested for this study as
additional data fields to the FIA plot location, all the studied databases’ forest type and

the other selected stand level attributes could be added to the FIA plot level record.

In this study, Mr. Geoff Holden, North Central Research Station (NCRS), carried out this
data compilation. Data extraction at the FIA plot locations was performed using two
approaches depending on the data format. For vector data (01 and NRIS-FSVeg), a
point in polygon overlay was performed using the "Intersect" tool in ArcGIS 9.0 to pick
up the 01 and NRIS-FSVeg forest classification. For raster data (IFMAP), the "Extract
Values to Points" tool (from the Spatial Analyst extension in ArcGIS 9.0) was used to
pickup the IFMAP classification. In this way, Mr. Holden carried out theoverlay
operation and combined the NRIS-FSVeg, 01 and IFMAP information at the exact
location of the FIA permanent plots. The resulting database table had a plot number
(sequential, l to n), the FIA field crew defined (FLDTYPCD) classification, the 01 or
FS-Veg classification, and the IFMAP classification. Hence, each plot/record had data

from three sources.

In this study, the approach to study remote sensing thematic mapping classifications was

adopted to examine the site-specific accuracy of classification of both the satellite

imagery and the 01 and NRIS-FSVeg forest classification databases.

56

An error matrix compares the reference condition to the classified condition. Assume that

n samples are distributed into k2 cells where each sample is assigned to one of k

categories in map or database and independently (usually in the. rows), to one of the same

k categories in the reference data set (usually the columns) (Figure2).

i = rows

(classification)

Column total

n+j

 

 

 

 

 

 

j= columns
(reference) row total
2 ni+

H11 H12 H13 nlk n1+
H21 H22 1123 112k n2+
H31 1132 1133 113k n3+
nkl 11kg nk3 nkk nk4-
n+1 n+2 n+3 n+k n

 

 

 

 

 

 

Note: Adopted from Congalton and Green (1999).

Figure 2. Illustration of an Error Matrix or a Confusion Matrix

Following Congalton and Green (1999), let ni- denotes the number of samples classified

into categories i (i = 1, 2, 3,. k) in the remotely sensed classification or any

classification under comparison (for e.g. OI, NRIS-FSVeg and IFMAP) and category j (j

=1,2,3, ...... k) in the reference data set (FIA data).

k
Let ni+ = Znij (column sum)
Fl

57

be the number of samples classified into category i in the remotely sensed
k
classrficatron, and n+ j = Zlnij (row sum)
I:

be the number of samples classified into category j in the reference data set.
The overall accuracy between the defined classification (IFMAP, NRIS-NRIS-FSVeg

and OI) and the reference data (FIA) was computed as follows:

 

 

k
2 “ii
Overall accuracy 2 1:1 (1)
n
Producer’s accuracy can be computed by
n if
Producer’s accuracy j: —— (2)
And user’s accuracy can be computed by
“if
User’s accuracy ,- = (3)
ni+

The user’s accuracy (UA) and producer’s accuracy (PA) can be used to calculate the error
of commission and error of omission. Errors of omission (EO) refers to the samples of a
certain class of the reference data that were not classified as such and errors of
commission refers to the samples of a certain class of the classified data that were

wrongly classified ( J anssen and Van der Wel 1994).

58

Campbell (2002) applied the reference classification column sum as the denominator to
calculate the Errors of Commissions however, J anssen and Van der Wel (1994) published
the following relationships
User’s accuracy % = 100% - Errors of Commission (%), and (4)
Producer’s accuracy (%) = 100% - Error of Omission (%). (5)
A commission error is simply defined as including an area into a category when it does
not belong to that category and an omission error is excluding that area from the category
in which it truly does belong (Congalton and Green 1999). In this way, “every error is an
omission from the correct category and a commission to a wrong category” (Campbell

2002).

Cohen (1960) developed a coefficient of agreement (called Kappa statistic) for nominal
scales which measure the relationship of two classifications beyond chance agreement to
expected disagreement. This measure of agreement uses all cells in the matrix, not just

diagonal elements.

Again, let p ,1. denote the proportion of samples in the i,jth cell, then corresponding to n ,1. .
In other word
Pij = ”if, n.
Then let pi, and p“. be defined by
k

Pi+ = Zpij and
i=1

59

k
p+j = Zprj
i=1

The estimate of Kappa (K) is the proportion of agreement after chance agreement is

removed from consideration, that is,

K = (P0 _ pc) (6)
(1'- pc)
in which

p0 = proportion of units which agree

p, = proportions of units for expected chance agreement, and

X,-
po = Zpii , Pc = 291417”), pij : %
+ represents summation over the index (Rosenfield and Fitzpatrick-Lins 1986, Congalton

and Green 1999) where N = total number of counts in matrix and X i}. = number of counts

in ijth cell.

For computation purposes

k
”Znii - Zni+n+i
i=1

[2 .-= (7)

"2 " Zni+n+i

 

with "ii , ni+ and n+i as previously defined (Congalton and Green 1999).
The KHAT values are a measure of agreement or accuracy (Congalton and Green 1999).

Cohen mentions I? = 0 when obtained agreement equals chance agreement. Positive

60

values of Kappa occur from greater than chance agreement; negative vales of Kappa are
from less than chance agreement (Rosenfield and Fitzpatrick-Lins 1986). Congalton and
Green note the KHAT values can range from +1 to -1. However, since there should be
positive correlation between the forest management databases (NRIS-FSVeg, 01 and

IFAMP) and the FIA, positive KHAT values are expected.

The approximate large sample of Kappa is computed using the Delta methods as follows.

 

 

)4 620— 612+) 2(1— 62X26262— 62) (1— 62) (24-40;)

Var(K
n (1" l92)2 (1" 492)3 (1 92)4

(8)

1
where 01 = — 211,-,-

1 k
91=—Z"ii
"1:1

1 k k
64: "32 Zn nij(nj++n+i)2
i=1j=1

The variance of KHAT was calculated by using LabView 7.1 (National Instruments) Mr.
Murari Regmi, a physics graduate student of MSU, wrote the program as per the author’s

instruction.

61

The test statistic for testing the significance of a single error matrix is expressed by
131
A - (9)
t/Varl K1 ’ ‘

Z is standardized and normally distributed. Given the null hypothesis H0 : K1 = 0, and the

Z:

 

alternative Hl : K 1 at 0, Ho is rejected if Z 2 20,,2 , where 012 is the confidence level if

the two-tailed Z test and the degrees of freedom are assumed to be cc (infinity).

The KHAT values are a measure of agreement or accuracy. A KHAT value is computed
for each error matrix, and it is a measure of how well the classification of a forest
management database (NRIS-FSVeg, OI and IFMAP) agrees with the FIA forest
classification. This provides a means for testing the significance of the KHAT statistics
for a two independent KHAT values, and therefore two error matrices that are
significantly different. With this test, it is possible to compare each forest, management
database with the FIA forest classification and the classification with higher accuracy was

identified.

The test statistics for testing if two independent error matrices are significantly different
is expressed by
|K1 - K2|

Z = JVar(I?l)+ Var(K2).

 

(10)

Z is standardized and normally distributed (Congalton and Green 1999). Given the null

hypothesis H0 :(Kl — K2) = 0, and the alternative Hl :(Kl — K2) at 0, H0 is rejected if Z

62

_>. Z a / 2 , where (1/2 is the confidence level if the two-tailed Z test and the degrees of

freedom are assumed to be cc (infinity).

Chapter Summary

In summary, this chapter outlined the procedures to assess the accuracy of the thematic
maps or databases with any reference data sets, which may be ground truth field data or
another thematic map or any other source of reliable information. In this study, data were
obtained from the MDNR, the Spatial Data Library of Michigan and national forests of
Michigan. These data were compared for accuracy using the FIA field crew defined
forest type classification (FLDTYPCD) as the reference for comparison. The research
team requested the Forest Service North Central Research Station (NCRS) to compile
each forest type classifications (FIA, FS Veg and IFMAP in national forest and FLA, 01
and IFMAP in state forest) taking FIA exact location as the point of reference for
comparison. The objectives of consistency assessments of the NRIS-FSVeg in the
national forest lands and 01 in the state forest lands and IFMAP in the public forest land
of Michigan were carried out by computing the error matrix and computation of the
overall accuracy, producer’s accuracy and user’s accuracy and Kappa statistic. Each pair
of forest classification NRIS-FSVeg, 01 and IFMAP was compared for the agreement in
classification by comparing the KHAT statistic. The Z test was carried out to test two
independent KHAT values, and therefore two error matrices, for significant differences.
By performing this test, the forest classification of higher accuracy was identified both

for national forest land and state forest land.

63

CHAPTER 4

RESULTS AND DISCUSSION

There are four primary sources of forest management information for public forest land
in Michigan, namely IFMAP, 01 (for state forest lands) and NRIS-FSVeg (for national
forest lands) and FIA, each with different characteristics (Table 6). IFMAP is a statewide
raster map at the resolution of 30 meters by 30 meters and has information on land cover
classification. 01 is a forest resource management database developed by the MDNR
with the objective of supporting its day-to-day operational activity relating to resource
management issues and land use (MDNR n. d.). OI has detailed information in databases
covering information ranging from trees resources to special wildlife practices. Similarly,
NRIS-FSVeg which is one of several natural resource information systems (NRIS)
developed by USDA-FS, has stand-level information for the federally managed national
forest land. NRIS-FSVeg has stand examinations, inventories and regeneration survey
information to support management of national forests. It has wide range of information
ranging from the trees to animal habitats. Besides these, the FIA sample plots have

extensive information at the individual plot level.
Selected aspects of these four databases are summarized in Chapter 1. The overall

comparison of these forest management information databases in terms of defining forest,

coverage, objective and other details is summarized in Table 6.

64

8088802 0:": ::0 03:08:): 00:0": 0: SEE": ”0:02

 

A050:

 

:30: 00:: :0:0:00000:0 :80
80:0 :85 “000808: 03:00:80: 80880:: .wdv >858 .m.D
0: :00 8: .:0:0w0:ww0 80809808 80:: 80809808 :00:0: 8:08:80 0:: 000:00 :0:00:::000_0
:0>0: 00:000 0:02 3:80:30 :80 0:005 0:08 :80 :0:0:0Q 0800 :80 :0=0:0Q :00:0: :0 09¢.
08800

380:0 80808802

380:0 80808802

880:0 80808802

038080 80:: “003:0:
0.0 8088:: 800

89:08: :0 08:0 Z

 

 

 

 

 

 

 

 

 

 

 

 

 

as. 03:00:.
:00:0: 0:0:0 :80 000:0 :09»... 380800 00:>:0m 00:0":
080w 8:23: 98:20:: .2; :3 :0w0:08 <QmD .3 :0w0008 000:0:0Q :0 0w0:0>00
88:82 :0 0:0:m :88: 3:300:05 0:00:8 :0:0:0Z .m.D _0::0:::00 0:::m
::080m0:08 80:0»0000 :00:0: 3:30
8:00:08 8752 0:80 :0 80808808 :00:0: 08:0: 0:00:0: .m.D :0
:0 80802: 088 00:00.00: :88: 0 :0 0:00: 80808808 080:: :00 080:0 0:: :0
:0:::0::0 0:: 80800 0 H :0 :00: 0:: :008 0 2:. 008000: :00:0: :008 0 H 8:08:88 000000 0 H 80:30: :0 03:00.30
:: cm:
:000: :0 :0 523 :30:0 :00:0: 00 8:00:
823 8088:: oz 0 5:3 :38: :0 088m 0: 823 8:882
3:00 VNNNd: 0800000 0:00 0:0 3:80:00 0:00 0:0 3:80:00 0:00 0:0 088 :0 00:0 8:882
.368 :0 A30:
:0 88m: 3:80:03: 80:03 :80 803: 00:
0%: 3:08:80 :0 80:80: :00:8:0: :0: 8:803:
:0>00 :0 088080908 0: =3: :0:: :8: :0>00 8:080 :0: :80
058 82:: :0>00 :8: :00» 000:: :03. :0: 3:088: .:0>00 00:: :000 :0:
:88» 0:: :0 00mm :0: :: :0 on 80:80:: :05 :8: 8:00:08 .080 w::>0: 3:088: :0 .080
:000: :0 80080 30:00 :0 050800 8:0: :00 80 :0 000:: :00:0: a: 80 :0 000:: 088 a: 00:0": :0 :0::8:0Q
000:: :0 8:80:08 3:090 80.8 :000: :< 8:080 8 o: :000: :< 3:080 :0 0.8 :000: :<
:0>00 80:82:95: :0 J0>meEZ Sn: 0:0:0Q
3

8085:): 8 0000:80Q 8:08:88 ::080w0:0:2 00:0": :0 80.88800 .0 030:.

65

80035 8:33? 0: mg :80 ::080w0:0:>: 0::": :80 80:08:): :00:0n: 0: 2.1.9:": ”807:

 

:0::00::000:0 E :0>Q:

03:00.30 ::080w0:08

 

 

 

 

 

 

 

 

 

 

:0 0:0: :00:00 38080: 02:00.30 ::080m0:08 :00:0: :000: 8800:0000
0:: w:080 :00:0000 0008000: :0::0: :008 0: 80:0 08:0 :0>0: :0:: :0 808000 808000
:0 :0>0: 0::0:::000< :008 0: 808000 3:0,: :0: 808000 0:30:00008 :008 :80 ::0:0n: :0 :0>0: 3:080:80:
:0::08000000 008000:
:80 80808808 80:80:80
8800:0000 8.0: 0:83 0:; ::08:0::>:0
08:00:80 :0:00:: :80 SEE»: 875:): 0:: 0:83 mmédmb 0:: 000:0:0: 0::: :0
8:3 mm: 008808.: 8:85 :00: 3:80:00 8:85 :00: 3:80:00 0:00: :0 8:80: 080.: 8:00:83~
:0::00:0
:00:0:0w0> 0:00:80 :0: 00 :80? .0088 :800
: 0:0: :0::0 .0008 .w0x008 .wo: 800:: .80
3:83.30: \ :0::0 80:9: :08: 80:30: 00:03 .0::::0 .0w:::0n:0 :00:00:0 :07: 0:80 :0:00:::000:0
85888 00000088.: 80:08 0:00: .0w0m .080: 3:080:07: 07: 82080: 8:38: :07:
:98 0:2 :88 :0 0050.082 S: 0:500
3:

30.800: .0 030:.

66

In defining the forest area, FIA, NRIS-FSVeg and OI are based on the stocking of the
forest trees. OI has the requirement of 20 cu ft of timber production per annum per acre.
In contrast, IFMAP defines a forest based on the percentage of tree canopy cover. As
discussed in Chapter 1, IFMAP and OI have forest types for the lower density cover or
stocking of trees. Regarding the objective of each database, 01 and NRIS-FSVeg are very
similar to each other. On the other hand, the FIA objectives are very different, and
IFMAP’s objectives are very broad. Similarly, regarding the coverage, the IFMAP is a
“wall-to-wall” coverage of the entire state of Michigan, and 01 and NRIS-FSVeg covers
the state forest land and national forest land, respectively. Although, the FIA has only

sample plots, it covers the continental US.

From an inventory point of view, 01, NRIS-FSVeg and IFMAP are very close to each
other in that they yield a map. In contrast, FIA is extensive in terms of coverage and
intensive at the sample unit level. In terms of forest type classification, FIA and NRIS-
FSVeg are more detailed and 01 forest classification is more specialized to suit the needs

of forest management in Michigan. IFMAP has Level II and HI land cover classification.

The users of these databases are different; 01 and NRIS-FSVeg are used within the state
and federal forest management agencies. IFMAP and FIA data can be used for a variety
of reasons by large numbers of users. The anticipated levels of accuracy of these
databases are very close to each other. It will be fair to say that the accuracy of these
databases are more related to amount of resources spent to create them. In summary,

there are multiple sources of forest management information in Michigan, developed for

67

specific purposes. Prospective users should use either one or multiple sets of databases
based on their coverage and objective of use. Only FIA and IFMAP information have

complete statewide forest coverage in Michigan. To the extent'that these databases are
compatible to each other, they will provide more robust information for forest resource

and wildlife habitat management and for sustainable use of ecosystem services.

Using FIA Forest Types as the Reference Classification

FIA has sample plot information, which are called “conditions” that were derived by “the
discrete combination of landscape attributes that define the condition” (Miles et al. 2001)
of the particular sampling unit. Forest type is one of the important information items
included in the condition attributes. In addition, FIA has tree information that describes
each tree over 1 inch in diameter found on subplots. In this study, the FLDTYPCD (a
forest type assigned by FIA field crew) information from the condition attributes of FIA
plots was used as the ground truth forest classification for assessing the accuracy of

IFMAP, NRIS-FSVeg and 01 forest type information.

The FIA plots were established using a systematic grid approach (USDA-F8 2004). The
sample units for this study were chosen exactly at the FIA plot center location, so the
sample units selected for this study could be taken as a sub-population of systematic
sampling. The sets of data were filtered by a complex process of FIA data filtering before
release. After completing the data filtering process of FIA, NCRS released the

classification of 239 FIA sample plots (out of 397 possible) for state forest land and 523

68

sample plots (out of 753 possible) for national forest land (National FIA Database
System). The 239 datasets for state forest land was due to the partial digitization (only
40%) of the state forest land database in 01. Accuracy assessments of this study were

based on these sample units.

As discussed in Chapter 1, there were differences between the FIA, OI/NRIS-FSVeg and
IFMAP classification schemes. While developing the crosswalk table between
classifications, one-to-one relationships between the reference classifications to the
“map” classification were expanded to include cases where categories were “acceptable”
or “probably right”. In a previous study, Congalton and Green (1999) used a similar
approach in the California hardwood rangeland monitoring project. In this way, one-to-
many relationships between the reference classification and “map” classification were
identified for some reference classes. The most likely category is referred to as the
“primary classification”, and “acceptable” or “probably right” category is referred as the

“secondary classification”, of classified map hereafter.

In this study, for example, Jack pine in FIA may be reasonably comparable to Jack pine
and Oak-Jack pine (secondary) in NRIS-FSVeg. The author had only the FIA field crew
defined forest classification, the IFMAP classification and OI or NRIS-FSVeg

classification. Due to FIA confidentiality concerns, further details from any data sources

would have significantly reduced the number of sample plots available.

69

Congalton and Green (1993) identified eight factors that can affect error matrix or
difference matrix results, and these factors provide useful framework for interpretation of
the results of this study. The factors were: (1) land-cover change between the time of
reference data acquisition and satellite data acquisition, (2) reference sample location
error, (3) reference label data entry error, (4) reference label photo interpretation error,
(5) inconsistent labeling of reference data due to land-cover heterogeneity surrounding
the sample location, (6) difference in map and reference data registration, (7 ) map
delineation error, and (8) map classification error (W ickham et al. 2004 cited as
Congalton and Green 1993). Wickham and his colleagues (2004) also realized the
usefulness of these factors for the interpretation of the results of error matrix. In this
study, mainly factors (4) and (5) were important. To deal with these factors, measures
were taken while developing the forest classification crosswalks between the reference
classification and the classified map or database classification. This measure permitted
matches for a number of reference forest types to multiple “map” forest classification
types as a potential agreement between the two classifications. In other word, this
assumption had made one reference forest classification to many classified forest
classifications as “acceptable” classifications. For example, the Black Spruce of FIA was
matched with the Lowland Coniferous Forest, Lowland Shrub and Mixed Non—Forested
Wetland. These types of relaxation in matching of two classifications may overstate the

accuracy of a map or database.

In addition, FIA represents “the four 1/24th acre” subplots, IFMAP represents 900 square

meters of area and the 01 and NRIS-FSVeg represents an entire stand with a particular

70

forest type. In other words, the relative size of the area represented by these
classifications was different. A recent study by Smith et al. (2002) revealed that
“accuracy decreases as land cover heterogeneity increases and-as patch size decreases”.
Further the author emphasized these landscape variables remain significant factors in
explaining classification accuracy. As in this study, both the information on classified
map/database and the reference map/database were secondary source information, so the

author had to rely on a crosswalk between classifications to make comparisons.

In the forest classification crosswalk tables (Appendix A), there were multiple matches of
FIA forest classification with a classified classification of IFMAP, 01 and NRIS-FSVeg
and OI/NRIS-FSVeg as reference classification with IFMAP as the classified
classification. This type of one-to-many relationship between reference and classified
map limited direct formulation of the Error matrix and consequently the assessment of

the chance agreement and calculation of the KHAT statistic.

The accuracy assessments in this study were made from the difference classification
matrix. A user’s accuracy of 98.0% (Table 7) for Lowland Conifer Forest means that 50
of 51 of the pixels classified as Lowland Conifer Forest were lowland conifer forest types
according to FIA. User’s accuracy is sometimes called “reliability” in the comparison
scheme (J anssen and Van der Wel 1994). For example, 18 of 37 Jack pine plots (FIA,
101) are classified as Pines or acceptable Lowland Coniferous Forest (Table 7).
Similarly, dividing the number of correctly classified samples by the total column total

yields the producer’s accuracy: it indicates the percentage of samples of certain

71

(reference) class that were correctly classified in the comparison scheme (J anssen and

Van der Wel 1994).

Accuracy Assessment for State Forest Lands (MDNR)

To assess the accuracy of the IFMAP classification, FIA and 01 classifications were
considered as reference classifications. And to assess the agreement between 01 and FIA
forest classification, the FIA forest classification was used as reference classification. In
other words, the accuracy of IFMAP and 01 classification was assessed based on FIA
classification (Tables 7-9). IFMAP overall accuracy was found to be 63.6% with FIA as
the reference classification and 60.3% with 01 as the reference classification. The overall
accuracy of 01 was 84.5% compared to FIA as the reference classification. There were
differences between the FIA, 01 and IFMAP classification schemes that limited the
formulation of the error matrix and simultaneously calculating the chance agreement and

calculation of KHAT statistic.

The match between the reference classification and classified classification were not
along the diagonal as usual. The matched cells were shaded to identify them (Tables 7-
12). Producer’s and user’s accuracy are in the last row and column of the difference
classification matrix. The user’s accuracy for the Lowland Conifer was more than 94% in
comparison with FIA and OI (T ables 7-8). Thus, IFMAP did well in identifying the
Lowland Conifer Forest type. The user’s accuracy for Oak Association was as low

a323.5% in comparison to the FIA classification (Table 7) and 29.4% in comparison to 01

72

582 .25 ”8533 2&5 8m .25
525 322208832. swam .8» ”Boga; 2.2% 852 5% 532.58 8mg; .mon $80 32 .5 gases. ea
82% x85 mNfi an Ema—am Nanci and .NA: REE x02. .2: ”movoo <3 $8.58“ {8: .<D S858“ {8355 .45 “802

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

v.23 seam 23m he Sm Ba “22": Co E32 8:20th .5 22¢

850.8 13‘ 0.0m ode mNn od 1:. NS 09 Www m.mm w.wm QNV es <m
>858.“ =Eo>0 0N N m aN o m vm d mN _ 2 3 5332
Nm fl QMN om v n O». _ h an _ 0N m 2 NM .88.
ode m m _ ~ N _ 98:03 388-:02 U222 om
w.mm N. 2 v m a c ~ 2.5% 9:233 wN
93 On gm .Nm ~ 2 fl _ u 680m anchor—50 cam—33 mN
od 0 v N m _ 620m 3022009 6:233 «N
9% N. 3 m ~ ~ _ N v 320% @822 9833 NN
ed 0 _ _ muotcou can—ab @822 N
v.2. 5N fin N _ _ m 2 E 3:5 a
0.8 _N mm “N N N. _ N ~ _ seafloommxx 5&3. @—
mdN v S 2 N v confloom?‘ Mao 3
wNm mm ow m N m wN ~ _ couwmoommxx 3
803633 505.52
06 o cm 0 v ooh Nm
£2333 \ seam Baa:
od o 2 N a a 2 22:80 asoaeom 2
935 :2“: fig :5 4 8m . 8m 8w .on .. 1 8m . .5. mos. , m2. 92 ENE .N ..__.§ ,w ,. .. $.23. ., .. 88
$96 932 cop—880% n52 BEmmflu

73

5&2 .88 ”@529; 2%.: 8m .SN
525 32328333? swam .§ ”80322 22% 352 .on ”beige 8.3:? New ”58 32 .NN~ gasefi .eN_
603mm xoflm mN~ "E Ema—mm _N~ 65m com .NS 65a :03. .2: ”380 <5 .3538 €me J5 S0238 €02.85 .<m ”802

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

swam N.NN 9% 82 0.3 od o2 N.NN 2: NS N.NN 3N 3N 9N <m

N858“ =§>o mm N m «m o N mm _ _N 2 2 NN 6.22

NON NNN an a. m 9. _ N am _ 0N N : NN 38
NS ON _N 2 2 o N 5:80 0

953m c222
ode n n m 555 2233 4
od o _ _ €32; Neaam m
32 N N N am Bo; o
N.NN mm we R N n _ _ N _ N :83 <
N.NN :. 3. m N n E cooauam Eofioz 2
2% N N_ _ N 4 x8 0
N.NN N ”N mN m 380 0
2x: a _ _ c.8355 N
32 N N N 88% 335 m
N.NN 2 M: N 2 N 2E Em m
EN NN .m _ m NN 2% 8.2: N
N. <2 53: 38. 8a 8m 8» SN ONm mom NN. 0N. mNN ,5 N2 :: 538530 8 88
GEE NEE oococflom 838mm 35820

€53 awoken 88m 5.. Sm 98 HO mo x532 855:5 .w 2an

74

28825.2. . .2. "ooEqm 2225 .m 2223.5 22222223012 {2 32220 .0 60032832 502222072 .2 622222 3m .2 6:222 282. .2. 68392222 aEaBm .22
220222200 953m 3x22 20 ”wom c822. .Q 2.222on .0 E2232 .< $350 20 4822222822 M2822 .<D 2:222 2222222822 M23255 .<n2 202072

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

$26 od 82 82 N22. mdN Nee N62. od 22.222 Nee now N.NN 92. <2
522222822 22.22220 22 N m m 2m N2 22 o 2.2 N mN eN , 228.222 280.2.

3.2 amN 2 2. m N2 $2 2 2m 2 2N m wN we 2230.2.
odv N m 2 2 2 2 2 222222203 282022-252 322 cm
w.mm 2. m2 . 2 v 2 2 N v 2222.225 222222302 wN
22.2% N2. on e 2 2.2 N N N 28292 38022280 22222233 mN
dd o m 2 2 2 N 2820.12 3026289 222222302 X
9% N 2.2 2 N 2 v 2 m 2882 @9222 22822222 NN
od o 2 2 30222200 2x22222222 232222 2N
3N eN 3 N 2 N2 22 N 2 N 8:22 2
ode 2N mm N a 2 2 2 2N 2522222082 2222922 e2
de m 2.2 m N2 seafloommax 2220 2
m.N2. aN ov mN 2 02 220222228322 22

28032522 22222252
od o 2 v e 8022 N2
5288-33 2 222% 22.222
od o 2 N m 2. m 9222222220 222288222022 02
is 202.2% 2.25.2. .2‘ m: .2 o 2 2 2 m2. .0. .Q o s<. . 22.2.22. .. use
2202 823222229 85.20232 22222 32.22350

32222.2 282cm 83m .82 20 can n22”: 20 “22.22222 8:80.225 6 0322.2.

75

classification (Table 9). Thus, IFMAP did a poor job in identifying the Oak Association.
It was confused with Aspen in both comparisons. Regarding the producer’s accuracy,
IFMAP did well in identifying the Northern White Cedar (87.1%) and Black Spruce
(88.5%) of the FIA classification and Cedar (89.3%) and Mixed Swamp Conifer (81%) of
the 01 classification. However, for Northern White Cedar and Black Spruce of FIA and
Cedar and Mixed Swamp Conifer of 01, the crosswalk allowed multiple IFMAP
classifications as “acceptable”. This is one of the reasons for the small error of omission
in these categories of classification. Due to lack of further information, the author could
not verify any existence of map classification error in these forest classification

categories.

IFMAP was inaccurate in identifying Aspen (38.2 % PA for 01 and 44.1% PA for FIA)
and Jack pine (48.6 % PA) for FIA and Oak (41.7% PA) for 01. In classifying Aspen,
IFMAP had confusion with the Northern Hardwood and the Oak Association (Table 9)
and the Oak Association and low density cover forest types (Table 7). However, due to
the small number of samples in the Oak classification, the results of this study were not
conclusive and require further verification with adequate number of samples to infer

about Oak confusion with other categories of IFMAP.

From the difference matrix of 01 and FIA, for most of 01 classification there were
similarities with the FIA classification. From an 01 data user’s perspective, Oak had the
least user’s accuracy. The Oak classification of OI had confusion with the Jack Pine. In

nature these two species are often found together. This type of Oak trees in Jack pine

76

forest type of FIA was reported in Michigan Forest Statistics (Table 47), 1993
(Leatherberry and Spencer 1996). However, the small number of sample units and lack
of other stand information from the OI database relative to FIA plots, made it difficult to
derive further conclusions. There was also evidence of confusion in OI in classifying Red
pine and Jack pine. Similarly from the producer’s perspective, 01 had confusion in
identifying Jack Pine with Oak, Red Pine and Aspen, and Red Pine had confusion with

Jack Pine.

Accuracy Assessment for the National Forest Lands

To assess the accuracy of IFMAP classification, the FIA and NRIS-FSVeg classifications
were regarded as reference classification. And to assess the agreement between NRIS-
FSVeg and FIA forest classification, the FIA forest classification was used as the

reference classification (Table 10-12).

The overall accuracy of IFMAP, without reducing chance agreement, was 64.8% taking
FIA as the reference. The agreement of forest classification between the NRIS-FSVeg

and FIA without reducing chance agreement was 82.2%.

By comparing user’s accuracy, IFMAP was very accurate for the Lowland Coniferous

Forest (89.6%), the Northern Hardwood Association (86.0%) and the Pines (82.7 %)

relative to FIA (Table 10). And it did well for Lowland Deciduous Forest (87.5 %), and

77

692%. 228 2.2 2 28 223 gm an 32.2; 2 v.8 2.32 2 2E2 223 .29.
x285 2E U2.6 .82. ”53223 Eofioz .82 82% 28:2 .22 83% 223 .815 Egan 222282522 Egan .82
”0:222 833 EBmam 82 ”02:22 2232 .N02 6:222 2222. .2222 “mo—”coo «2.2 30225822 M283 .<D 2238258 9206:2205 .32 ”202

 

3
2:3 :28 22232 52 Sn 2.5 “252,222 2c 5822 08202.25 .2: 2.3

78

dogmas. 28 22252222: 2 2222: 2.32
6222228 232222220 5825:2292 8285

Sw 22.85.23 2 2&5 Eu: .822 ”2225 30:02 2 .288 2 2&5 $5 .28 aaeo 222222 2 58m 2 2%: .822
20N U228 2.0.2 E0582 6% 328 82223 .va 28268 <2n2 22258.2 M283 .<D ”208:8“ $283.92 .32 “3672

o

 

 

 

 

 

 

 

 

 

 

 

 

  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

23.2% a.wN QmN N.2a N.vN m.mw N.w2 odw QmN § <2

2858“ 22890 NN N 22 82 m N v N 228222

amm H mNm oN w N2 922 o 2 2 w w .502.
mdm a _. 02 N 2252220? 2882-52 26322 cm
N.Nw a _, N2 m N £55 25233 MN
9% mv wv m N 23.52 962022250 2.522302 mN
QmN N m 235m m=o=2228n2 22:22:02 vN
N.2m Om N 28292 2.3222 2.52 D NN
od . 202280 22 D 852 N
ed 82280 2.5: D 550 cm
N.Nw 62.2 m 2 2 85.2 a2
od 2 38289 E2 22 8522 M:
0.2 a 2m a 2 2 en 2 25222226894. 222232. 02
N mv 02 mm m 2 2 25222683. 622.20 2
o cm 222 aN2 v2 2 .H 2i22. via H v 220222003422 25032222222 E0552 E
o N 2 80.: 229.028.2502 2 235m 2652 D N2
30322522 222

. H 0250

 

 

 

_ .32

22322802 .02 22222.2.

79

Table 11. Difference Matrix of NRIS-FSVeg and FIA in National Forest Lands
(a)
Classified Reference ( FIA)
101 102 103 105 121 122 125 127
1 ack Pine . 45 5 l
2 Pine 3 l
3 Pine
5
11 Fir-
lack
l4 Wh Cedar
18 S conifer
19 -
Northern
w

ack Pine-Oak
Pine-Oak
lack Oak
Red Oak
Oak
I Ash-Elm-R

Beech -YB
Basswood

)

-beech-YB/red

Upland

l
l l
otal 62 98 1 1 31 13
33 13
A % 87.1 85.7 1 83.9 82.5 1 75
Note: PA, Producer's Accuracy; and UA, User's Accuracy. FIA codes: 101, Jack pine; 102, Red
pine; 103, Eastern white pine; 105, Eastern hemlock;121, Balsam fir;122, White spruce; 125,
Black spruce; 127, Northern white-cedar; 400, Oak Pine Group; 401, White pine / Red oak /
white ash; 503, White oak I red oak/ hickory

 

80

Table 11. (Cont'd).
(b)

l 800 801 805

1 ack Pine
2 Pine
3 Pine
5
11 Fir-
12 lack
l4 Wh Cedar
18 S conifer
19 -

Northern

w
48 ack Pine-Oak

49 Pine-Oak

53 lack Oak

55 Northern Red Oak
59 Oak

7 l I Ash-Elm-R

81 Beech -YB
82 Basswood
84 )

85

87 maple-beech-
birch/red

89 Upland 47

91 57
93 14
95 9
97 9
99 8
otal 5 ll 12 8 7 523 4
, 5 3 l 8 2 61 Overall
A % 75. 1 27.3 66.7 89 66.7 25.0 80.3 82.2%
Note: PA, Producer's Accuracy; and UA, User's Accuracy. FIA codes: 504, White oak; 505,
Northern red oak; 701, Black ash / American elm / red maple; 800, Maple / Beech / Birch
Group; 801, Sugar maple / beech / yellow birch; 805, Hard maple / basswood; 809, Red maple.
/ upland; 901, Aspen

 

81

22220 2.2222 .an 262220 2.32 E05202 .nm 2622.20 U2222.2 .mn 222220.322 2.32 .av .2220- 0:22 2282 .222. 232222 E05522 2.22232 .2N
222.222.2232. .3260 .222 2..”22228 nEuBm 22222 .222 522200 223 E05202 .222 283.2% U222:2 .N2 292-2225 2:32am. 22 222828022
2222222822 222622222822

      

82

 

8222832 2822232220
2222

.283 285.2 2282 82 832-322 2.5 .2222": 2o Ea: 822202.25 .2 2.2.2.2.

:80 8 Ewan 2.5233 .3 2.2m 83% 3.23.. .3 2:222 £8.32 .8 2.28.... 2:22.30 .3 632.: 2.5.2.5
263222 .3 232.2% 220222225 3232282262928 szm .5 2 2922 Swan .mwbevoEaE 2.32 .vw22oooammam 62222222 m .Nw
222% I 22039022222222 m .2w 22.222222 2.822222 5 .2 N 28268 wo>mu2-w2~22 28222822 M2822 .<D 22622222822 “2.2322255 .<N2 ”Boz

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

some can N.NN N8 Nam 22.8 NS N8 3N N8 82 QNN 22.8 2858.. £82.82
22.58.. 22226 22 a N 2.2 Nm N2. o E m N No 2. 220.2.
Nmm mNm 2. N e m 2N mm 2. N N N Nm N 22%:
ON N2 2 v N 2 2.252203 228.82 2.9.22 222
N8 22 N2 2. N 2 2 28% 2233 MN
as am 222. 2 N 2 2. 2 2852 32022.50 25233 N
3 N m m N 2852 2.82.282 2.5233 N
22.8 9. mm m 2 m N N N N 2852 2.2222 2.5.2.2: NN
22.22 c o 2 20.22228 2.5.2.2: 2.2.222 2N
22.22 o 2 2 2 20.22250 2.5.2.2: .226 ON
now N2: 22 2 2 2 N 2 N. 85.2 2
22.22 c 2 2.82289 2.5.2.2: 2.9.22 222
§N 2 2m 2 2. w 22 N o 2 a 823283.... 5.232 2
2.2m 2: mm N o 2 82228.2 28 n2
N.Nw 82 N2 2 2 2 222 N N «N N N a. N 822.2822 2832522 50552 2.2
22.22 c N 2 .35 22228-33 2 23% 2.5.2.5 N2
S. o N 2 2 2.222280 88032022 2
s <22 2222.22 2282. 8 Na 3 8 2a N22 N22 3 32 N22 222 2N .2222": 88
32

52.22502 .N2 0322.2.

83

the Northern Hardwood Association (83.7 %), and the Lowland Coniferous Forest
(81.3%) in comparison to NRIS-FSVeg. IFMAP did a very poor job for the Aspen
Association (17.6 %), the Lowland Deciduous forest (25.0%) and Oak Association (45.7

%) in comparison to FIA classification.

IFMAP had confusion with Aspen in classifying the Lowland Deciduous Forest and the
Oak Association in comparison to the FIA classification. Similarly, it was inaccurate for
the Aspen Association (29.4%) and the Oak Association (51.4%) in comparison to NRIS—
FSVeg classification. In classifying Aspen it had confusion with Sugar Maple-Beech-
Yellow Birch and Oak, and in classifying the Oak Association it had confusion with
Quaking Aspen and Bigtooth Aspen classification of NRIS-FSVeg.In this way, IFMAP
did a more accurate job in classifying the Northern Hardwood Association and the
Lowland Coniferous Forest in comparison with both classifications (FIArand NRIS-
FSVeg). Similarly, it was inaccurate for the Aspen Association and the Oak Association

in comparison with both classifications.

The difference matrix of NRIS-FSVeg and FIA revealed that there was excellent
correspondence between most of the forest types between two classifications except the
Hemlock forest type defined by NRIS-FSVeg (T able 11). The Hemlock forest and
Lowland brush of NRIS-FSVeg had indication of confusion with the Aspen forest type of
FIA. However, due to small numbers of sample units in these forest types, these pieces of
evidence are not enough to support the hypothesis of confusion. Similarly, by comparing

the producer’s accuracy, it can be inferred that Red maple (upland), Black ash/American

84

elm/Red maple of FIA classification had lower match with the corresponding NRIS-
FSVeg corresponding classification. Red maple (upland) and Black ash/American
elm/red maple had confusion with Quaking Aspen and Sugar maple/beech/YB. However,
due to small numbers of sampling units in these classifications these hypotheses were not
conclusive. Similar to OI and FIA comparison, the NRIS-FSVeg Red pine, classification
has confusion with Jack pine, and Jack pine classification has confusion with Red pine. In
a case of NRIS-FSVeg classification, there were Jack pine-Oak and Red pine-Oak
classifications which gave more opportunity for matches for Red pine or Jack pine
classified in FIA classification. However, in the OI classification, there were no such
transitional forest classifications so there was marked confusion between both Red pine

and Oak (Table 8).

Three-way Accuracy of FIA/OI/IFMAP Assessment for State Forest Lands

The three-way accuracy of the 239 sample units in the state forest land was assessed by
computing the producer’s accuracy of FIA classification and user’s accuracy of OI and
IFMAP. The overall accuracy in these three classifications was found to be 54.8%, that
is, on average 54.8% of the plots would have comparable classifications in all the three

schemes (Table 13).

The producer’s accuracy of FIA gave information about how well each FIA classification
matched concurrently with OI and IFMAP. The percentage match of the Northern White

Cedar forest type of FIA was highest followed by White/Red oak hickory and Black

85

spruce. Among the poorly matched forest types of FIA were Jack pine, Balsam fir, and

Aspen. The user’s accuracy of OI and IFMAP might be useful to infer about the errors of

commission in these classified forest categories. The Aspen classification of OI had the

least user’s accuracy followed by Jack pine and Oak. The best performing classifications

of OI were Black Spruce, Cedar and Lowland Brush. There was similar evidence from

Table 13. Three way accuracy assessment of FIA, 01 and IFMAP for State Forest

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Lands
(a) Producer's accuracy
' Overall
Code Description Total Match Column Total accuracy %
101 Jack Pine 12 37 32.4
102 Red Pine 9 17 52.9
121 Balsam Fir* 1 3 33.3
125 Black Spruce 17 26 65.4
126 Tamarack“ 1 1 100.0
127 NW cedar 34 39 87.2
503 W/R oak/hickorfi 5 7 71.4
520 Mixed upland hardwood“ 0 1 0.0
Sugar maple/beech/yellow .

801 birch 23 40 57.5
805 Hard maple! basswood* 3 5 60.0
809 Red maple(upland)* 2 4 50.0
901 Aspen 24 59 40.7
Overall accuracy % 131 239 54.8

(b) User's Accuracy 01

User's
Code Description Total Match Column Total accuracy %

A Aspen 24 68 35.3
C Cedar 23 28 82.1
D Treed Bog* 2 3 66.7
E Swamp Hardwds* 0 1 0.0
J Jack pine 12 31 38.7
L Lowland Brush* 4 5 80.0
M Northern Hardwood 28 44 63.6
0 Oak 5 12 41.7
Q Mx Swamp Conifer 16 21 76.2
R Red Pine 9 18 50.0
S Black spruce“ 7 7 100.0
T Tamarack* l 1 100.0
Overall accuracy % 131 239 54.8

 

 

 

 

 

*refers to small number of sample units

86

 

 

 

Table 13. (Cont’d).

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(c) User's accuracLIFMAP
User's
Code Description Total Match Column Total accuracy %
10 Herbaceous Openland 0 15 0.0
Upland Shrub / Low-density
12 trees 0 10 0.0
Northern Hardwood
14 Association 28 40 70.0
15 Oak Association 4 17 23.5
16 Aspen Association 20 35 57.1
19 Pines 21 34 61.8
21 Mixed Upland Conifers* 0 l 0.0
22 Upland Mixed Forest 5 14 35.7
24 Lowland Deciduous Forest* 0 4 0.0
25 Lowland Coniferous Forest 46 51 90.2
28 Lowland Shrub 6 13 46.2
30 Mixed Non-Forest Wetland* 1 5 20.0
Overall accuracy 131 239 54.8

 

 

 

 

 

* refers to small number of sample units

the OI-FIA difference matrix that, Aspen had the third lowest user’s accuracy (Table 8).

However, the three-way accuracy assessment was dependent on all three sets of

However, the three-way accuracy assessment was dependent on all three sets of

classifications.

From IFMAP user’s accuracy assessment, Oak had the least accuracy of 23.5%. From

this evidence it could be inferred that when IFMAP classified the Oak Association, then

there would be less than 25% likelihood that the two other classifications also classify the

given sample unit as Oak Association. IFMAP had the highest user’s accuracy for

Lowland Conifer followed by the Northern Hardwood Association. These conclusions

from the three-way comparison were very similar to the previous IFMAP-FIA

comparison (Table 7).

87

 

Three-way Accuracy of FIA/NRIS-FSVeg/IFMAP Assessment for National

Forest Lands

The three-way accuracy assessment of 523 sample units in the national forest land was
carried out by computing the producer’s accuracy of FIA classification and user’s
accuracy of IFMAP and NRIS—FSVeg classifications. The overall agreement among these
three classifications was found to be 58.5%, which was slightly higher than the similar

comparison in state forest land (Table 14).

Producer’s accuracy of FIA in this classification was useful to provide information about
how well the other classification matched with FIA plots and the user’s accuracy of two
classified maps or databases were useful to provide information about how well these
classifications correctly classified the corresponding FIA information. The producer’s
accuracy of FIA classification was excellent for Hard maple/basswood (91.7%) and good
for Northern red oak (80%). The producer’s accuracy was low for Black ash/American
elm/Red maple (18.2%), Red maple/upland (25%), and Aspen (26.3%). In these
categories the NRIS-FSVeg and IFMAP were unable to correctly classify these FIA

forest types in their corresponding classified map category.

The user’s accuracy of IFMAP was excellent for Lowland Conifer Forest (81.3%) and the
Northern Hardwood Association (79.8%). However, the user’s accuracy of IFMAP was
very low for Other Upland Conifers (0.0%), the Aspen Association (15.7%), and

Lowland Deciduous Forest (25%). The IFMAP classifications were also not so promising

88

Table 14. Three-way accuracy comparison of FIA, NRIS-FSVeg and

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

IFMAP in National Forest Lands
(a) Producer's accuracy of FIA classification
Total Column Producer’s
Code Description Match Total, Accuracy %
101 Jack pine 32 62 51.6
102 Red pine 63 98 64.3
103 Eastern white pine 0 l 0.0
105 Eastern hemlock 1 1 100.0
121 Balsam fir 1 0.0
122 White spruce l 0.0
125 Black spruce 24 31 77.4
127 Northern white-cedar 30 40 75.0
400 Oak Pine Group 0 l 0.0
401 White pine / Red oak/ white ash 2 0.0
503 White oak / red oak / hickory 10 13 76.9
504 White oak 3 8 37.5
505 Northern red oak 4 5 80.0
701 Black ash / American elm I red maple 2 11 18.2
800 Maple / Beech / Birch Group 3 6 50.0
801 Sugar maple / beech / yellow birch 101 146 69.2
805 Hard maple / basswood ll 12 91.7
809 Red maple / upland 2 8 25 .0
901 Aspen 20 76 26.3
Overall accuracy 306 523 58.5%
( b) User's accuracy IFMAP
Total User’s
code Description Match Row Total Accuracy %
10 Herbaceous Openland 0 9 0.0
12 Upland Shrub / Low-densitLtrees 0 2 0.0
14 Northern Hardwood Association 103 129 79.8
15 Oak Association 16 35 45.7
16 Aspen Association 8 51 15.7
18 Mixed Upland Deciduous O 1 0.0
19 Pines 95 133 71.4
20 Other Upland Conifers O 10 0.0
21 Mixed Upland Conifers 0 6 0.0
22 Upland Mixed Forest 28 58 48.3
24 Lowland Deciduous Forest 2 8 25.0
25 Lowland Coniferous Forest 39 48 81.3
28 Lowland Shrub 9 17 52.9
30 Mixed Non-Forest Wetland 6 16 37.5
Overall accurgy 306 523 58.5%

 

 

 

 

 

89

 

 

Table 14 (Cont’d)

(c) User's accuracy of NRIS-FSVeg

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Total Column User’s
Code Description Match Total Accuracy %

1 Jack Pine 27 51 52.9
2 Red Pine 59 86 68.6
3 White Pine 2 0.0
Hemlock 4 4 100.0

11 Balsam Fir-Asp-PB 9 0.0
12 Black Spruce ll 13 84.6
14 Northern Wh Cedar 19 19 100.0
18 Mix Swamp conifer 17 23 73.9
19 Cedar-Aspgn-PB 2 2 100.0
21 Mixed Northern Hdw 0 l 0.0
48 Jack Pine-Oak S 10 50.0
49 Red-Pine-Oak 4 4 100.0
53 Black Oak 1 5 20.0
55 Northern Red Oak 3 4 75.0
59 Mixed Oak 13 25 52.0
71 BI Ash—Elm-R Maple 2 4 50.0
81 S Maple-Beech -YB 47 67 70.1
82 S Maple-Basswood 6 7 85.7
84 Red Maple(Dry) 0 3 0.0
85 Sugar Maple 26 34 76.5
87 Sugar maple-beech-yellow birch/red spruce 4 6 66.7
89 Mixed Upland Hdwd 31 47 66.0
91 Quakig Aspen 12 57 21.1
93 Bigtooth Aspen 3 14 21.4
95 Asp-W Spruce BF 5 9 55.6
97 Lowland Brush 5 9 55.6
99 Open 8 0.0
Overall accuracy 306 523 58.5%

 

 

for Mixed Non-forest Wetland (37.5 %), Oak Association (45.7%) and Upland Mixed
Forest (48.3%). These conclusions were very consistent with the results of IFMAP—FIA

comparison in the national forest land (Table 10).

The user’s accuracy of NRIS-FSVeg was found excellent for Northern White Cedar,
Cedar-Aspen-Paper birch and Hemlock. The user’s accuracy of NRIS-FSVeg forest types

was poor ranging from 20 to 22% for Black oak, Quaking aspen, and Bi gtooth aspen.

90

Similarly, the NRIS-FSVeg performance for Jack pine, Jack pine-Oak, and Mixed Oak
was not so promising. However, these results were not similar to the NRIS-FSVeg-FIA
comparison except for the Black oak (Table 11). The lower overall accuracy and lower
user’s accuracy in the above mentioned classifications of NRIS-FSVeg were mostly due

to mismatch with the similar forest types of IFMAP classification.

Comparison of Classification Agreement Between IFMAP and FIA for State

Forest Land

The well accepted method for comparing two classifications is to formulate an error
matrix or confusion matrix and to calculate of a Kappa statistic, a coefficient of
agreement between two classifications in comparison to random classification. As
already noted, there were marked differences between the classification schemes of these
four forest classification systems. When there were multiple categories of classified
classifications which were acceptable or probably right to one category of reference
classification, then formulation of a square error matrix with a matching category
diagonal became difficult. The choices for formulating the error matrix were examined:
(1) keeping the user’s accuracy for each classified map constant, (2) keeping the
producer’s accuracy constant for each reference map, and (3) aggregating classified forest
type categories with multiple matches in the reference classifications. In this study, these
three approaches were adopted to assess effects of accuracy of IFMAP in comparison
with FIA forest classification for both state forest land (in this section) and national forest

land (in the next section). Overall, this provides a partial assessment of the utility of

91

IFMAP information. After formulation of error matrices using the three approaches, the
first two approaches yielded the same overall accuracy as the difference matrices. For the
last approach, aggregating the classified map categories, the overall accuracy will
increase, but details of forest classification relationships will be lost. For many users, a

loss of detailed information may be unacceptable.

In the UA constant approach, the “acceptable” or “probably right” classified map
categories were moved left or right in the error matrix to the appropriate diagonal cell
(Table 16). As a result the user’s accuracy remained unchanged. However, the producer’s
accuracies (column) may be modified. Similarly, in the PA constant approach, in the
“acceptable” or “probably right” map categories were moved up in the error matrix to the
appropriate diagonal cell (Table 17). The user’s accuracy of the “acceptable” matches or
primary classification of IFMAP may be modified, but the producer’s accuracy does not
change. For example, in the UA constant approach, the PA for Upland Mixed Forest,
Lowland Shrub, and Mixed Non-forest Wetland was 100 % (Table 15), which was not
true in the original difference matrix (Table 7). And for the PA constant approach, the
UA for the Pines, Lowland conifer, Northern Hardwood increased considerably (Tables 7
and 16); this was not the casein the original difference matrix. From the user’s
perspective, the approach of UA constant was found relatively fair for the “acceptable”
match or primary classification types; however, this approach overestimated the
producer’s accuracy of the “probably right” or secondary classification types. In the UA

constant approach, the classified map label was not changed as movements in the error

92

2822222822.. 2.2.282 02 2.220222 <3 9222 52222203. 22.282285 02 20.22 <22 2802

 

82280232 32.223225

225222228 <3 22.5212 2882 025m 82 2.2.5-522 2c 2.5222 282m .2 0222.22.

93

282222203. 22.283 222 2.2.2028 <3 22:22 322228222 22.282285 02 £22.22 <22 202222

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

8.8 22225822 22820 38.22 .5252
322 22.22 22.2w 2.2.2. 2..2N 2.8 2.2. <2
%N 8 2 2.m 2% N N2. .32;
22.22 N 22 2 2 2252203 2852.872 2.222: on
22.22 22 22 2 2. 2 222.2% 225233 N
2222 222 2.2m 2 28822 282022228 225233 N
22.22 2. 2 N 2 28522 2.8222822 222222233 2.N
22.22 N 2 o 22 222222 222222 2222222222 NN
22.22 2 22 2 2522228 2252222: 2292222 2N
2222 mm 2. N N 2 222222 N2
8 222. 2 2 2 N N 22 82222822 .2223 2
N.NN 222 22 m N 8222228322 25 m2
N.NN 22. N m 2.2.. 822.228.22.22 2.2
25032232.— 222052072
22.22 222 22 2. c 822 2222522 N2
-33 2 222:5 225222222
22.22 2 2 N2 N 22.22.22.220 m222282.22225 222
.22. <22 2.28.2. 22m N N 2.N NN 2N N2 2 2 2.2 N2 222 222.22 080
22222 3228832 222222 222.2685

22229228 S22 8223 282cm 33 82 2.222-222.2222 222 22.5.22 28m .22 222222.

94

matrix were made only along the row and consequently the user’s accuracy of the
primary classification were remained the same. For the user’s accuracy, the Northern
Hardwood (82.5) remained the same and the producer’s accuracy of Upland Mixed,
Lowland Coniferous and Mixed non-forest Wetland forests were 100% (Tables 7 and

15).

In the aggregated approach the 12X12 matrix was reduced to 6X6, merging the six
IFMAP categories to two new forest types, namely Pines, Conifers and Lowland Low-
density Forest and Upland hardwood. The first new types were formed merging Pines,
Lowland Conifers and Lowland Shrub and Mixed Non-forested Wetland groups and
second by merging Aspen, Oak, and Northern Hardwood associations and Mixed Upland
Forest (Table 17).

Table 17. Error Matrix of IFMAP-FIA in State Forest Lands (aggregated)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Classified Map Reference map (FIA)
Code IFMAP ‘ ,. ,“A’ 3,21 24 1012 Total UA%
Pines, Conifers, and Low density
A Lowlands 94 9 103 91.3
B Upland Hardwoods 11 95 106 89.6
21 Mixed Upland Conifers 0 l 1
24 Lowland Deciduous Forest 1 3 4
10 Herbaceous Openland l3 2 15
12 Upland Shrub / Low-density tree 4 6 10
Total 123 116 239
PA % 76.4 81.9
KHAT 0.628 Over all accuracy 79.1 %

 

 

 

 

Note: PA, Producer's accuracy; UA, User's accuracy; and new forest types A: Pines(19)+
Lowland coniferous forest (25)+ Lowland Shrub(28) + Mixed Non-forest Wetland (30)

and B = Northern Hardwood (14)+ Oak (15)+ Aspen (l6) associations and + Upland Mixed
Forest (22). '

By merging IFMAP classification the information on these classifications was lost. A

number of sample units which were misclassified by IFMAP were now in as matched due

95

to new classification. There were no sample units in four categories in so the cells in
these categories are empty. The overall accuracy was 79.1% for the aggregated
classification for the state forest land. The user’s accuracy of the poorly classified forest
types went up and well classified forest types of IFMAP went lower. For example the
user’s accuracy of Oak Association, a component of Upland Hardwood Forest, went up
from 23.5% to 89.6% and Lowland Conifer Forest, a component of Pines, Conifers and

Lowlands Low-density Forest, user’s accuracy was lowered from 98.0% to 91.3%.

When interpreting the KHAT, the error matrix can be compared to how much better the
classification was than the random allocation. The value of KHAT can range from -1 to
+1 (Congalton and Green, 1999). To quantify the strength of agreement for the
comparison of categorical data, Landis and Koch (1977) described the ranges for KHAT
values into several groupings: a value greater than 0.80 (i.e., 80%) almost perfect; a value
between 0.61 and 0.80 substantial; a value between 0.41 and 0.60 moderate; and less than
0.40 represents fair and slight agreement. Based on this classification of strength of
agreement, there was a moderate level agreement between FIA and IFMAP classification
for both UA constant and PA constant approaches (Table 18). However, for the
aggregated forest type approach there was substantial agreement between FIA and

IFMAP classification.

Table 18. Individual Error Matrix Kappa Analysis Results for State Forest Land

 

Classification Comparison N KHAT Var (Khat) Z statistic
IFMAP-FIA UA Constant 239 0.5683 0.001250 16.05882
IFMAP-FIA PA Constant 239 0.5695 0.001251 15.65486

IFMAP-FIA Aggregated Forest Types 239 0.6284 0.001772 14.92579
Note: UA refers to User’s accuracy and PA refers to Producer’s accuracy.

 

96

The comparison of the error matrix using the Z-test confirmed that all the three
approaches of classification were significantly different than the random classification.
Similarly, the aggregated forest types approach had highest KHAT value; it inferred that
this approach of classification was 63% similar to FIA classification. The UA fixed and
PA fixed approaches had almost the same KHAT value. The third approach raised the
KHAT value due to aggregation of the forest categories; this approach made some of the
misclassified sample units in the difference matrix as matched after redefining new
categories. The limitation of this approach was that neither the user nor the producer had
much information on the components of the merged forest types. The Z statistic values
for all classifications are greater than 14 and it confirms that these classifications are

better than the random classification.

Kappa analysis can compare two error matrices at a time to determine if they are
significantly different. This test was based on the standard normal deviate and the fact
that although remotely sensed data are discrete, the KHAT statistic is asymptotically
normally distributed (Congalton and Green 1999). At the 95% confidence level, the
critical value would be 1.96. Therefore, the results of the pairwise test of three matrices
revealed that the comparisons were not significantly different, as the KHAT values of the
UA constant and PA constant approaches were very close to the pairwise comparison Z
statistics was found small for these two pairs (Table 19). Hence, aggregation improved
overall accuracy, but not in a statistically significant sense. However, Z statistic for the
comparison between aggregated and PA fixed were large but not enough to make them

significantly different at 95% confidence interval. The Z values results were consistent

97

Table 19. Kappa Analysis Results for the Pairwise Comparison of the Error Matrices

 

Z values
Approach UA constant PA constant Aggregated
UA constant 0.280893 1.102931
PA constant 0.280893 1.358207
Aggregated 1.102931 1.358207

 

Note: UA refers to User’s accuracy and PA refers to Producer’s accuracy.
with pairwise Kappa test (Table 19). The strongest agreement was observed between the

PA constant and UA constant approach.

Comparison of Classification Agreement between IFMAP and FIA for

National Forest Land

A process identical to one used for state forest lands was employed to analyze national
forest lands. Changes in UA and PA reflect cases for which there were multiple IFMAP
matches to the FIA forest type (Table 10). For example, the user’s accuracy of the Oak
Association, Pines and Lowland Deciduous Forest remained the same (47.5%, 82.7%,
and 25%, respectively) in both approaches (Tables 20 and 21) and producer’s accuracy
for the Pines and Lowland Deciduous Forest remained the same (68.3% and 18.2%,
respectively). Similar to the state forest land, the aggregation approach increased the
overall accuracy of IFMAP in national forest land from 64.8% to 73.6%. By aggregating
the IFMAP classification the user’s accuracy of Hardwood Forest (new aggregated
classification) becomes of 82.8 %, lower than its Northern Hardwood Association
component and higher than its Aspen and Upland Mixed Forest components. In this way,
the user’s accuracy of well matched classified map category may be reduced and the

poorly defined classified map category may increase.

98

 

2322222822 328: J5 2222222222822 {822220222 6.22 28222

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

c2.22.222 22258.2 22.2226 $2222 $252

22222 22222 oNN N222 22222 22.22 22.22 322.2 2.2.2 n22 N.2N 92222

22.2.22 NNW 22 22 2.m 22 222 N N 2222 8 N 82 282;
new 222 a 22 22 22 2 2. N 2252203 222222.252 222222 2222
N.NN N2 22 2.2 22 22 2 m N 222.22% 2252.33 N
2.32 222. 22 22 N22 22 22 m N 2852 28322822 225233 N
22.2 22 22 22 22 , N 2 m 222222288222822 225233 2.N
N.2m 22m 22 22 2 2. 2222. N N 22222.2 2222222 225222222 NN
22.22 c 22 22 22 22 m 2 22022222222 225222222 82222 N
22.22 222 22 22 22 22 22 2 2 22022.28 225222222 222222 8
N.NN m: 22 22 22 2 2 2222 22 2 2. 2.22222 222
22.22 2 22 22 22 22 2 22522222822 225222222 82222 222
2.: 2m 22 22 N 22 N 2 a 2. mm 2822222822 .2222 222
N22. mm 22 22 22 22 m 2 e2 2 2522222822 22220 2
22.222 2N2 22 22 22 2. 22 2.2 222 25222822 2.2

2200322222: 222052222
22.22 N 22 22 22 22 2 2 2822 N2

22288-33 2 22825 225222222
22.22 a 22 22 22 22 m 2 2 225222220 382222522 222

22.225 2922a, 22m N N 2.N NN 2N ON 82 222 22 2 2.2 N2 222 . , 222222; 0822
NE 22222223532220 oozeohom 2222 222.2885

22522228 22222 22222.22 28.52 2222522222 :2 222222222222 222 2222222222 222cm .8 0222222.

99

222222222822 {cab <3 2222“ 22822222822 22.2022222222222 <n2 2802

 

couaoEmmflo 8228832 222.2320

22522.28 «.222 22252.2 2382 222822222 .222 222.222.22.22 222 2225.22 282m .2N 222$

 

.80 2222222203 2moSN2éoZ 220222: + 88 2222225 222222222512
£va 282cm 222520222200 22222223222 u R: 235m 2222222253 52222203 33 2252 822202222220 222222223012 22:22 ANNV 2mo20n2 2202222 222222.223 + $22 =o22222oomm<
:o%< + 3.22 =o22u2uomw< 250252222222 E05202 n 28202 2200322322 022.22 2320a 3072 22222.. 2202222822 @233 <3 22822222822 22.282.252.22 <2 2082

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

2.2.2.2222 2222822282222 8:20.232

2223822232 8.23 222222 2222522222 82 2222222222 2225 52.22.2222": .NN 0223

22220.2. 20222258 22222920 2N2o.o .2.<22v2

new N.w2 36 W20 2.? 2» <2

22. 22 N m 202 0N 3N 2228.2.
2.: 2w , N w 2.2 222222222252212 3

2 222292022 250.2 20:22 302822.200 222222233
QWN w 2. N . 2 m unouonm m220=EOOQ “22222233 VN
22.22 2.2 n 2 £022.50 222222. 3 228222 2N
22.22 222 .2 .2... 22 N 22022228 225.2 22 see 8
N.Nw mm2 w 2 2 a 2 N2 822222 2
od 2 2 c 302222283 2222222223 222222 222
2.622 mm m v.2 . 02 22022222083... 22220 m 2
wNm NMN m m N NN o 3.2 a 2822i 2202225222223 0
06 N 2 2 8022 222522.302 2 2222225 2222222 3 N2
06 a w m 2 2222222: 0 95822222022 02

.22. 2.22 2.22.2.2. 22 . 2.N .2N... 22|N . .. . a. .. 22,2. ... 1m . 0 N2 . g .222 . , , , .22., I. , 02.8
22222 82.222220

101

The KHAT value was 0.569 for the UA constant approach and 0.554 for the PA constant
approach (Table 23). For both the PA constant and UA constant approaches, the KHAT
value was at a moderate level. The KHAT value of the error matrix for aggregated
categories was 0.612, a substantial level. The aggregated IFMAP classification was

61.2% better than the random classification.

Table 23. Individual Error Matrix Kappa Analysis Results for National Forest Land

N KHAT Var (Khat) Z statistics
IFMAP-FIA UA Constant 523 0.568698 0.000602 23.17888
IFMAP-FIA PA Constant 523 0.554815 0.000607 22.51480
IFMAP-FIA Aggregated 523 0.612099 0.000736 22.56409

 

 

The comparison of the error matrix with Z-test confirmed that all the three approach of

classification were significantly different than the random classification (Table 24).

Table 24. Kappa Analysis Results for the Pairwise Comparison of the Error Matrices

 

Z value
Approach UA constant PA Constant Aggregated
UA constant 0.399238 1.186575
PA Constant 0.399238 1.563060
Aggregated 1.186575 1.563060

 

The results of the pairwise test of three matrices for significance at 95% confidence
interval revealed that the comparisons were not significantly different. The KHAT value
of the UA constant with PA constant approach was close to the pairwise comparison Z
statistics. However, the Z statistic for the comparison between aggregated and PA
constant were higher, but not enough to make them significantly different at the 95%

confidence level.

102

A previous study carried out by Space Imaging (2004) found the overall accuracy of
IFMAP as 67.9% and KHAT value 0.60 for state forest land (N =789). This study found
the overall accuracy of IFMAP as 63.6% in State forest land‘and 64.8% in national forest
land. Similarly, the KHAT values were almost 0.56 for both state forest and national
forest lands. In both studies the FIA sample plots were used as the reference
classification. The previous study also had additional information of “canopy call”, a
measure of crown cover percentage use. But in this study, robust crosswalk tables were
used with some flexibility for including to the low density crown cover forest types of

IFMAP.

Due to lack of additional information on the crosswalk table, it is difficult to critically
compare the results across studies. For example the user’s accuracy of the Oak
Association and the Aspen Association in the previous study was 35.6% and 44.9%, and
from this study Error Matrix keeping the UA constant user’s accuracy for these forest
types were 23.5% and 60% in state forest land and 45.7% and 17.6% for national forest
land respectively. Both studies conclude that IFMAP had difficulty in classifying the Oak
and Aspen associations. Similarly, the overall accuracy of Oak and Aspen associations
from the three-way accuracy assessment user’s accuracy of IFMAP were 23.5% and
57.1% for state forest land and 45.7% and 15.7% for national forest land, respectively. In
this way, IFMAP had difficulty in classifying the Oak Association and Aspen
Association. The unavailability of the crosswalk table of previous study between the FIA

classification and IFMAP classification precludes more detailed comparisons. Findings

103

from the previous study and this study confirm that the overall KHAT values of IFMAP

are in the range of 55% to 60%.

Explanation of Differences in Classifications of IFMAP, 01 with FIA

The final study objective focuses on explaining the differences in classification for the
state forest lands. Analyses are limited due to the unavailability of plot and stand level
characteristics. However, efforts were made to explain the difference in classification
based the interpretation of the difference matrix and the personal communications with
the FIA unit at NCRS. For a number of forest type categories, the reliability of the
IFMAP was very poor (e.g. Oak Association, Aspen Association, Upland Mixed Forest,
Mixed Non-forest Wetland, and Lowland Shrub). The findings of this study and the
previous study were similar for these forest types. Thus, the satellite imagery
classification techniques used for IFMAP classification needs to be further improved to

enhance the producer’s accuracy of these forest classifications.

The other reason for lower overall accuracy may be due to incompatibility between FIA
and IFMAP classifications. To explain the differences in classification between IFMAP
and FIA in state forest land, a number of questions were asked to FIA unit at NCRS. The

questions clarified the relationships between the FIA plots and IFMAP classes.

The difference matrix of IFMAP and FIA (Table 7) was used to explain the differences in

IFMAP with FIA classifications. There was noticeable misclassification error for Jack

104

pine, Black spruce and Aspen of FIA classification into appropriate IFMAP
classifications. Aspen Associations of IFMAP wrongly classified inappropriate FIA

forest types into this classification.

In the 10 sample locations, Jack pine (FIA) was classified into Herbaceous Openland and
in four locations into Upland Shrub/Low-density tree (IFMAP) (Table 7). The
observation of FIA plot data for additional details about the age or stocking of trees
indicated “all but one of 14 plots is in the small diameter size class. Ages are generally
young (5-15 years), but a few were 30 plus. Algorithm typed those stands a variety of
types including Northern red oak, other pine/hardwood, post oak/black jack oak, aspen,
black cherry” (Holden pers. comm.). Thus, IFMAP and FIA classification approaches
were non-compatible to one another for forest classifications of the non-forest types, less
than 25% trees canopy cover, defined of IFMAP. The information of 13 plots in small
diameter indicates that most of the trees in this location were smaller in size and younger
in age and heterogeneous in vegetation. This explanation apparently seems to corroborate
the IFMAP classification but not sufficiently confirm that in these locations the tree
canopy covers were less than 25%. Thus the differences in defining forest areas and type

classifications may have lowered the accuracy for low canopy cover classes of IFMAP.

Four misclassified plots of Jack pine (FIA) were called Upland Mixed Forest (IFMAP).
Close examination of FIA plot data and FIA algorithm forest type classes indicated “only
one of four plots was typed as ‘other pine/hardwood’ (409) by the algorithm” (Holden

pers. comm.). So, IFMAP clearly misclassified these plots in three of four cases.

105

In other instances of misclassification, Black Spruce (FIA) was classified into Lowland
Shrub or Mixed Non-forest Wetland (IFMAP) in eight sample units. The FIA plot-level
information gave an impression that these stands “are between 25 and 70 years old, four
of which are over 50 years. Half of stands are in the small diameter size class. The other
half is distributed between large and medium. Most stands are medium or poorly
stocked” (Holden pers. com.). This additional information supported that IFMAP was

right in 50% of these sample units.

In another example of misclassification of Aspen (FIA) into Oak and Northern Hardwood
associations (IFMAP) in 16 sample units, out of total 59 Aspen forest types classified by
FIA in this study (Table 7), the FIA plot-level information indicated that “[t]he algorithm
typed almost all of these stands as aspen. Two were typed as Oak forest types and one as
other pine/hardwood” (Holden pers. com.) This information only supported IFMAP
classification of Oak at two stands of out of total 11 stands; therefore IFMAP

misclassified 9 sample stands.

The other misclassification of Aspen forest (FIA) into Herbaceous openland or Upland
shrub/low-density (IFMAP) in eight sample units, the FIA plot-level detail information
pointed out “four of eight stands are less than 10 years old. A couple of the older stands
are poorly stocked. One stand was typed as ‘white oak/red oak/hickory by algorithm, also
a pretty young stand” (Holden, pers. comm). In this case, IFMAP was right in half of the
stands, however the FIA does not have compatible forest classifications with lower

canopy cover or sparsely vegetated forest types. This non-compatibility between the

106

reference and classified map in lower canopy cover of sparsely vegetated categories may

have lowered the overall accuracy of IFMAP.

In case of Aspen Associations (IFMAP) misclassified as various oaks or Northern
Hardwoods (FIA), the further query of FIA plot-level information implied that “[m]ost of
these stands are 50 plus years old and all fall into the large to medium diameter size class.
Two of 12 stands are typed as Aspen by the algorithm” (Holden pers. comm.). In this

way, out of 12 stands, IFMAP was wrong in 10 stands.

Similar efforts were made to explain the misclassification between the 01 and IFMAP
(Table 8). In case of misclassification of Jack pine (FIA) into Red pine (OI) in three
sample units (Table 8), the FIA plot level information hinted “[o]ne of three plots was
half red pine” (Holden pers. comm.). In this way, 01 was only right in one of three
misclassifications of Jack pine. Similarly in case of misclassification of Jack pine (FIA)
into Oak (OI) in four sample stands, the FIA plot-level information supported “Two of
four plots had alternative forest type conditions from algorithm. One was other
pine/hardwood and the other was oak/ black jack oak” (Holden pers. comm). In this way
01 was completely right in one of four misclassifications and partially right on one
additional sample stands. Out of four misclassified Jack pine stands by 01, OI was wrong

for two stands.

In addition to misclassification of Jack pine (FIA) into Aspen (01) in three stands, the

FIA plot-level information provided additional details “[o]ne of three plots typed as other

107

pine/hardwood by algorithm. Some aspen on each plot (Holden pers. comm)” In this
way, the FIA algorithm typed another forest type supported the classification of OI in one

plot, and the occurrence of aspen partially supported 01 classification.

Misclassification of Red pine (FIA) into Jack pine (01) occurred. The FIA plot-level in
three sample units data indicated “[o]ne of three plots typed as jack pine by algorithm and
most of plots are composed of jack pine” (Holden pers. comm.). In this case, the 01 was
completely right in one sample units. This example provided the errors of
misclassification of reference forest types by FIA field crew due to the mixed vegetation

of Red pine and Jack pine.

In case of misclassification of Black spruce (FIA) into Treed bog and Lowland brush (OI)
in eight sample stands, the FIA plot-level data hinted “[s]tand ages on eight plots range
from 25-80 and stand size classes (from algorithm) are large to medium and small
diameter. Field crew tended to call plots small to medium. Growing stock codes were
poor to medium stocking. One was considered overstocked” (Holden, pers. comm.) In
this way, these sample stands which were poorly stocked or small diameter, the 01
classification may be right. However, the non-compatibility of OI and FIA in forest

classification and definition of forest limited the accuracy assessment.
In five sample units, the Sugar maple/beech/yellow birch (FIA) was classified as Aspen

by 01. The request for further information from FIA in this misclassification indicated

“[s]tand ages on five plots range from about 45-70. Stand size class mostly large

108

diameter. Few aspen on any of the plots” (Holden, pers. comm.) In this way, this
explanation revealed that in these plots there were mixes of both sugar maple and aspen,
and the plot-level heterogeneity has made it difficult to assign the reference forest
classification. As discussed earlier, the relative size of FIA plots and 01 stands are

different. The relative size of patch size may have biased the two classifications.

The non—compatibility of reference and classified map classification in lower canopy or
stocking class of 01 and IFMAP, the relative patch size, and the heterogeneity in the
sample stands creates difficulty in correctly classifying the reference classification. In
addition, the IFMAP is relatively more inaccurate in classifying the Aspen, Oak
associations, Upland Mixed Forest and Lowland shrub; this may have aggravated the
differences in classification of IFMAP and FIA. These issues may have consequently

lowered the consistency of three-way forest classification.

109

CHAPTER 5

SUMMARY AND CONCLUSION

In Michigan, there are multiple sources of forest management information, developed for
specific purposes. There are differences in defining forest, forest types, objectives of their
development and details in defining the non-timbered forestland classes, application, and
anticipated level of accuracy among these databases and map (Table 6). OI and NRIS-
FSVeg are more similar to each other than the other two in terms of their objective,
nature of inventory and application. IFMAP is the only database which can provide the
“wall-to— wall” information in Michigan. FIA has sample plots throughout Michigan,
which are used to statistically assess the status and trends of forests in Michigan.
Application of FIA data for sub-state areas produces results with lower levels of
accuracy. However, they provide very accurate point-level data for the reference data for
this study. IFMAP uses a number of low density cover forest types based on a Level H
classification scheme, and aggregates forest classes relative to the finer detail in FIA, 01,

and NRIS-FSVeg.

The results of the consistency among FIA, NRIS-FSVeg and IFMAP were calculated by
computing the three-way overall accuracy. Based on the sampling information of 523
sample points on national forest lands at the exact FIA plot locations, the overall
accuracy was found to be 58.5%. Among the forest classification with a large number of
samples, the Lowland Coniferous Forest and Northern Hardwood Association of IFMAP

did well in user’s accuracy or reliability and the Aspen Association, Oak Association and

110

Upland Forest did poorly in terms of user’s accuracy. These findings were similar to

those found in a study conducted by Space Imaging in 2004.

The findings of the consistency among FIA, 01 and IFMAP on state forest land revealed
that the overall accuracy of three-way classification was 54.8% (N =239). The three-way
matching of Lowland Conifer and Northern Hardwood were promising on state forest
land. The study indicates the performance in the Oak Association and Aspen Association
were not very accurate. However, the strength of this conclusion is limited due to small

number of sample plots in these categories.

The objective of comparison of the IFMAP and FIA classifications for national and state
forests was limited due to the incompatibility of the two classifications. A crosswalk table
was prepared, and based on that the overall accuracy were calculated. The overall
accuracy of IFMAP in comparison to FIA on national forest land was 64.8% and on state
forest was 63.6%. To compute the KHAT values and assessment, “acceptable” and
“probably right” matches had to be re-assigned to other classes. First, the user’s accuracy
of IFMAP classification was kept constant while developing the error matrix. In this
process all of the secondary FIA forest type matches were shifted to the diagonal of the
error matrix. Second, the producer’s accuracy was kept constant; in this approach, the
secondary matches were added to the FIA primary classification to keep the PA constant.
Third, all of the IFMAP classifications with matches in multiple categories were merged
to new categories. After formulating the error matrix the KHAT values were computed,

the KHAT values from the first two approaches were very close, 0.55 and 0.56. For the

111

third approach, using aggregated IFMAP categories, the KHAT value was 0.612 for
national forest land and 0.628 for state forest land. The third approach resulted in slightly
higher KHAT values due to aggregation of categories. The problem with the third
approach is that it gave larger overall accuracy and larger KHAT values at a cost.
Specifically, the users will not be aware of the user’s accuracy in the category of IFMAP

for which information was aggregated.

The major reason for the difference in the classification was due to poor performance of
IFMAP in classification of two important forest types, the Oak Association and the
Aspen Association. The other’reason for the lower overall accuracy and KHAT value was
due to the inconsistency in labeling the reference data due to land cover heterogeneity
surrounding the sample location. IFMAP has Level 11 classification for the low canopy
cover land use classification, and FIA has a finer detailed forest classification for land
with lower canopy densities. For example, regenerating jack pine is Jack pine in FIA

plots, but non-forest in the IFMAP classification.

Conclusions

With the creation of IFMAP land cover information, many planners, resources managers
and researchers will likely use it for landscape-level studies because it is the “best
information available”. This study highlights some of the shortcomings of IFMAP, OI
and NRIS-FSVeg. These shortcomings should be caveats on the landscape-level studies.

The standard of accuracy set forth by the USGS for the generalized first and second

112

levels thematic map is 85% to 90%, IFMAP needs to improve its producer’s accuracy to

meet this requirement. For some forest types, it meets this standard.

Limitations

The findings of this study are based on the number of small sample plots from the NCRS
FIA unit. As the rule of thumb, the number of sample size required for a map covering
large area or the classification with large number of forest cover types (i.e. more than 12
categories), the minimum number of samples should be increased to 75 or 100 samples
per category (Congalton and Green, 1999). The producer’s accuracy, user’s accuracy
derived for the map category with lower number of samples may or may not be

representative of the public forest lands of Michigan.

The findings of this study were based on the FIA field crew defined forest types. The
heterogeneity in the sample area may have affected the reliability of the reference map
classifications. In this study there was some evidence that supports the heterogeneity in
the sample units and inconsistency in defining the reference forest classification (FIA
field crew and forest typed by algorithm). FIA reporting of forest resources 1993 as well
as 2003 clearly mentions the heterogeneity in FIA sample plots. For example the growing
stock of Jack pine forest types only contains about 61% Jack pine, 11% Red pine, 10%
other softwoods and other species growing stocks (Table 47 in Leatherberry and Spencer

1996, Table 6 in Leatherberry et al.2005).

113

In this study the patch size, a smallest unit for which the forest classifications were
compared, are different. For FIA patch size is the four subplots; for IFMAP it is a 30m X
30 m pixel; and for the 01 and NRIS-FSVeg it is the stand. Previous research found that
as the patch size decreases, the accuracy decreases (Smith et al. 2002). Small patches

may have affected the accuracy results in this study.

This study was only confined within the public forest lands of Michigan. The accuracy of
the IFMAP may be similar for public and private forest lands, but this was not examined

due to the public lands focus of this project.

This study was based on the current vegetation types reported on the forest management
databases (FIA, OI, NRIS-FSVeg) and IFMAP land-cover 2001. Vegetation classes may
change over time, and data from different times may lead to misclassifications. Data in

this study were ca. 2000, but from different years.

As there are differences in defining forest and forest types among databases, this may
have generated bias on accuracy estimation for the forest classifications with lower

canopy cover and between the databases with differences in definition of forest.

In this study, the crosswalk between the reference and IFMAP classifications allowed
multiple matches to the “acceptable” and “probably right” classification of IFMAP. This
assumption may have over-estimated the accuracy for these forest types which were

allowed for multiple matches by the crosswalk tables.

114

Policy Implications

Results of this study have several policy implications. Some of the important policy

implications are noted below.

For state forest lands, OI provides more accurate information than IFMAP land cover,
and 01 information should be used by policy makers and planners when possible.
Similarly, for national forest lands, the NRIS-FSVeg information is more accurate than
IFMAP, so priority should be given to using this information. However, for all forest
lands, public and private, the only available information is either IFMAP or FIA

information.

There are undoubtedly many detailed implications of the reliability of forest management
data. A few examples provides ideas regarding some concerns that may arise. User’s
accuracy-Jack pine and Red pine account for 80% user’s accuracy for IFMAP’s Pine
category (Table 7). By using the IFMAP land cover map for Pine category, users will
have non-pine types 20% of the time. If these point data were expanded to the landscape,
say 500,000 acres, the actual Pine forest area may be 100,000 acres less. Even this high
reliability of Pine could yield significant challenges in the field ranging from issues of

short-term timber/habitat availability to long-term sustainability.

Producer’s accuracy for Jack Pine is fairly low, 48.6% (Table 7). Over 25% of the Jack

Pine plots are classified as Herbaceous Openland by IFMAP. In most cases, these are

115

regenerating stands, but this is not evident from the simple comparison of forest cover;

hence, producers must communicate potential problems of this type with users.

Misclassification or confusion among Upland hardwoods is common. For example, 21
Aspen stands are classified as Aspen Associations, but seven Sugar maple/beech/yellow
birch stands are also classified as part of the Aspen Association (Table 7). Again these
misclassifications can have significant habitat management implications, and users must

be aware of these classification issues.

Additional Research

Based on the experiences of this study, the following areas of future study will be useful

to expand the understanding of accuracy of the IFMAP land cover raster map.

(1) Study on impact of patch size on accuracy of IFMAP

(2) Study on impact of landscape heterogeneity on accuracy of IFMAP

(3) Study on crosswalk formulation between forest classifications

(4) Compare the acreages classified by the four databases, using FIA expansion
factors

(5) Examine experiences from other states in improving raster land cover map

accuracy.

For several studies (1), (2) and (4), the research will be greatly enhanced if researchers

have direct access to the exact FIA plot locations and plot-level attributes.

116

REFERENCES CITED

Anderson, J. R., Hardy, E.E., Roach, J .T and Witrner, RE. 1976. A land use and land
cover classification system for use with remote sensor data, (a revision of the land
use classification system as presented in US. Geological Survey Circular 671),
USGS Publications Warehouse Geological Survey Professional Paper 964 In:
http.‘//www.wsdot.wa.gov/environment/envinfo/docs/RSPrj_USGS_lulcclass.pdf
(downloaded 06-15-05).

Berlanga-Robels, C. and Ruiz-Luna, A. 2002. Land use mapping and change detection
in the coastal zone of northwest Mexico using remote sensing techniques. Journal
of Coastal Research 82517-522.

Bishop, Y. M. M., Feinberg, S. E. and Holland, P. W. 1975. Discrete Multivariate
Analysis-Theory and Practice: Cambridge, Mass, The MIT Press.

Brand, G. 2005. Forest inventory and analysis sampling hexagons. FIA fact sheet series,
USFS. In: http://www.fia.fs.fed.us/library/fact-sheets/data-
collections/Sampling_hexagons.pdf (downloaded 06-15-05).

Brand, G.J., Nelson, M.D., Wendt, D.G., and Nimerfro, K. K. 2000. The hexagon/panel
system for selecting FIA plots under an annual inventory. In Proceedings of the
First Annual Forest Inventory and Analysis Symposium (McRoberts, R.E., Reams,
G.A., Van Deusen, PC, 2000 Eds.) Gen. Tech. Rep. NC-213. St. Paul,
MN: USDA Forest Service, North Central Research Station. pp. 8-13.

Burkman, B. 2005. FIA Inventory and Analysis Phase 2 and Phase 3: Ground
Measurements. FIA fact sheet series. USFS. In: http://fia.fs.fed.us/library/fact-
sheets/datacollections/Phase2_3.pdf (downloaded 06-15-05).

Campbell, J .B. 2002. Introduction to Remote Sensing (third edition). Guilford
Publications Inc., New York.

Chuvieco, E. and Congalton RC. 1988. Using cluster analysis to improve the selection
of training statistics in classifying remotely sensed data. Photogrammetric
Engineering and Remote Sensing 54:1275-1281.

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

Congalton, RC. 1981. The Use of Discrete Multivariate Analysis Techniques for the
Assessment of Landsat Classification Accuracy. MS Thesis. Virginia Polytechnic
Institute and State University, Blacksburg, Virginia, USA.

Congalton, RC. 1991. A review of assessing the accuracy of classifications of remotely
sensed data. Remote Sensing of Environments 37:35-46.

117

Congalton, RC. and Green, K. 1993. A practical look at the sources of confusion in

error matrix generation. Photogrammetric Engineering & Remote Sensing 59:641-
644.

Congalton, RC. and Green, K.l999. Assessing the Accuracy'of Remotely Sensed Data:
Principles and Practices. Lewis Publishers, Boca Raton. 137 p.

Congalton, R.G., Oderwald, R.G., and Mead, RA. 1983. Assessing Landsat
classification accuracy using discrete multivariate analysis statistical techniques.
Photogrammetric Engineering & Remote Sensing 49: 1671-1678.

Foody, GM. 2003. Status of land cover classification accuracy assessment. Remote
Sensing of Environment 80: 185-201.

Goodwin, RF. and Hudson, W.D. 2002. Comparison of Air Photo and Satellite Image
Sources for Updating Land Cover and Land Use Maps. Center for Remote Sensing
and GIS, Michigan State University June 2002. In:
http://www.rsgis.msu.edu/pdf/comparison.pdf, downloaded 06-15-05.In:
http://www.globalchange.msu.edu/pdfs/Forecast_Michigan_Project.pdf
(downloaded 06-15-05).

Hahn, J .T. and Hansen, M.H. 1985. Data Bases for Forest Inventory in the North Central
Region. Gen Tech. Rep. NC- 101. St. Paul, MN: USDA, Forest Service, North
Central Forest Experiment Station. 57 p.

Holden Geoff. Forest Inventory and Analysis, North Central Research Station. USDA-
FS. St. Paul, MN.

Homer, C., Haung, C., Yang, L., Wylie, B., and Coan, M. 2004. Development of a 2001
national land-cover database for the United States, Photogrammetric Engineering
& Remote Sensing, 70:829-840. In: httpzllwww.mrlc.gov/pdfs/July_PERS.pdf
(downloaded 06-15-05).

Hord, R. M. and Brooner, W. 1976. Land-Use map accuracy criteria. Photogrammetric
Engineering and Remote Sensing 42:671-677.

J anssen, L.L.F. and Van der Wel, F.J.M. 1994. Accuracy assessment of satellite derived
land-cover data: a review. Photogrammetric Engineering & Remote Sensing, 60:
419-426.

Jennings, M. 1993. Modified UNESCO Natural Terrestrial Cover Classification. Gap
Analysis handbook. In:
http://www.gap.uidaho.edu/handbook/LandCoverMapping/UNESCO/default.htm
(downloaded 06-15-05).

Katila, M., Heikkinen, J. and Tomppo, E. 2000. Calibration of small-area estimates for
map errors in multisource forest inventory. Canadian Journal of Forestry

118

Research 30: 1329-1339.

Kurvonen, L. and Hallikainen, M.T. 1999. Textural information of multitemporal ERS-l
and JER-l SAR images with applications to land and forestry type classification in
boreal zone. IIE Trans. Geosci. Remote Sensing 37:680-689.

Landis, J .R. and Koch, CG. 1977. The measurement of observer agreement for
categorical data. Biometrics 33: 159-174.

Lawson, Glenn H. n. d. Use of NR1 Point Data in Image Processing. USDA,
NRCS/NRIA In: http://www.nriai.nrcs.usda.gov/technical/analysis/glenn/
(downloaded 06-16-2005. 2005).

Leatherberry, E.C., Haugen, D, and Brand, G. 2005. Michigan’s Forest Resources in
2003. Resour. Bull. NC-245. USDA-PS, NCRS, St. Paul, MN. 39 p.

Leatherberry, E.C., Spencer, J .S, Jr. 1996. Michigan forest statistics 1993. Resour. Bull.
NC-170. USDA-FS, NCRS, St. Paul, MN. 144 p.

Liu, C., Zhang, L., Davis, C.J., Solomon, D.S., Brann, TB, and Caldwell, LE. 2003.
Comparison of neutral networks and statistical methods in classification of
ecological habitats using FIA data. Forest Science 49:619-631.

Loveland, TR, and Shaw, D.M. 1996. Multiresolution land characterization: building
collaborative partnerships, Gap Analysis: A landscape Approach to Biodiversity
Planning (Scoot, J .M., Teat, T., and Davis F. Eds.) Proceedings of the
ASPRS/GAP Symposium, Charlotte, North Carolina, pp 83-89.

Martin, M. E., Newman, S. D., Aber, J. D., Congalton, and R. G. 1998. Determining
forest species composition using high spectral resolution remote sensing data.
Remote Sensing of Environment 65:249-254.

Mayaux, P., and Lambin, ER 1997. Tropical forest area measured form global land
cover classifications: inverse calibration models based on spatial textures. Remote
Sensing of Environment 59:29-43.

McRoberts, RE. 1999. Joint annual forest inventory and monitoring system, the North
Central perspective. Journal of Forestry 97:27-31.

MDNR. 2002. Michigan’s Growing and Expanding Forest Resource. State of Michigan.
Release date April 01, 2002. In
httpzllwww.michigan.gov/printerFriendly/O,1687,7-153--23833--,00.html
(downloaded on 06-15-05).

MDNR. 2003a. Center for Geographic Information, Department of Information

Technology, Michigan (http://www.mcgi.state.mi.us/mgdll) (downloaded on 06-
15-05).

119

MDNR. 2003b. IFMAP / GAP Lower Peninsula Land Cover, Michigan Spatial Data
Library. In:
http://www.dnr.state.mi.us/spatialdatalibrary/sdl2/land_use_cover/2001/IFMAP_1p
_landcover.htm. (downloaded 06-15-05) 12 p.

MDNR. 2003c. IFMAP / GAP Upper Peninsula Land Cover, Michigan Spatial Data
Library. In:
http://www.dnr.state.mi.us/spatialdatalibrary/sdl2/1and__use_cover/2001/IFMAP_u
p_landcover.htm. (downloaded 06-15-05) 12 p.

MDNR. 2003d. Decision Rules for classes, Classification system key for remote sensing
and ground data gathering. Spatial Data library, MDNR. In:
http://www.dnr.state.mi.us/spatialdatalibrary/sdl2/land_use_cover/2001/IFMAP%
20Classification%20Rules.htm (downloaded on 06-15-05).

MDNR, 2003c. Description of the Classes Used in the Michigan Statewide Map,
Michigan Department of Natural Resources. In:
http://www.dnr.state.mi.us/spatialdata]ibrary/sdl2/1and_use_cover/2001/IFMAP%2
0TM%20Classification%20Description.htm (downloaded 06-15-05) 4 p.

MDNR. n. d. Operations Inventory Manuals, Michigan Department of Natural
Resources, Lansing, MI. (Unpublished Technical Guideline).

Miguel-Ayanz, J .S. and Biging, G.S. 1997.Comparison of single-stage and multi-stage
classification approaches for cover type mapping with TM and SPOT data. Remote
Sensing of Environment 59292-104.

Miles, P.D., Brand, G.J., Alerich, C.L., Bednar, L.F., Woudenberg, S.W., Glover, J .F.,
and Ezzell, EN. 2001. The Forest Inventory And Analysis Database: Database
Description and Users Manual Version 1.0. Gen Tech. Rep. NC-218. St. Paul,
MN: US. Department of Agriculture, Forest Service, North Central Research
Station. 130 p.

MIRIS. 1981. Current Use Inventory Classification System Definitions (May 1981).
Michigan Department of Natural Resources. In:
http://www.rsgjs.msu.edu/pdf/lclu/CUI.pdf (downloaded 06-15-05).

MRIS. 2000. Land Cover/Use Classification System, Version 11. MIRIS, Michigan
Resource Information System, Resource Mapping and Aerial Photography
Section, Land and Mineral Services Division, Department of Natural Resources In:
httpzllwww.rsgis.msu.edu/pdf/lclu/MIRIS_2_definitions.pdf (downloaded 06-16-
05).

Nelson, R., Case D., Homing, N., Anderson V., and Pillai, S. 1987. Continental land
cover assessment using Landsat MSS data. Remote Sensing of Environments
21:61-81.

120

NRIS FSVeg. 2004. Common Stand Exam Users Guide Version 1.6 July 2004. Natural
Resource Information System: Field Sampled Vegetation, USDA Forest Service.

NRIS FSVeg. 2005. FSVeg Data Dictionary Version 1.7, January 2005. Natural
Resource Information System: Field Sampled Vegetation, USDA Forest Service.

Powell, R.L., Matzke, N., de Souza, C.J., Clark, M., Numata, I., Hess, LL, and
Roberts, DA. 2004. Sources of error in accuracy assessment of thematic land-
cover maps in the Brazilian Amazon. Remote Sensing of Environment 90:221-234.

Fax-Lenny, M., Woodcock, C.E., Macomber, S.A., Gopal 8., Song C. 2001. Forest
mapping with a generalized classifier and Landsat TM data. Remote Sensing of
Environment 77:241-250.

Pedersen, Larry. 2000. Contrasts of State Forest FIA and 01 Data: Can they be
reconciled? Michigan Department of Natural Resources (Unpublished report). 9 p.

Pedersen, Larry. Forest Minerals and Fire Management, Michigan Department of
Natural Resources. Lansing, MI.

PMR. 2001a. Integrated Forest Monitoring Assessment and Prescription IFMAP,
Review of Remote Sensing Technologies for the IFMAP Project, 2nd Revision.
(prepared for) State of Michigan, Michigan Department of Natural Resources,
Pacific Meridian Resources (PMR), Ann Arbor, Michigan. 78 p. In:
http://www.dnr.statemi.us/spatialdatalibrarv/sdlZ/land use cover/2001/IFMAP%
20Remote%20Sensing%20Report.doc (downloaded 06-15-05).

PMR. 2001b. Michigan DNR Integrated Forest Monitoring Assessment and
Prescription, Current Field Inventory Design. (prepared for) State of Michigan,

Michigan Department of Natural Resources, June, 2001. Pacific Meridian
Resources (PMR), A Space Imaging Company, Ann Arbor, Michigan. 56 p.

Rosenfield, OH. and Fitzpatric-Lins K. 1986. A coefficient of agreement as a measure

of thematic classification accuracy. Photogrammetric Engineering & Remote
Sensing 52:223-227.

RS and GIS. n .d. RS&GIS History in Land Cover/Land Use Issues. Remote sensing
and Outreach, Remote sensing and Geographic Information system, Michigan
State University, East Lansing, MI, 48823. In:
http://www.rsgis.msu.edu/pdf/lclu/lulc_history.pdf (downloaded 06-15-05).

Saatchi, SS. and Rignot E. 1996. Classification of boreal forest cover types using ASR
images. Remote Sensing of Environment 60:270-281.

Schmidt, T.L., Spencer, John S.J., and Bertsch, R. 1997. Michigan’s forests, 1993: an
analysis. Resour. Bull. NC-179. St. Paul, MN: US. Department of Agriculture,
Forest Service, North Central Forest Experimental Station. 96 p.

121

Schreuder, H.T., Bain, S., Czaplewski, RC. 2003. Accuracy Assessment of Percent
Canopy Cover, Cover Type and Size Class. Gen. Tech. Rep. RMRS-GTR-108.
Fort Collins, CO: USDA-FS, Rocky Mountain Research Station. 10 p.

SI. 2004. Integrated Forest Monitoring Assessment and Prescription (IFMAP), Review
of Remote Sensing Technologies Used in the IFMAP Project, Final Report.
(prepared for) State of Michigan, Michigan Department of Natural Resources
(10/25/2004). Space Imaging (SI), Ann Arbor, Michigan.

Silbemagel J ., Chen J ., Gale, M.R., Pregitzer, KS, and Probst, J. 1997. An
Interpretation of Landscape Structure from Historic and Present Land Cover Data
in the Eastern Upper Peninsula of Michigan. Gen. Tech. Rep. NC-192. St. Paul,
MN: USDA-FS, North Central Forest Experiment Station. 30 p.

Skole, D.L., Batzli, S., Gage, S., Pijanowski, B., Chomentowski, W., Zhou, Y., Rustem,
W., and Salas W. 2002. Land Use Issues and Policy: Forecasting Land Use
Change in Michigan, August 2002. A Forecast Michigan Project, Center for
Global Change and Earth Observations, Michigan State University.

Smith, J H Wickham, J .D., Stehman, S.V., Yang, Lirnin. 2002. Impacts of patch size
and land-cover heterogeneity on thematic image classification accuracy.
Photogrammetric Engineering & Remote Sensing 68:65-70.

Smith, W. B., Miles, P.D., Vissage, J .S., and Pugh, SA. 2004. Forest Resources of the
United States, 2002. Gen. Tech. Rep. NC-241. St. Paul, MN: USDA, Forest
Service, NCRS. 137 p.

Stehman, S. V. 1999. Basic probability sampling designs for thematic map accuracy
assessment. International Journal of Remote Sensing 20:2423-2441.

Story, M. and Congalton, RC. 1986. Accuracy assessment: a user’s perspective.
Photogrammetric Engineering & Remote Sensing 52:397-399.

USDA-PS. 2004. Forest Inventory and Analysis National Core Field Guide Volume 1:
Field Data Collection Procedures For Phase 2 Plots Version 2.0. In:
http://www.fia.fs.fed.usllibrarv/field-guides-methods-proc/docsfiore ver 2-

O 01 04.pdf (downloaded 06-05-05).

Wang, C., Qi, J ., Skole, D.L., Cochrane MA. and Matricardi, E.A.T . 2003. Tropical
Forest Degradation and Recovery: Estimation and Validation of Canopy Cover
with Landsat and IKONOS Imagery. CGCEO/RA01-03/w Michigan State
University, East Lansing, Michigan.

Wessels K. J., De Fries, R.S., Dempewolf, J., Anderson, L.O., Hansen, A.J. Powell,
S.L., Moron, ER 2004. Mapping regional land cover with MODIS data for
biological conservation: examples from the Greater Yellowstone Ecosystem, USA
and Para State, Brazil. Remote Sensing of Environment 92:67-83

122

Wickham, J .D., Sethman, S. V., Smith J. H., and Yang, L. 2004. Thematic accuracy of
the 1991 National Land-Cover Data for the western United States. Remote Sensing
of Environment 91: 452-468.

Wynne, R.H., Oderwald, R.G., Reams, and Scrivani, J. A. 2000. Optical remote sensing
for forest area estimation. Journal of Forestry 98:31-36

123

APPENDIX

124

APPENDIX A

Table A1. NRIS-FSVeg FIA Crosswalk

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

FSVeg FIA
Code NRIS-FSVeg Classification CODE FIA Classification

1 Jack Pine 101 Jack pine

2 Red Pine 102 Red pine

3 White Pine 103 Eastern white pine

5 Hemlock 105 Eastern hemlock
11 Balsam Fir_Asp-PB 121 Balsam Fir
12 Black Spruce 125 Black Spruce
14 Northern Wh Cedar 127 Northern white Cedar
18 Mix Swamp conifer 127 Northern white Cedar
18 Mix Swamp conifer 125 Black Spruce
19 Cedar-Aspen-PB 127 Northern white Cedar
21 Mixed Northeren Hdw 520 Mixed Upland Hdwd
48 Jack Pine-Oak 101 Jack pine
49 Red pine-Oak 102 Redpine
53 Black Oak 515 Chestnut oak lblack oak/scarlet
55 Northern Red Oak 503 White oak / red oak / hickory
59 Mixed Oak 503 White oak / red oak / hickory
59 Mixed Oak 504 White oak
59 Mixed Oak 505 Northern red oak
71 BI Ash-Elm-R Maple 701 Black ash/ American elm / red maple
81 S Maple-Beech -YB 800 Maple / Beech / Birch Grogp
81 S Mme-Beech -YB 801 Sugar maple / beech/ yellow birch
82 S Maple-Basswood 801 quar maple / beech/ yellow birch
82 S Maple-Basswood 805 Hard maple / basswood
84 Red Maple(Dry) 809 Red maple upland
85 Super Maple 800 Maple / Beech / Birch Group
85 Sign Maple 801 Sign maple / beech/ yellow birch
85 Sugar Maple 805 Hard maple / basswood

Sugar maple-beech-yellow birch-

87 red spruce 801 Sugar maple/ beech/ yellow birch
89 Mixed Upland Hdwd 520 Mixed Upland Hdwd
89 Mixed Upland Hdwd 801 Spugar maple / beech/ yellow birch
89 Mixed Upland Hdwd 805 Hard maple / basswood
89 Mixed Upland Hdwd 809 Red maple / upland
90 Mixed Upland Hdwd 805 Hard maple / basswood
91 Quakifl Aspen 901 Aspen
93 Bigtooth Aspen 901 Aspen
95 Asp-W Spruce BF 901 Aspen
97 Lowland Brush 125 Black spruce
99 Open 999 Nonstocked

 

125

 

APPENDIX A

Table A2. IFMAP-FIA Crosswalk

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

IFMAP FIA
code IFMAP Classlflcation Code FIA Classification
10 Herbaceous Openland 999 Nonstocked
12 Upland Shrub /Lowdensity trees 999 Nonstocked
14 Northern Hardwood Association 105 Eastern hemlock
14 Northern Hardwood Association 800 Maple / Beech / Birch Group
14 Northern Hardwood Association 801 Sugar maple / beech / yellow birch
14 Northern Hardwood Association 805 Hard maple / basswood
14 Northern Hardwood Association 809 Red maple / upland
15 Oak Association 503 White oak / red oak / hickory
16 Aspen Association 901 Aspen
19 Pines 101 Jack pine
19 Pines 102 Red pine
21 Mixed Upland Conifer 101 Jack pine
21 Mixed Upland Conifer 102 Red pine
21 Mixed Upland Conifer 103 Eastern white pine
21 Mixed Upland Conifer 400 Oak Pine Group_
22 Upland Mixed Forest 503 White oak / red oak / hickory
22 Upland Mixed Forest 801 Sugar maple / beech / yellow birch
22 Upland Mixed Forest 901 Aspen
22 Upland Mixed Forest 800 Maple / Beech I Birch Group
22 Upland Mixed Forest 809 Red maple / upland
21 Mixed Upland Conifer 401 White pine / Red oak / white ash
24 Lowland Deciduous Forest 701 Black ash /American elm / red maple
25 Lowland Coniferous Forest 121 Balsam fir
25 Lowland Coniferous Forest 125 Black spruce
25 Lowland Coniferous Forest 127 Northern white-cedar
28 Lowland Shrub 125 Black spruce
28 Lowland Shrub 127 Northern white-cedar
30 Mixed Non Forest Wetland 127 Northern white-cedar
30 Mixed Non Forest Wetland 125 Black spruce

 

126

 

APPENDIX A

Table A3. Ol-FIA Crosswalk

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

OI FIA
Code OI Classification Code FIA Classification
A Aspen 901 Aspen
C Cedar 127 Northern White-cedar
D Treed Bog 125 Black spruce
E Swamp Hrdwds 700 Elm / Ash / Cottonwood Group
Black ash /American elm / red
E Swamp Hrdwds 701 maple
F Spruce fir 121 Balsam fir
F Spruce fir 125 Black spruce
J Jack pine 101 Jack pine
L Lowland Brush 125 Black spruce
M Northern Hardwood 520 Mixed upland hardwoods
M Nothern Hardwood 801 Sugar maple / beech / yellow birch
M Nothern Hardwood 805 Hard maple / basswood
M Nothern Hardwood 809 Red maple/ upland
0 Oak 503 White oak / red oak / hickory
Q Mx Swamp Cnfr 121 Balsam fir
Q Mx Swamp Cnfr 125 Black spruce
Q Mx Swamp Cnfr 126 Tamarack
Q Mx Swamp Cnfr 127 Northern White-cedar
R Red pine 102 Redpine
S Black Spruce 125 Black spruce
T Tamarack 126 Tamarack

 

127

 

APPENDIX A

Table A4. IFMAP-OI Crosswalk

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

IFMAP OI
code lFMAP Classification Code OI Classification
12 Upland Shrub/low density trees U Upland Brush
14 Northern Hardwood Association M Nothern Hardwood
15 Oak Association O Oak
16 Aspen Association A Aspen
19 Pines J Jack pine
19 Pines R Red pine
21 Mixed Upland Conifers W White pine
22 Upland Mixed Forest A Aspen
22 Upland Mixed Forest M Nothern Hardwood
25 Lowland Coniferous Forest C Cedar
25 Lowland Coniferous Forest D Treed Bog
25 Lowland Coniferous Forest 0 Mixed Swamp Conifer
25 Lowland Coniferous Forest S Black Spruce
25 Lowland Coniferous Forest T Tamarack
28 Lowland Shrub S Black Spruce
28 Lowland Shrub L Lowland Brush
28 Lowland Shrub Q Mixed Swamp Conifer
30 Mixed Non-Forest Wetland L Lowland Brush
30 Mixed Non-Forest Wetland O Mixed Swamp Conifer

 

128

 

Appendix A

Table A5. Crosswalk lFMAP-NRIS—FSVeg

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(a)
IFMAP
code lFMAP Classification FSVeg NRIS-FS Veg Classificaiton
1O Herbaceous Openland 999 Nonstocked
12 Upland Shrub /Lowdensity trees 999 Nonstocked
14 Northern Hardwood Association 5 Hemlock
14 Northern Hardwood Association 81 S Maple-Beech -YB
14 Northern Hardwood Association 82 S Maple-Basswood
14 Northern Hardwood Association 84 Red maple(dry)
14 Northern Hardwood Association 85 Sugar maple
14 Northern Hardwood Association 85 Sugar maple / beech / yellow birch
14 Northern Hardwood Association 85 Sugar Maple
14 Northern Hardwood Association 90 Mixed upland hardwoods
15 Oak Association 59 Mixed Oak
16 Aspen Association 11 Balsam Fir_Asp-PB
16 Aspen Association 91 Aspen
16 Aspen Association 93 Bigtooth Aspen
16 Aspen Association 95 Asp-W Spruce BF
19 Pines 1 Jack pine
19 Pines 2 Red pine
19 Pines 3 White pine
19 Pines 48 Jack Pine-Oak
19 Pines 49 Red Pine-Oak
20 Other Upland Conifer 5 Hemlock
20 Other Upland Conifer 11 Balsam fir-Asp-PB
20 Other Upland Conifer 8 White spruce
20 Other Upland Conifer 19 Cedar-Aspen-PB
21 Mixed Upland Conifer 3 White pine
22 Upland Mixed Forest 5 Hemlock
22 Upland Mixed Forest 48 Jack Pine-Oak
22 Upland Mixed Forest 53 Black Oak
22 Upland Mixed Forest 59 Mixed Oak
22 Upland Mixed Forest 81 S Maple-Beech -YB
22 Upland Mixed Forest 85 Mar Maple
Sugar maple-beech-yellow birch/red
22 Upland Mixed Forest 87 spruce
22 Upland Mixed Forest 89 Mixed Upland Hdwd
22 Upland Mixed Forest 91 QuakingAspen
22 Upland Mixed Forest 93 Bigtooth Aspen
22 Upland Mixed Forest 95 Asp-W Spruce BF
24 Lowland Deciduous Forest 71 BI Ash-Elm-Red maple
24 Lowland Deciduous Forest 91 Quaking Aspen
25 Lowland Coniferous forest 11 Balsam fir-AspoPB
25 Lowland Coniferous forest 12 Black spruce

 

 

 

 

129

 

 

Table 5. (Cont’d).
(b)

 

 

 

 

 

 

 

 

25 Lowland Coniferous forest 14 Northern Wh Cedar
25 Lowland Coniferous forest 18 Mix Swamp Conifer
25 Lowland Coniferous forest 19 Cedar-Aspen-PB
28 Lowland Shrub 12 Black Spruce

28 Lowland Shrub 14 Northern Wh Cedar
28 Lowland Shrub 18 Mix Swamp Conifer
28 Lowland Shrub 97 Lowland Brush

30 Mixed Non-Forest Wetland 12 Black spruce

 

30

Mixed Non-Forest Wetland

18

Mix Swamp Conifer

 

30

Mixed Non-Forest Wetland

97

Lowland Brush

 

 

30

 

Mixed Non-Forest Wetland

 

99

 

Open

 

130

 

i

{a ’34 v1.1.5}. ii

1 &.1\ ‘13er ‘QIArleciL. viii; \L,I.A..\\

f t. O . .1 .. Y-'.dr.1i..‘1

44 w

   

llllll1lllllll‘lllijljllllljfl