. ”an.
W‘s”:

.5?»
. A r z. .
:1 . 2.4..v‘.‘
.: :fi 11.
.3... .

 

5%

.

3:...
. . .12.

I 21...... arid,

.....1....m.. .
wan

 

 

 

 

Sari‘... .57.? .,

Meals

 

 

' LIBRARY f
54395; I 93 Michigan State "
— University ——~

This is to certify that the
dissertation entitled

IDENTIFICATION AND MAPPING OF AGRICULTURAL
TILLAGE METHODS UTILIZING REMOTELY SENSED DATA

presented by

Stephen R. South

has been accepted towards fulfillment
of the requirements for the

Doctoral degree in Geography

 

 

’K/lajor %fessor’s Signature

(fizz/03

Date

 

MSU is an Affirmative Action/Equal Opportunity Institution

 

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

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 cJCIRC/DataDue.p65-p.15

 

IDENTIFICATION AND MAPPING OF AGRICULTURAL TILLAGE METHODS
UTILIZING REMOTELY SENSED DATA

By

Stephen R. South

A DISSERTATION

Submitted to
Michigan State University
In partial fulfillment of the requirements
For the degree of

DOCTOR OF PHILOSOPHY

Department of Geography

2003

ABSTRACT

IDENTIFICATION AND MAPPING OF AGRICULTURAL TILLAGE METHODS
UTILIZING REMOTELY SENSED DATA

By

Stephen R. South

The impact of widespread land use/land cover conversion from natural
ecosystems to managed agricultural production has significantly altered long
term ecological processes and balances. The methods employed in crop
production, encompassing hundreds of thousands of individual fields, collectively
' account for 199 million acres of current land use. In order to assess the impacts
of cropping methods, fundamental spatial data are required. Data regarding the
total spatial extent and distribution of cropping methods would provide much
needed data to assess and monitor the environmental impacts of widespread row

crop agriculture.

Remotely sensed data provides a means to quickly and cost effectively
monitor cropping methods over large areas. Increased use of conservation tillage
methods may enhance carbon sequestration rates in soils and significantly
reduce erosion of topsoil, the nations largest contribution of non point source

pollution. The differentiation of cropping methods using remotely sensed data

would provide current estimates of environmental impacts, and data for use as
input into environmental models, to predict future consequences and impacts of

man’s influence on the environment.

This study investigates the use of remotely sensed data to identify
conventional and conservation tillage methods. Landsat 7 ETM + data, covering
a 180 x180 km study area over portions of Michigan, Indiana, and Ohio were
analyzed to document and map agricultural cropping practices. Several
classification techniques, including spectral angle methods, were examined to
assess their suitability to differentiate conventional and conservation tillage
practices. The results indicate that of the 3.6 million acres of agricultural land
use/land cover identified within the study area, 1.8 million acres (52%) were
cropped using conservation tillage methods. The current yearly carbon
sequestration potential of the study area conservation tilled fields is estimated at

228,490 metric tons/year.

In summary, remote sensing techniques and methods used in this study
have the potential to provide a great deal of data regarding the implications of
cropping methods. The results of this study, scaled up to larger regional or
watershed scales, would provide a time and cost effective method to assess

cropping methods and their associated environmental impacts.

Between the idea and the reality
Between the motion and the act
Falls the shadow

TS. Eliot — The Hollow Men

TABLE OF CONTENTS

LIST OF TABLES ............................................................................................... Vll
LIST OF FIGURES ............................................................................................ Vlll
CHAPTER 1 INTRODUCTION ......................................................................... 1
1 .1 Introduction ............................................................................................ 1
1.2 Land Use and Environment ................................................................... 2
1.3 Cropping Methods .................................................................................. 4
1.3.1 Conventional Tillage ........................................................................... 5
1.3.2 Conservation Tillage ........................................................................... 7
1.4 Research Needs .................................................................................. 11
1.5 Statement of Problem .......................................................................... 12
CHAPTER 2 LITERATURE REVIEW .............................................................. 14
2.1 Digital Remote Sensing ....................................................................... 14
2.2 Spectral Characteristics/Considerations .............................................. 16
2.3 Crop Residue Spectral Characteristics ................................................ 17
2.4 Soil Spectral Characteristics ................................................................ 19
2.5 Classification Methodologies ............................................................... 20
2.6 Classification Decision Rules ............................................................... 26
2.6.1 Minimum Distance ............................................................................ 26
2.6.2 Mahalanobis Classification ............................................................... 26
2.6.3 Maximum Likelihood ......................................................................... 26
2.6.4 Spectral Angle Mapping ................................................................... 28
2.6.5 Cosine of the Angle Concept ............................................................ 29
2.7 Accuracy Assessment .......................................................................... 32
CHAPTER 3 MAPPING TILLAGE PRACTICES ............................................. 34
3.1 Study Area ........................................................................................... 35
3.2 Study Area Physical Characteristics .................................................... 37
3.4 Study Area Agriculture ......................................................................... 41
3.5 Ground Truth Data Collection .............................................................. 46
3.6 Laboratory Radiometric Sampling ........................................................ 47
3.7 Field Based Radiometric Sampling ...................................................... 48
3.8 Ground Truthing — Driving Transects ................................................... 49
3.9 Temporal Considerations ..................................................................... 56
3.10 Imagery Acquisition .............................................................................. 58

3.11 Imagery Accessibility Considerations ................................................... 60
3.12 Study Area Imagery ............................................................................. 61
3.13 Imagery Processing ............................................................................. 63
3.14 Radiance .............................................................................................. 63
3.15 Atmospheric Correction ....................................................................... 64
3.16 Atmospheric Correction -- 55 Method ................................................... 68
3.17 Classification Methodology .................................................................. 71
3.18 Classification Accuracy Assessment .................................................... 73
CHAPTER 4 RESULTS ................................................................................... 76
4.1 Spectral Properties of Tillage Practices ............................................... 76
4.1.2 Field Based Spectral Properties ....................................................... 79
4.1.3 Satellite Spectral Properties ............................................................. 84
4.1.4 Spectral Characteristics at Differing Spatial Scales .......................... 85
4.2 Classification Training Sites ................................................................. 89
4.3 Classification Methodology Comparison .............................................. 89
4.4 Classified Imagery Results .................................................................. 98
4.5 Cosine of the Angle Concept — Spectral Angle Mapping ................... 101
4.6 Angular vs. Geometric and Probability Based Classification .............. 103
4.7 Validation ........................................................................................... 105
4.8 Potential Sources of Classification Error ............................................ 111
4.8.1 Soil Background Effect ................................................................... 111
4.8.2 Mixed Pixels ................................................................................... 113
CHAPTER 5 CONCLUSIONS / DISCUSSION .............................................. 114
5.1 Carbon Cycle Implications ................................................................. 118
5.2 Regional Carbon Accounting ............................................................. 123
5.3 Carbon and Tillage ............................................................................. 126
5.4 Study Area Total Carbon Sequestration ............................................ 128
5.5 Regional Estimates of Carbon Sequestration Based on Tillage ......... 129
5.6 Soil Erosion and Tillage ..................................................................... 130
5.7 Scaling Up - Regional Measurements and Estimates ....................... 133
5.8 Further Research ............................................................................... 135
BIBLIOGRAPHY ............................................................................................... 139

vi

LIST OF TABLES

Table 1.
Table 2.
Table 3.
Table 4.
Table 5.
Table 6.
Table 7.
Table 8.
Table 9.

Table 10.
Table 11.
Table 12.
Table 13.
Table 14.

Study Area Acreage and Production Value .......................................... 37
ETM+ Sensor Characteristics ............................................................... 62
55 Derived Correction Parameters for ETM+ data. .............................. 71
Maximum Likelihood Classification Error Matrix. .................................. 91
Mahalanobis Distance Classification Error Matrix. ............................... 92
Minimum Distance Classification Error Matrix. ..................................... 93
Spectral Angle Mapping Classification Error Matrix .............................. 94
Cosine of the Angle Concept Classification Error Matrix. ..................... 95
Classification Results. .......................................................................... 96
Results of Classification Statistical Significance Pairings. .................. 97
Study Area Land Use/Land Cover. ................................................... 100
Validation Results of the Cosine of the Angle Classification Routine.111
Yearly Carbon Sequestration Potential ............................................ 130
Temporal Image Acquisition for Crop Identification. ......................... 135

vii

LIST OF FIGURES

Figure 1. Classification Decision Space. ........................................................... 23
Figure 2. Cosine of the Angle Concept Decision Space .................................... 30
Figure 3. Diagram of Methods ........................................................................... 34
Figure 4. Study Area Location. .......................................................................... 36
Figure 5. Study Area Precipitation and Potential Evapotranspiration. ............... 38
Figure 6. Study Area Yearly Temperature Graph. ............................................. 38
Figure 7. Driving Transect Data Points. ............................................................ 51
Figure 8. Ground Truth Data Collection Flow Chart. ......................................... 52
Figure 9. Landsat 7 P21 R31 Band 5 May 27‘“, 2002. ...................................... 62
Figure 10. 55 Atmospheric Correction Flowchart. ............................................. 69
Figure 11. Laboratory Reflectance Readings. .................................................. 77
Figure 12. Field Based Spectral Reflectance Measurements, .......................... 80
Figure 13. Crop Residue - Photograph of Com Residue .................................. 81
Figure 14. Crop Residue, Photograph of Corn Residues. ................................. 81
Figure 15. Crop Residue, Photograph of Soy Bean Residues. ......................... 82
Figure 16. Field Based Measurements of Reflectance ..................................... 83
Figure 17. Spectral Reflectance derived from ETM+ Data ............................... 84
Figure 18. Classified Data Based on the COTA Classification Routine ............. 99
Figure 19. Study Area Land Use/Land Cover in Acres .................................... 101
Figure 20. Soil Types and Ground Truth Collection Locations ........................ 112
Figure 21. Agriculture and Carbon. ................................................................. 124

viii

CHAPTER 1 INTRODUCTION

1.1 Introduction

A country drive through any midwestern state reveals that one land use
dominates the landscape, row crops. The three primary row crops corn,
soybeans, and wheat collectively account for 199,000,000 acres of land use
(USDA, 2000). Corn and soybeans each account for 73 million acres and wheat
production occurs on 53 million acres (USDA, 2000). Collectively these three

crops account for 36.5 billion dollars of agricultural production (USDA, 2000).

The implications and consequences of cropping methods, comprising
such a large area, are far reaching. The conversion of native ecosystems to
managed agricultural production, consisting primarily of row crops, exhibits many
environmental consequences. Cropping methods affect carbon sequestration
rates, soil erosion and in turn water quality. Additionally, pesticide and herbicide
applications vary according to cropping methods. Conservation tillage techniques
exhibit many environmental benefits over traditional tillage techniques. Of these
199 million acres of row crops, how many acres are grown under conventional or
conservation tillage cropping practices. To date, the only method of estimating
the total acreage by each cropping method is through the use of driving
transects. Driving transects are time consuming, costly, and most importantly,

incomplete.

 

ad

the

the
agr

OIv

While there are many articles in the current literature related to the
environmental benefits of conservation tillage cropping systems there remains a
significant gap in the literature regarding the spatial extent and location of these
areas. Currently no data exists regarding the spatial distribution or total area of
agricultural fields utilizing conservation tillage techniques. The primary objective
of this study is to investigate the possibility of utilizing remotely sensed data to

differentiate cropping methods.

1.2 Land Use and Environment

Natural physical and biotic processes have been the primary mechanisms
shaping the surface of the earth for millions of years. Man’s impact on the
environment has substantially altered the balance of natural processes across
the globe. In all but recent history, the activities of man took place on limited
areas, and the effect of human settlement was relatively small compared to the
total earth surface. Increasing human population and technological
advancements (agricultural and industrial revolutions) have significantly changed
the processes that shape the earth’s surface. The surface of the earth is in a
constant state of flux, and humans are playing an increasingly greater role in
these processes. Land use/land cover change associated with row crop
agriculture, deforestation, and urbanization major agents of widespread change

of what was once a natural landscape.

 

The carbon cycle is a particular aspect of the ecosystem that has been
significantly altered due to anthropomorphic processes. Several studies have
shown that the overall atmospheric concentration of carbon dioxide has been
dramatically rising over recent decades (Fan et al., 1998, Appenzeller, 1998). A
significant amount of scientific research has been undertaken to better
understand the carbon cycle and man’s impact and influences on the natural flow
of carbon through the atmosphere, land surface, and oceans (Tans 1990,
Houghton 1999, Houghton 1999a, Field & Fung 1999). To better understand
man’s role in altering earth’s processes, a better geographic understanding is
required. To what extent has man changed the natural balance, where are these
changes taking place, what is the total area of change, and how will these

changes affect our environment, now and in the future.

The 199 million acres of cropland scattered throughout the United States
used for agricultural production of corn, soybeans, and wheat accounts for a very
large area spatially. In fact, if the total acreage were aggregated together, the
resulting area would completely cover the states of Illinois, Indiana, and Iowa.
Yet little to no data exists regarding the spatial distribution and total area of
production practices used to raise the annual crops. The two primary
management practices, differing significantly in terms of environmental impact,
are routinely in use today but no accurate estimate of total acreage or the spatial
distribution of cropping methods across the landscape exists. A single 100 acre

field has the potential to sequester approximately 12 million grams (12 metric

tons) of carbon annually and significantly reduce soil erosion. Applying the same
carbon sequestration potential of 30 g/C/mZ/year to the 199 million acres of row
crop agricultural production results in a staggering carbon sequestration potential
of 24,159,732 metric tons of carbon per year. In order to better understand the
impacts of cropping methods on the natural environment basic spatial information
regarding the spatial distribution and total acreage of the two primary cropping
methods are required. Quantifying the acreage of each cropping method will
allow for the accurate estimate of carbon sequestration and soil erosion across

agricultural lands as well as other environmental effects and benefits.

1.3 Cropping Methods

Conventional and conservation tillage methods are the primary agricultural
production methods found throughout the Midwestern United States. Each
cropping method has associated trade offs and benefits that every producer must
evaluate based on environmental, conservation, social and economic
implications. Generally, factors affecting cropping method decisions are first and
foremost financial, with soil conservation and other environmental benefits being
secondary. Producers however, also realize there is a temporal element in
maximizing economic gains. Short-term economic gains, yearly crop production,
may result in long-term environmental consequences such as topsoil erosion,
ultimately leading to diminished yields and profits. Producers must weigh the

short-term economic benefits or advantages against long-term sustainability.

While the decision as to which production method to utilize ultimately lies
in the hands of the producer, society has an interest as well. “The negative
impacts associated with agricultural production, and the use of conventional
tillage systems in particular, include soil erosion, energy use, leaching and runoff
of agricultural chemicals and carbon emissions “ (Uri, 1999). Clearly, the impact
of agricultural production methods, made by thousands of individual farmers,

reaches well beyond the scope of single fields and farms.

The consequences of agricultural production affect not only farmer’s fields
but also, more importantly, the larger shared environment. Production methods
vary significantly regarding environmental costs and benefits such as soil
erosion. Soil erosion from agricultural fields is the largest source of non point
source water pollution in the United States. The federal government currently has
programs in place to create incentives to protect highly erodible lands but the
future of conservation incentives may lie with the reduction or sequestration of

greenhouse gasses.

1.3.1 Conventional Tillage

Conventional or traditional tillage consists of a combination of plowing,
disking, and tilling to condition and loosen the soil, mix in organic matter, and to
prepare a flat and uniform seed bed. Conventional tillage practices create a soil
surface highly susceptible to erosion until the crop has grown sufficiently to

provide protection from environmental conditions such as wind and precipitation.

Soil erosion is the primary disadvantage of conventional tillage from a producer’s
point of view. Conventional tillage methods also release C02 to the atmosphere,
contributing to global warming. Moldboard plowing mixes oxygen into the soil
which oxidizes soil organic matter carbon, releasing carbon to the atmosphere as
carbon dioxide. Conventional tillage techniques are the primary mechanism for
carbon dioxide release from soils (Reicosky, 1996). Conventional tillage,
however, does provide some measure of weed control in the field. Spring plowing
kills early emerging weeds and brings weed seeds to the surface where they are

susceptible to late frosts, and other weather conditions.

Conventional tillage was initially required to convert native ecosystems
such as grassland and forested areas into agricultural production. Additionally,
before the advent of effective herbicides, moldboard plowing was the primary
method of initial weed control in agricultural fields. The tradition of moldboard
plowing remained a standard method of field preparation throughout most of the
twentieth century. Farmers engaged in field crop production utilized moldboard
plowing and the technique was passed from generation to generation. There was
no reason to abandon conventional tillage techniques, it provided plentiful yields
and was a technique with which farmers were familiar. It wasn’t until the 19705
that the technology for conservation tillage methods became widely accessible in
the marketplace. Most family operated farms saw the cost of a specialized seed
drill, required for conservation tillage techniques, as cost prohibitive. However,

technological advancements in machinery, crop varieties, and agricultural

chemicals have significantly altered the economics of crop production and

conservation tillage methods are quickly spreading throughout the land.

1.3.2 Conservation Tillage

Conservation tillage practices are characterized by leaving the stubble
from the previous year’s crop on the soil surface. The stubble from the previous
year’s crop, left on the soil surface, dramatically reduces soil erosion. The
stubble binds highly productive topsoil in place on the farmer’s field preventing it
from blowing or washing away over the years. Carbon sequestration is an
additional long-term benefit associated with conservation tillage cropping
methods. Conservation tillage cropping has been shown to enhance carbon

sequestration rates in agricultural soils (Robertson et al., 2000).

Conservation tillage cropping methods, which leave plant litter in place,
generally improve soil conditions. Plant litter positively affects soil conditions by
affecting soil properties such as water infiltration, evaporation, soil porosity and
soil temperatures (Reicosky, 1996). Litter left in place on the soil surface
influences the flow of nutrients, carbon, water, and energy in terrestrial
ecosystems. The decay of litter adds nutrients to the soil, improves soils structure

and reduces soil erosion (Aase and Tanaka, 1991).

Not only does conservation tillage cropping exhibit positive environmental

benefits but economic benefits as well. Conservation tillage cropping methods

increase overall cost efficiency by way of reduced operation costs. Conventional

tillage practices require the additional steps of plowing, cultivation, and disking as
compared to conservation tillage cropping techniques. The costs of operating the
tractor, the time to perform these tasks, and the cost of the implements are saved

under conservation tillage cropping methods (Parsch, 2001).

While the yearly operating costs are lower with conservation tillage
cropping methods, they do require an initial capitol outlay to purchase the
specialized planting equipment required to successfully raise crops using
conservation tillage techniques. The longer term economic advantages must be
weighed against the short term capitol expense of a specialized planter or “drill”.
A seed drill cuts through the previous year’s crop stubble and the unplowed soil
surface beneath to plant the seed at the appropriate depth below the soil surface.
Most farmers already own a traditional planter and see the cost of a specialized
conservation tillage planter as cost prohibitive to adopting conservation tillage
cropping techniques. In the past conservation tillage yields were generally 5 — 10
bushels per acre less than conventional cropping techniques. Technological
advances in machinery, agricultural chemicals and round-up ready crop strains
have closed the gap in yield between the two cropping methods. Farmers, as
well as research in the literature, report that difference in yield between the two
cropping methods is now negligible. (Parsch, 2001). In fact, research has shown
that continued conservation tillage methods result in higher yields due to

enhanced soil fertility (Boyd, 2001). Agricultural fields must be cropped by

 

conservation tillage methods for six years in order to raise soil fertility levels

enough to positively effect yields (Boyd, 2001).

Conservation tillage techniques have, in the recent past, required more
herbicide treatment than conventional cropping techniques thereby increasing
cost. The advent of GMO (Genetically Modified Organism) hybrids, such as
round-up ready corn and soybeans, has changed this necessity. Herbicide
treatment on GMO soybeans has become extremely effective, requiring only one
or two herbicide applications to effectively control all weeds in a field. GMO crops
are genetically designed to work with a specific herbicide; the herbicide kills
everything but the genetically engineered plant. These advances in agricultural
technology have given conservation tillage cropping methods the upper hand in
production cost. Conservation tillage cropping methods are quickly spreading
throughout the Midwest as the cost savings of conservation tillage cropping over

conventionally cropped methods increases.

The associated benefits of conservation tillage are meaningful at a variety
of scales. At the individual farm level conservation tillage production methods
result in:

- Reduction of labor and time spent preparing fields for planting
- Reduced costs in maintaining and operating farm machinery
- Improved water infiltration and storage

- Improved soil fertility

- Increased profit

 

Scaling up to the regional level reveals that the benefits of conservation
tillage are not reaped by the farmer alone, but also benefit the immediate
surrounding area and region as evidenced by:

- Reduced soil erosion

- Improved water quality in lakes, rivers and streams

- Less flooding due to improved water infiltration

- More stable water flows in rivers and streams due to improved water

infiltration and storage of water within the soil and water table

dramatically reducing soil erosion and runoff during rain events.
At even larger global level scales the impacts of conservation tillage can be
observed. Conservation tillage from a global perspective results in:

- Carbon sequestration in agricultural soils

- Reduced water pollution from both soil erosion and agricultural

chemicals,

- Recharge of aquifers through enhanced water soil infiltration rates

- Overall reduction in agricultural fuel use

The two primary tillage practices, encompassing such a large area
spatially, exhibit a great deal of influence on the environment. With agricultural
row crop production taking place over such a large expanse it seems the
research community would have a great deal of data regarding the spatial
distribution of cropping methods and acreage. This is not the case; fundamental
data regarding the spatial distribution of cropping methods does not exist. The
scaling up of field level research to larger regional areas allows for the
incorporation of results into larger modeling scenarios to better understand the

complex role man has in changing and shaping the environment. In order to

better understand the impacts and influences of tillage methods results from the

10

field level must be scaled to larger areas to accurately model or portray the

influences field level cropping practices have on the larger environment.

1.4 Research Needs

Clearly, to gain a better understanding of the environmental impacts of
row crop agriculture and the techniques used to raise these crops, a better
understanding of the spatial distribution of cropping methods is required. Remote
sensing data and analysis, properly applied, can provide an invaluable tool to
assess, map, and monitor agricultural areas over a variety of scales. The results
of such analysis will provide fundamental data regarding the spatial distribution

and total area of agricultural cropping methods.

To better understand the environmental implications of man’s influence on
the environment, including the carbon cycle, and especially the missing sinks,
more detailed data are required. Agricultural production encompasses a vast
area, yet little to no data exists regarding the spatial extent and distribution of
agricultural crops and the associated cropping methods used in production.
Knowledge of the spatial extent and distribution of agricultural tillage practices
will provide new insights into complex environmental issues including carbon
dynamics. Such data will add another piece to the carbon cycle conundrum,

including perhaps, some insight into the missing carbon sink.

11

n)

(n

C(

An additional requirement of this research is to ensure that the findings
are repeatable as well as accessible to a wide variety of researchers. The
methods used in this study attempt to offer an accurate and cost effective
alternative to current collection methods such as driving transects. The methods,
software, and data requirements for this study were selected with these

characteristics in mind.

1.5 Statement of Problem

While it is clear from the literature that conservation tillage techniques
exhibit many environmental and societal benefits, there are no data regarding the
spatial extent and location of conservation tillage acreage. In order to accurately
estimate carbon sequestration potentials, erosion reduction, and other
environmental benefits, conservation tillage acreage needs to be accurately
identified and mapped. Remote sensing provides the tools to identify and map
the spatial distribution of conservation tillage cropping acreage over local and
regional scales quickly, cost effectively, and by the use of easily repeatable
methods. Additionally, remote sensing data creates a unique and permanent
record that can be used to monitor conservation tillage over a variety of temporal

scales.

The spatial extent and location of conservation tillage acreage will allow

for the scaling up of previous studies to more accurately estimate the impacts

conservation tillage farming practices have on the environment. The successful

12

identification and mapping of conservation tillage acreage allows for examining
temporal change as well. Current results can be compared with past or future
data to monitor change in the spatial extent and location of conservation tillage

acreage.

The specific research objectives to be addressed by this study are to: 1)
Characterize the spectral properties of agricultural fields with crop residues,
(conservation tillage), and fields composed of bare soil (conventional tillage); 2)
Determine the most appropriate method of classification based on overall
classification accuracy and data availability; and 3) To estimate the extent

(acreage) of conventional tillage versus conservation tillage within the study area.

13

In

In

(n

n)

CHAPTER 2 LITERATURE REVIEW

2.1 Digital Remote Sensing

Remote sensing may be defined as techniques that utilize electromagnetic
radiation to detect and quantify physical, chemical and biological properties about
objects that are not in contact with the sensing apparatus (Jahne, 1997). Remote
sensing platforms commonly in use include digitally based aerial and satellite
systems designed to view the earth’s surface from an overhead perspective. The
resulting digital images consist of data (digital numbers) based on reflected or

emitted electromagnetic radiation from the surface of the earth.

The resulting digital image formation is a combination of information layers
consisting of digital numbers (DNs) representing the relative reflectance of
surface elements per spatial unit area, pixels. The data layers are based on the
available spectral bands of the remote sensing instrument and commonly include
spectral bands beyond the visible spectral range. The width of the spectral
bands, spectral resolution, describes the sensors sensitivity to particular
wavelengths, or more commonly, bands of wavelengths. For example, a near
infrared band may be sensitive to electromagnetic radiation ranging in
wavelength from 700 — 900 micrometers, not individual specific spectral

wavelengths such as 751 micrometers.

14

0c
er
IOI

all

Get
I93
anc

USE

Ont

Each layer or spectral band is comprised of individual pixels representing
a spatial area on the surface of the earth. Each individual pixel encompasses a
predetermined spatial area, commonly referred to as the spatial resolution, or
pixel size, of the data. An image product comprised of 30-meter by 30-meter
pixels means that each pixel covers a 30-meter by 30-meter spatial area on the
surface of the earth. The spatial resolution of image data is commonly referenced

to the nadir position, the point falling directly beneath the sensor.

Because both the spatial and spectral aspects of the data collection are
not infinitely small, both information layers or contents are products of averaging,
both spatial and spectral. For example, if an object on the surface of the earth
occupies a 5 meter by 5 meter area the corresponding 30 meter pixel
encompassing this area is termed a mixed pixel. The resulting reflectance value
for this particular object will be a combination of the reflectance from the object

and the surrounding 25 square meters of the earth’s surface.

A great deal of design work over the decades has resulted in more
detailed data. Digital satellite based information is now commonly available to the
research community in a rich assortment of varying spectral, spatial, radiometric,
and temporal resolutions. The decision as to which remote sensing instrument to
use for data collection must be carefully assessed based on the phenomenon to
be observed and the sensor’s ability to provide data on the phenomenon based

on the spatial, spectral, radiometric, and temporal resolutions of the sensor. Data

15

ct

3V

accessibility and cost are other commonly considered criteria when selecting a
remote sensing data product. If the data are not publicly available or the cost of
data are prohibitive, these factors may play an overriding role in the decision

process.

2.2 Spectral Characteristicleonsiderations

The successful differentiation between conservation tillage and
conventional tilled fields lies in the spectral separability of the two classes. With
only a few minutes of training, it is possible to differentiate the two tillage
methods with the naked eye from ground level. Conservation tillage fields contain
stubble and plant residue from the previous years crop on the soil surface,
whereas conventionally tilled fields do not. This fundamental difference between
the two cropping methods will be used to differentiate the cropping methods in

remotely sensed imagery.

There exists a fundamental lack of knowledge regarding the spectral
characteristics of crop residues under field conditions. Only a few articles are
available in the literature regarding the reflectance properties of standing crop
residues under field conditions. These articles focus on alternative methods such
as fluorescence and laser based technologies for detecting crop residues. The
research community has seemed to accept outright the findings of laboratory
studies concluding that differentiation between crop residues and soils based on

spectral reflectance in the visible to short wave infrared (commonly available

16

spectral bands on current and future sensors) portions of the electromagnetic

spectrum are unattainable.

2.3 Crop Residue Spectral Characteristics

Previous studies investigating the spectral reflectance properties of crop
residues are found in the literature, Gausman and Gerberrnan 1975, Daughtry et
al., 1995, Daughtry et al., 1996, and Nagler et al., 2000. These studies concluded
that across the visible, near infrared, and shortwave infrared portions of the
electromagnetic spectrum the spectral reflectance of soils and crop residues
were similar. Therefore, classification based on these findings would result in
ambiguous conclusions. The data indicate that crop residues generally exhibit
higher reflectance than bare soils. However, soils or crop residues could exhibit
higher or lower reflectance across the visible, near infrared and shortwave
infrared based on the complex interaction of many factors, including organic
matter, moisture, texture, iron oxide content, and surface roughness (Daughtry et

al,1996)

These studies, however, do not directly address the possibility of
classifying remotely sensed data under field conditions. These studies included
several shortcomings including sample size, residue condition and lighting
conditions. The sample sizes of these studies were unrealistically small,
measuring 6 — 20 cm in diameter. The crop residues were also cut and/or ground

to fit them into the sampling apparatus. The sample preparation and size are

17

both unrealistic representations of standing crop residue under actual agricultural
field conditions. Crop residues are a complex assemblage of stems, stalks, and
leaves which overlay and cast shadows, conditions which are not accurately

represented in the laboratory setting.

Furthermore, these studies only investigated each target individually. That
is, measurements of crop residues directly overlying soils were not presented.
This is of minor consequence, due to the fact that even if these studies had
investigated crop residues directly overlying soils the small sample sizes along
with the sample preparation (chopping and grinding) would not have been
representative of agricultural field conditions. Agricultural fields and associated
management practices take place at a scale where texture and shadowing play a

significant role in overall spectral reflectance.

An additional consideration affecting the spectral reflectance of crop
residues not taken into account with previous studies is the overall condition of
crop residues under field conditions, especially given the temporal requirements
of imagery acquisition. To identify conservation tillage from conventional tillage
techniques data collection must occur after field preparation in the spring. Crop
residues used in these previous studies indicate that recent crop residues were
used as opposed to crop residues under field conditions in the spring. From the
time of harvest to early spring, crop residues have been subjected to

environmental conditions such as wind, snow, and rain, as well as

18

decompositional processes. These environmental effects and decomposition
significantly weather the crop residues and reduce the overall reflectance as

compared to freshly cut crop residues found in the fall after harvest.

A recent article by Bricklemyer et al., (2002) investigated the use of
Landsat ETM+ imagery to document no-till and conventional tillage practices in
Montana. The article only addresses wheat cultivation and the research was
performed relying on a single producer and fields under the producer’s control.
The paper illustrates the need to produce low cost and efficient methods to map
and quantify tillage practices but is limited in scope and application. The article
illustrates the suitability of using Landsat ETM+ data to classify cropping
methods, albeit on a limited scale. The research demonstrates the importance of
assessing spectral response patterns of tillage methods under field conditions.
Bricklemyer successfully classified cropping methods, based on spectral
reflectance under field conditions, used in wheat production. The research
indicates that under field conditions wheat fields utilizing conservation tillage
methods exhibit lower DN (digital number) values, or less reflective. The
research provides an important example of the use of remotely sensed data to

capture data relating to carbon sequestration and cropping methods.

2.4 Soil Spectral Characteristics

The spectral characterization of soils underlying crop residues or

completely bare soils are important to characterize because they make up a

19

larg
earl
to It
481
soil
IGIIl
abs
dor

dor

I881
SDe

Site

2.5

large part of the overall observed reflectance from row crop targets, especially
early in the growing season. Stoner and Baumgardner (1981) conducted a study
to investigate the reflectance of soils. The study examined 246 soil series from
481 sites throughout 39 of the 48 contiguous states of the US consisting of 485
soil samples. They found that the 485 soil samples could be described by five
reflectance curves identified by curve shape and the presence or absence of
absorption bands. The five representative reflectance curves were a) Organic
dominated, b) Minimally altered, 0) Iron affected, d) Organic affected, e) Iron

dominated (Stoner and Baumgardner, 1981).

They found that all Vertisol soil samples and a majority of Mollisol soil
samples exhibited the organic dominated curve form. Alfisols, Entisols,
lnceptisols, Mollisols, Spodosols, and Ultisols under an aquic moisture regime
also exhibited an organic dominated or organic affected curve. Alfisols and
Ultisols with a humid moisture regime follow the iron affected curve form. This
research shows that while there are a great deal of soil types, when viewed
spectrally, this great diversity of soil types can be simplified to five representative

spectral curves.

2.5 Classification Methodologies

Image classification is the procedure of grouping pixels into spectral
clusters based on spectral characteristics (Strahler, 1990). Image classification,
i.e. categorization of pixels based on their spectral characteristics, is one of the

fundamental analysis techniques for remotely sensed data, with land cover

20

 

n)

mapping arguably being the most frequent application (Cihlar et al., 1998).
Classification of image pixels based on spectral response allows for the grouping
of pixels based on class membership, thus producing a new data product that
identifies pixels of similar values into unique classes. Classification routines are
commonly used in remote sensing studies to aid in image interpretation and are

a valuable tool for information extraction.

Digital image analysis, including classification routines, offers many
advantages over conventional interpretation, primarily due to the numeric nature
of digital data sets. The use of computers provides the ability to perform rapid
mathematical computations within or between spectral bands with little to no
computational error. An additional advantage is the ability to perform a series of
analyses, successful or othenivise, without altering the original data set. The

ability to discard or save results offers a powerful analytical tool in its own right.

Several classification routines exist for the classification of multispectral
imagery including maximum likelihood, minimum distance, and spectral angle
mapping classifiers. The classification routines fall into three main categories;
distance Based, probability based, and angular based decision rules. New
classification routines or, more commonly, modifications to existing classification
routines are frequently reported in the literature, an indication that there is no one
ideal classification routine to suit all needs and requirements (Kaminsky et

al.,1997, Mural and Omata 1997, Cortijo and Perez de la Blanca 1998, Cihlar et

21

8C

C0

CIa

IOU

al., 1998, Erol and Akdemiz 1998 Kartikeyan et al., 1998). Each classification
routine consists of a series of trade offs regarding processing time, model
complexity, and classification accuracy. The decision regarding the most suitable
classification technique is made on a case by case basis with factors such as
spectral seperability, number of classes to be identified, processing time, and

model complexity each playing a role in the decision process.

Geometric or distance based classifiers such as minimum distance rely
primarily on mean values, ignoring variance within classes. The primary
advantage of geometric processing routines lies in the relatively quick processing
time to perform the classification. In the recent past, during the widespread
adoption of desktop computing, when processing speed governed by the CPU
(central processing unit) and RAM (random access memory) were limited,
processing time was an important consideration when deciding on a classification
routine. The quickly advancing technology of both CPU processing speed and
availability of RAM, coupled with decreasing cost has negated this advantage.
The speed and memory of most common desktop computers are now able to run
advanced classification routines without the need for several hours or days of

computational time.

Statistical classification routines offer, in most cases, higher degrees of

classification accuracy over geometric classifiers. Statistical classification

routines incorporate both the mean and variance of the data set into the

22

classification decision rule. The utilization of variance into the classification
decision rule provides additional data on which to base the classification, thereby
improving overall classification accuracies. An illustration of the classification
decision rule for both statistical and geometric classification routines is provided

in Figure 1.

 

 

Band Y
Band Y

 

 

 

 

 

 

 

 

Band X Band ><

Minimum Distance

 

 

 

 

 

Band Y
Band Y

"'3

 

 

 

 

 

 

 

 

Band X Maximum Likelihood Band X

 

 

 

Figure 1. Classification Decision Space.

Image classification, or more specifically, digital image classification, uses
the spectral information contained in the remotely sensed data to transform the
data from digital numbers to information classes, a thematic data product. Data

regarding the spectral response across each band for each pixel is used to

23

classify pixels in the imagery based on spectral information. The resulting
classified image is a representation of the original data merged into classes, the
end result being a thematic map of the original image. The generated thematic
map provides a means of representing spatial information in meaningful

informational classes (land use/land cover).

The derivation of spectral classes is commonly performed using two
primary procedures, supervised and unsupervised classification routines.
Supervised classification techniques rely on the skill and knowledge of an analyst
to identify representative training sites of various surface cover types or
information classes. The basic steps to perform a supervised classification are:

-Determine ground cover type classes

-Establish training sites

-Signature collection of training sites

-Classification

-C|assification accuracy assessment

-Presentation of classification results
The training sites, areas of homogenous pixels representing a particular
informational class, are used to develop a spectral signature for each class. The
analyst supervises the classification by identifying training sites and
corresponding informational classes. Supervised classification techniques require
ground truth information and/or knowledge and familiarity with the geographic
area to accurately relate training sites with a particular land use/land cover type.

The spectral information for pixels contained in the training sites are then used to

develop spectral signatures for each training site. The software uses the spectral

24

tn

signatures derived from training sites to classify pixels in the remote sensing data

set based on a variety of classification decision rules specified by the analyst.

Unsupervised classification uses clustering algorithms to partition the
remote sensing data in multispectral space. These clustering algorithms are
based on numerical and statistical measures resulting in classes that are based
on natural breaks in the data, the basic steps for performing an unsupervised
classification are:

-Specify number of output classes

-Perform unsupervised classification

-Assign/label classes to a land use/land cover

-Merge or expand classes based on results

-Classification accuracy assessment

-Pre5entation of classification results
Clustering algorithms are used to aggregate groups of pixels with similar spectral
patterns into classes. These classes are only representative of class membership
and may or may not represent individual informational classes (land use/land
cover) on the ground. The classes are then examined by the analyst to determine
if a particular clustering represents an informational class of interest. The only
guidance or input to the classification scheme the analyst provides is the number
of output classes prior to the clustering operation. It is common practice to
specify a number of classes greater than the expected informational classes to
ensure adequate separation of informational classes. The classes identified by
the unsupervised classification are then examined by the analyst and associated

with land use/land cover types. The decision rule in both classification methods

can be statistically or geometrically based and the specific rules for assigning

25

In!

K)

IIkl

class membership can be explicitly stated in the classification routine or the

default values provided by the software may be used.

2.6 Classification Decision Rules
2.6.1 Minimum Distance

Minimum distance classification decision rules are based on simple
Euclidean distance. The mean value for each class is calculated and the
unclassified pixel is evaluated against these mean values. The unclassified pixel
is compared to the mean value of each class and assigned class membership

based on the closest mean class value, or minimum distance.

2.6.2 Mahalanobis Classification

The Mahalanobis classification decision rule bridges the gap between
simple Euclidean distance classification and probability based maximum
likelihood classification routines. The Mahalanobis classification decision rule
uses minimum distance as the main method of classification but also
incorporates a directional weighting component derived from the covariance

matrix based on a class average (Richards, Jia, 1999).

2.6.3 Maximum Likelihood
The maximum likelihood classification routine is generally regarded as a
standard to which other classification routines are compared. The maximum

likelihood classification routine is based on statistical probabilities. An unknown

26

pixel is compared to training sites statistically and assigned to class membership
based on probability theory. The maximum likelihood decision rule incorporates
statistical measures and probabilities utilizing the covariance matrix to assign
unclassified pixels to class membership. In order to sufficiently train the classifier
a statistically significant number of training pixels must be utilized to accurately
estimate the covariance matrix. To avoid a singular covariance matrix the
number of training pixels for each class must be at a minimum N+1 pixels large,
where N is the number of spectral bands. To further ensure accurate covariance
matrix estimation a commonly recommended minimum is 10*N pixels for each
class is desirable. Swain and Davis (1978) recommend a minimum of 10N pixels

for each class with 100M for each class being even more desirable.

The classification accuracy for maximum likelihood classification routines
depends heavily on the accurate estimation of the covariance matrix. To
accurately estimate the covariance matrix a sufficient number of pixels for each
class training site must be readily available. However, even if great care is taken
to ensure a statistically significant number of training pixels are used to generate
reference signatures, the underlying decision rule still requires an assumption of
data normality with well defined variances in each spectral band (Sohn and
Rebello, 2002). These assumptions are generally not met using multispectral
data sets; spectral bands obtained from remotely sensed imagery often are
skewed or exhibit non-normal distributions. Given the inherent limitations and

assumptions that must be met with maximum likelihood classification decision

27

rules
evld

clas

2.6.

spe
in n
Spe
dec
Spa

fete

the
he
Ber
Vec

alb.

Fete
her

mUI

rules in order to accurately estimate the variance and covariance matrix, it is
evident that spectrally similar reflectance patterns may well be inaccurately

classified.

2.6.4 Spectral Angle Mapping

Spectral angle mapping (SAM) is a classification decision rule based on
spectral angles formed between a reference spectrum and an unclassified pixel
in n-dimensional space where n is the number of spectral bands available.
Spectral angle mapping offers an innovative technique to classify data. The
decision rule is straightfonivard to understand; a vector is plotted in n-dimensional
space for an unknown pixel. The angle this vector forms with the vectors of
reference signatures is compared for each reference vector and the pixel is
assigned class membership to the smallest angle formed with respect to one of
the reference vectors. An advantage of this method is that SAM is relatively
insensitive to illumination and albedo effects (ENVI, RSI lnc., online help).
Because the vectors of both the unclassified pixel and the reference spectra are
vectors in n-dimensional space, a variation between pixels due to illumination or

albedo effects will have little effect on the vector.

The term spectral angle mapping is used loosely in the literature and may
refer to any one of several SAM variations. Some references to SAM in the
literature are referring to simple angular decision boundaries to partition

multispectral space. Other references to SAM refer to a classification routine

28

It

known as the cosine of the angle concept, a technique that builds upon SAM

principles.

2.6.5 Cosine of the Angle Concept

The cosine of the angle concept algorithm determines the spectral
similarity between spectra by calculating the angle between the vectors of the
spectra for each band in multi-dimensional space. The spectra are represented in
multidimensional space (determined by the number of bands) as vectors and the
angle between the reference spectra and target spectra are calculated similar to
the SAM technique. The cosine of the angle concept takes SAM a step further by
calculating the cosine of the angle formed between the vector representing the
unclassified pixel and the vectors representing the reference spectra (Figure 2).
The cosine of the angle concept effectively incorporates the length of the vectors
of the reference signature and the pixel to be identified by incorporating the
cosine of the angle. The cosine of the angle formed between the two vectors is
defined by the adjacent over the hypotenuse. Thus, the length of the vectors are
incorporated into the decision rule, adding additional information to the
classification decision rule over spectral angle mapping. Spectral angle mapping

uses only the angle formed between the vectors as the decision rule.

The angle that defines a spectral signature or class does not change and

the vectors forming the angle from the origin delineate and contain all possible

positions for the spectra (Sohn and Rebello, 2002). These parameters

29

9X
dis
IOL

The

E

encompass all the possible combinations of illumination for the spectra. The fact
that the spectra of the same type are approximately linearly scaled versions of
one another due to illumination and topographic variations is utilized to achieve
accurate classification results (Sohn et al., 1999). Cosine of the angle concept
relies on a normalized dot product, the cosine of the angle formed by the vectors
of the spectra, which can be utilized to overcome shortcomings of statistical and
geometric classification routines to successfully classify spectral groups (Sohn et

al., 1999).

 

nknown 1

"D
g»
8
'3’
Q9

 

 

 

Figure 2. Cosine of the Angle Concept Decision Space.

The spectral signatures of healthy green vegetation and soil provide an
example. The spectral signatures of the two targets, vegetation and soil, are
distinct enough to permit accurate classification with traditional classification
routines. Now compare the spectral signatures of soil and senescent vegetation.
The spectral signatures of the two targets are much more similar. It is possible,

due to illumination, topographic or soil moisture effects that variation between the

30

IC

III

f0
50
CI;

30;

two signatures may significantly overlap. If traditional classification routines,
which rely on means and variances are utilized, the two classes will be
misclassified in a number of instances. Cosine of the angle concept works
around the shortcomings of traditional classification routines by encompassing all
the possible illumination, topographic and soil moisture effects in a single class
(Sohn et al., 1999). The underlying signature as defined by the angle of the
vectors remains invariant; the vectors into multidimensional space form the
boundaries for all possible illumination and topographic effects (Sohn et al.,
1999). This method relies on the fact that the unique spectral signature as
defined by the cosine of the angle for each signature set does not change.
Changes in reflectance due to illumination effects are still within the class angle

only the magnitude of the vector changes (Sohn et al., 1999).

New data sources and sensors coupled with improved classification
routines show a great deal of promise in discrimination between conventionally
tilled and conservation tillage cropping techniques under field conditions. The
improved radiometric and spatial resolution of Landsat 7 ETM+ data provide a
more detailed data source than previously available to the research community
for large scale agricultural mapping studies (Goward et al., 2001). Improved data
sources together with enhanced classification routines provide the basis for
classifying previously difficult to discriminate spectrally similar targets such as

soils and crop residues.

31

2.7 Accuracy Assessment

The confusion matrix has been identified as a best practice standard for
assessing overall accuracy as well as identifying errors of omission and
commission (Foody, 2002). Recent literature suggests a commonly
recommended target of 85% classification accuracy (Foody, 2002). The United
States Geological Survey (USGS) has also set a guideline of 85% classification

accuracy for remotely sensed data products.

The confusion matrix, or error matrix, is a representation of classification
accuracy with the rows representing the findings of the classification and the
columns containing the reference data. The error matrix, if all goes well, is read
down the diagonal: the column and row totals for each class should match
identically, if not, a classification error exists. This method allows for the
identification of both errors of inclusion (commission errors) and errors of

exclusion (omission errors).

The practice of using training site pixels in classification accuracy
assessment is one that should be avoided. Campbell (1987) wrote regarding the
use of training pixels as classification accuracy assessment that the “biased
procedure is born of expediency and can have little use in any serious attempt at
accuracy assessment.” Pixels selected for classification signature collection

should not be used again in the classification accuracy assessment. A better

32

method is to compare classified pixels with known ground truth data points that

were not used in classification signature collection.

33

CHAPTER 3 MAPPING TILLAGE PRACTICES

In principle, land cover mapping from satellite data is straightforward and
consists of four steps: data acquisition, pre-processing, analysis/classification
and product generation (Cihlar, 2000). In addition to the aforementioned
procedures, a knowledge and familiarity with the landscape and management
practices operating across the landscape also provide a great deal of information
to be used in the overall analysis procedure. A diagram providing an overview of
the methods is provided in Figure 3. This section provides a description of the

study area and the methods utilized in data analysis and extraction.

Field Based .
Spectral Acquisitio

Lab Based Processing Transect
Spectral Acquisition Classification Data

- +
Satellite Based
Spectral Acquisition

 

 

 

 

 

 

 

 

 

Classified
Data

I

Map of
Tillage Practices d
-Results-

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 3. Diagram of Methods

34

th

M

Inc

ac

3.1 Study Area

The study area is delimited by Landsat path 21 row 31 an area that
encompasses a 180 km by 180 km area of southwestern Michigan, northern
Indiana, and a section of northwestern Ohio (Figure 4). The study area is
primarily dominated by agricultural land use with intermixed forested areas. The
study area also encompasses urban and suburban land uses as well but spatially
these areas are small in comparison to the surrounding countryside. A total of
4,356,644 acres of farmland lie within the 22 counties encompassed by the study
area. A cross section of the study area from Fort Wayne, Indiana to Lansing,
Michigan is representative of the region; the landscape is dominated by row
crops. Cropping systems in the area are typical of the US. corn belt, comprised
mainly of corn/soybean acreage. Alfalfa and winter wheat are also produced
throughout the region but spatially account for very little of the overall landscape.
The 22 counties encompassed by the P21 R31 swath collectively produced
$1 ,702,236,000 in agricultural sales for the year 2002 (Table 1) predominantly
through corn and soybean production (USDA, 2002). The Great Lakes and
Midwestern regions of the United States are some of the most productive areas
in the United States for agriculture production. Agriculture has been a dominant
industry in the region since settlement occurred with many of the farms under

active cultivation for over one hundred years.

35

 

Figure 4. Study Area Location.

Landsat 7 P21 R31

36

 

 

 

 

County State Acrgqe $ Production $/Acre

Paulding OH 209,983 $54,144,000 $257.85
Defiance OH 186,385 $45,248,000 $242.77
Williams OH 203,201 $51,383,000 $252.87
Steuben IN 123,953 $25,641,000 $206.86
Lagrange IN 189,932 $103,278,000 $543.76
Elkhart IN 182,771 $124,038,000 $678.65
St Joseph IN 154,142 $55,178,000 $357.97
Marshall IN 201,637 $62,187,000 $308.41
Kosciusko IN 246,907 $146,062,000 $591.57
Noble IN 181,963 $58,841,000 $323.37
Dekalb IN 162,936 $38,669,000 $237.33
Allen IN 276,385 $89,877,000 $325.19
Whitley IN 165,067 $51,930,000 $314.60
Allegan MI 236,936 $186,757,000 $788.22
Van Buren Ml 177,360 $100,641,000 $567.44
Kalamazoo MI 146,927 $105,494,000 $718.00
Calhoun MI 243,151 $60,985,000 $250.81
Jackson MI 181,287 $44,311,000 $244.42
Hillsdale MI 257,469 $71,729,000 $278.59
Branch MI 234,076 $77,249,000 $330.02
St Joseph MI 217,345 $81,103,000 $373.15
Cass MI 176,831 $67,491,000 $381.67
Total 4,356,644 $1 ,702,236,000 $390.72

 

 

Table 1. Study Area Acreage and Production Value. Source: USDA (2002).

3.2 Study Area Physical Characteristics

Annual rainfall measured at the Kellogg Biological Station averages 890 mm
y'1 (54) (LTER). Potential evapotranspiration (PET) exceeds precipitation during
the summer months with a maximum generally observed during the month of
July. Mean annual temperature is 9.7 °C, (Figure 6) and the average growing
season is approximately 140 — 160 days per year. The temperature and
precipitation of the region are well suited for agricultural production. Within the

Koeppen climatic classification system the climate of the region is characterized

37

as moist continental mid-latitude with cold winters. These climatic conditions are

well suited for corn, soybean and cover crop production.

 

Water Volu
ceases

 

 

w I T T 1' T I“ T l T w
F M A M J J A S O N 0
Month

“-6)

[j precipitation 0 potential evapotranspiration

 

 

 

Figure 5. Study Area Precipitation and Potential Evapotranspiration.
Source: KBS LTER

 

40“
d
1

A
AA AAA

     

6
L
93
B
‘3
a)
o.
E
o
'—
2
§
2

 

 

. Mean Monthly Minimum 0 Record Daily Minimum
A Mean Monthly Maximum A Record Daily Maximum

 

 

 

Figure 6. Study Area Yearly Temperature Graph.
Source: KBS LTER

38

DI
We
Ch
are
de,

Cor

The study area landforms and soils have been extensively shaped by
glacial processes. The last major glacial activity occurred during the Wisconsinan
glaciation of the Pleistocene epoch approximately 13000 to 16000 years B.P.
The study area encompasses a diverse, glacially-formed landscape including
such features as outwash plains, ground and end moraines, and ice contact

features such as kames and kettles (USGS, 1998).

Soils in the area developed on glacial till, and include well- and poorly-drained
alfisols, mollisols, and entisols. Most regional soils are sandy loam and silty
clayey loams of moderate fertility (KBS LTER). The major soil associations of the
study area are predominantly alfisols of which udalfs are the dominant suborder
(USDA, 2000). Loamy sands found primarily in outwash plains and sandy loams
or loamy sands commonly associated with end and ground moraines (USGS,

1998)

The soils of the moraines are typically well and excessively well drained.
Drainage conditions on the outwash are more variable, ranging from excessively
well drained to very poorly drained. Thick outwash deposits are usually
characterized by excessively well drained conditions. Shallow outwash deposits
are underlain in some places by bedrock or fine textured till and lacustrine
deposits, causing poor or very poor drainage conditions (USGS, 1988). On ice

contact topography, soils are typically excessively drained on the upland kames

39

 

and

draj
are;

Mar

3ng

and eskers and poorly or very poorly drained in the kettles and outwash channels

(USGS, 1988).

Soil textures range from sand to clay; the most common soil texture is
sandy loam on the moraine ridges and sand on the outwash plains. The glacial
drift that forms the moraines is largely derived from the local limestone bedrock.
In the ice contact areas, soils are sands and gravels, which are well drained and

not well suited for agricultural production.

Very poorly drained soils are common in the narrow outwash channels
between drumlins and in ice-block kettles or abandoned stream channels
(USGS, 1998). Poorly drained areas combined with significant inputs of organic
materials can lead to peat accumulations of up to 6 to 10 feet thick. Soils on the
end moraine and ground moraine are typically sandy loam or loamy sand, and
most are well drained. A well-developed argillic horizon is common in these soils

(USGS, 1998)

Poorly drained soils, if not reclaimed for agricultural production through
drainage activities, are generally forested or remain as wetlands or swampy
areas. Upland areas which tend to be well drained are dominated by agriculture.
Many of the historical wetland areas have been drained and converted to

agricultural production.

Soil erosion across the study area is well below the national average due

to a maximum average slope of 3.2 feet/mile (USGS, 2000). The small slopes

40

across the study area result in relatively small losses of soil from agricultural
lands of 1.0 - 5.0 tons/acre/year compared to other reports of 21 tons/acre/year
for a national average of erosional soil loss from agricultural lands (Aber, Melillo,
2001). For comparison, the estimated erosional soil loss from forested areas is

estimated at 0.001 to 0.06 tons/acre (Kimberlin and Moldenhauer, 1977).

3.4 Study Area Agriculture

Agricultural production across the study area, and the Midwestern United
States, is primarily comprised of corn and soybean rotations with lesser amounts
of secondary crops such as alfalfa and winter wheat. Corn and soybeans are
commonly grown on the same acreage alternating yearly. The practice of yearly
crop rotation between corn and soybeans is primarily driven by the need to
maintain soil fertility. Corn and soybeans are also readily marketable, ensuring
that both crops are widely produced. Corn and soybeans are used not only in
food production, both human and livestock, but also in a wide variety of corn and

soybean based products.

Following settlement the study area was systematically logged, cleared,
and in many cases drained to enable agricultural production. Historically crop
production throughout the study are encompassed many crop types grown on
small family owned and operated farms. Crops included corn, alfalfa, oats, wheat
and other small grain varieties with the vast majority of acreage used for small

grain and alfalfa production (USDA, 1993). Until mechanized machinery was

41

widely available, shortly after WWII, most crop acreage was produced in

relatively small fenced fields consisting of 5 — 10 acres (USDA, 1993).

Widespread adaptation of mechanized equipment used in agricultural
production resulted in a significant change in crops, field size, and environmental
impacts. As field sizes increased, tree lined fencerows that once separated
individual fields and provided a measure of erosion control were lost. Large
mechanized machinery resulted in a dramatic change in the amount of land that
could be worked and also led to an economy of scales. Larger fields enabled
farmers to make long uninterrupted passes through a field with implement and
tractor, fewer turns and maneuvering of machinery results in less time working
the land. Larger individual field sizes coupled with larger more powerful and
efficient equipment resulted in larger farms (more acreage) even though the total

number of farms was decreasing (Baker et al., 1998).

The shift in agricultural crops from small grains and alfalfa raised on small
fields to the widespread production of corn and soybeans on larger individual
fields has created a variety of environmental impacts. Small grains and alfalfa are
planted in very close proximity of one another, that is the spacing of plants and
overall density of plants is very tight and compact. The close plant spacing and
tight canopy provide protection from erosional processes such as wind and water
action. This is in marked contrast to corn and to a lesser degree soybeans.

Individual corn plants are planted at a spacing of 10” — 14” apart from one

42

 

In

In

another along a row and rows are generally 20” — 30" apart, leaving a large
portion of the soil exposed to erosional action. Even at full canopy development
the soil beneath a corn crop is bare and the only protection from erosional action
is the stalk and root system. Soybeans also leave a bare soil surface exposed to
the elements under a fully developed canopy. The row spacing of soybeans is
closer than corn, generally 18” —24", but still leaves a portion of unprotected soil

until full canopy development.

In addition to erosional losses due to a shift in cropping methods, crops
and field size associated with larger scale mechanized farming methods. Before
the widespread adaptation of mechanized farming and implements, carbon
implications associated with agricultural crop production were less severe than
presently observed with conventional tillage methods. Smaller field sizes, which
were required due to the amount of work that could be performed by an individual
farming family with manual labor and livestock, meant that less overall area was
impacted by farming during settlement and before widespread availability of
mechanized farming equipment. Additionally, due to what some would term the
inefficiency of horse drawn implements and manual labor a great deal of crop
residue was left on the soil surface following harvest. Even if the stalks of small
grain crops were harvested for straw the machinery of the day paled in
comparison to the efficiency of modern day agricultural machinery and left a
great deal of plant residue in place. Plowing was another aspect of farming that

has undergone significant change following WWII.

43

Before mechanized farm machinery was available plowing a field
consisted of either manual labor or livestock assisted plowing methods. Both
methods were time consuming and labor intensive work, limiting the amount of
soil that could be worked in a single year by a single farm. Plows designed to be
pulled with tractors quickly changed the amount of acreage that could be worked
by an individual or family farm. Tractors have become more powerful and

efficient over the years leading to larger and larger field sizes.

In addition to market demands, corn and soybeans are widely raised on
an alternating yearly basis for the benefits associated with soil fertility and long
term production capacities. Corn production is nitrogen intensive, a nutrient that
can be quickly diminished in topsoil if corn production is continuous on a
particular piece of land. Corn yield is closely related to soil nitrogen availability
and farmers routinely add nitrogen fertilizer to corn acreage on a yearly basis to
increase yield. To offset or minimize the loss of soil nitrogen due to corn
production farmers rotate crops between corn and soybeans on a yearly basis,
one-year producing com the next soybeans. The advantage of using a rotating
corn/soybean crop system is that soybeans effectively fix nitrogen from the
atmosphere and sequesters nitrogen in the soil. The added nitrogen from
soybean production from the previous years crop adds to soil fertility and

provides nitrogen so critical to achieving profitable corn yields.

44

pr

th

I6

39

CIIr
Il‘ve

fen

Winter wheat and alfalfa are two commonly occurring secondary
agricultural crops produced across the study area. Winter wheat and alfalfa
production occurs at a much smaller scale than corn or soybean production
primarily due to market demands and crop values. Winter wheat production does
not provide as much income on a per acre basis as compared to corn and
soybean production so the amount of winter wheat production is correspondingly
much lower throughout the study area. The majority of winter wheat production
has shifted to the great plains area where the soils are not of suitable quality for
corn and soybean production but are quite adequate for winter wheat production.
As the value of corn and soybeans has risen over the past decades agricultural
production has shifted to corn and soybean production on soils that will support
intensive row crop agriculture. The soils found on agricultural lands throughout
the study area are of sufficient quality to support corn and soybean production

resulting in less land used for winter wheat production.

Alfalfa is an additional secondary crop found throughout the study area but
again, as with winter wheat, does not provide the economic incentive for farmers
to raise the crop over large areas. Alfalfa production thrives in the soils and
climate found throughout the study area but is primarily used as feed for
livestock. The production value of corn and soybeans exceeds the economic
return for grazing and or feeding livestock and this is reflected in the land
use/land cover of the study area. Alfalfa production that does occur within the

study area is generally on smaller acreage fields in support of small family owned

45

and operated farms that use the alfalfa production to raise a small number of
livestock owned on the farm. Alfalfa is a crop in which a large portion of the entire
above ground biomass is harvested, not just the seed or grain as in corn,
soybean, and winter wheat production. Alfalfa, when harvested is baled, creating
a very heavy and bulky commodity. A commodity that is very expensive to
transport as opposed to corn, soybean and wheat production which is easily
handled and of which only a small portion of the harvested plant is used. Alfalfa
on the other hand is difficult to handle, expensive to transport due to the weight
and size of the bales and also relatively expensive to store due to the size and
bulk of bales. Additionally, alfalfa does not store or keep well for extended
periods of time. These limitations make for a crop that has limited marketability
and value, with farmers generally only producing limited amounts for personal on

farm use or for sale to the immediate surrounding area.

3.5 Ground Truth Data Collection

Several large farms employing conservation tillage techniques have been
identified through personal correspondence with agricultural field extension
agents in the Landsat 7 P21 R31 study area. These farms will serve as
classification accuracy assessment checks for image processing of Landsat 7
data. Driving transects were established to include these large farms. A subset of
these fields were used as end members for spectral signature collection while the

majority will be used as classification accuracy check of the classified image.

46

(I)

II
It

pl

be

3.6

21:
Th.
nm

ag

UIII
IeS

Me

The KBS LTER site provides field access to a variety of crops and
cropping methods and provides full access to agronomic logs detailing the
methods and dates of fieldwork. While the K83 LTER site manages the largest
repetitive experimental agriculture plots in the country, they are too small (100 m
by 100 m)to be used in a study relying upon Landsat 7 data (30 m by 30 m)pixel
size). The KBS LTER plots provide access for hand-held radiometric reflectance
readings, as well as a knowledgeable staff to provide assistance and advice
relating to cropping systems. Making use of both the LTER site and larger
privately held fields ensures field access and provides a target large enough to

be used with Landsat 7 data.

3.6 Laboratory Radiometric Sampling

Laboratory radiometer measurements were performed using an Analytical
Spectral Devices, Field Spec Pro. The Field Spec Pro collects spectral data in
2134 channels between 350 nm to 2500 nm of the electromagnetic spectrum.
The spectral resolution from 350 nm to 1050 nm is 3 nm and 10 nm for the 1050
nm to 2500 nm portion of the spectrum. This level of spectral resolution provides

a great amount of detail that has not been available until recently.

Laboratory measurements were taken from a height of 50 centimeters
utilizing a 24 degree fore optic resulting in a field of view (FOV) of 21 cm. Crop
residues of soybean and corn stubble were examined as well as bare soil.

Measurements were taken under a variety of conditions including crop residue

47

alone and crop residue directly overlying soil. Measurements were also taken of
samples consisting only of soil. Laboratory measurements were taken under dry
conditions, both the soil samples and crop residues were allowed to air dry for 48

hours before radiometric measurements were taken.

Radiometric sampling was performed using a direct overhead position,
commonly referred to as NADIR. The illumination source consisted of laboratory
grade halogen light source operated with a DC inverter. The DC inverter ensures
any electrical fluctuations in current or voltage common with AC electrical

sources are suppressed to minimize any potential fluctuations in illumination.

3.7 Field Based Radiometric Sampling

A backpack mounted hyper-spectral radiometer (Analytical Spectral
Devices, Field Spec Pro) was used to collect spectral data of conservation tillage
and conventionally tilled crops in situ. The data provided valuable information
regarding the spectral response of the two cropping methods under field
conditions. These data provide not only ground truth data but also a great deal of
data regarding the placement of spectral windows for future remote sensing
platforms. Additionally, field spec pro data can be utilized in the future to create
classification end member inputs for a variety of other remote sensing platforms.
The field spec pro data can be resampled and aggregated to approximate the

spectral resolution of a wide variety of remote sensing platforms.

48

3.I

wilt

QIOL

trans

The Field Spec Pro offers a variety of fore optics for controlling and
adjusting the field of view (FOV). To investigate the issues of scaling and
reflectance, spectral reflectance measurements were made using varying fore
optics (8 and 24 degree) while holding the height of the instrument constant
above the field surface at a height of two meters. Using the 8 degree fore optic at
a height of 2 meters results in a FOV of 28 cm, a 24 degree fore optic at the
same height results in a FOV of 85 cm. Overall field conditions were relatively dry
with no noticeable moisture present. Measurements were taken in the principal
plane at the NADIR position. Field measurements were calibrated using a
spectralon panel and converted to reflectance utilizing readings of the calibrated

reflectance panel.

3.8 Ground Truthing - Driving Transects

Driving transects consisting of approximately 960 miles were established
within the Landsat 7 P21 R31 swath (Figure 7). These transects provided the
ground truth data about crop type and cropping method. Along the driving
transects agricultural fields were identified and recorded using a GPS video
overlay system. This system incorporates a video camera and a GPS unit to
record fields along the driving transect and overlays the GPS positional
coordinates directly onto the videotape. This system provided digital ground truth
data incorporating position, date, time and a video recording of the field condition
observed during the driving transect. These transect data also provided ground

truth information for the classification accuracy assessment. A definitive accuracy

49

assessment requires the use of ground truth as part of the sampling design

(Magnussen, 1997).

Driving transects were established using Landsat 7 P21 R31 ETM+
imagery from June 6th 2001 and input from agricultural extension agents.
Imagery from 2001 was examined to identify areas within the scene that consist
primarily of agricultural land use. Additionally, transects were developed to
include areas identified by extension agents as places where conservation tillage
cropping has historically taken place. Through personal conversation with these
agents, many indicated that 50% - 80% of all agricultural fields in southwestern
Michigan are cropped using conservation tillage practices. Using these two
methods -- examining last years imagery to identify agricultural areas coupled
with input from agriculture extension agents ensured that the driving transects

sampled the appropriate areas.

50

 

|:] Indiana

,m Michigan

 

Ohio

 

Figure 7. Driving Transect Data Points.

Collecting ground truth data for use in conjunction with remotely sensed
data sets is commonly practiced. Ground truth data provide valuable information
to assess classification results as well as ensuring training sites for supervised
classification are accurately identified. The collection of ground truth data points
is often limited by time. Ground truth data should ideally be collected as close to

the time of imagery acquisition as possible, especially in areas with seasonal

51

temporal change such as in agricultural settings. This study relies on the coupling
of technologies to collect ground truth data quickly and accurately over large

areas.

The ground truth data collection system employed for the driving transect
portion of this study uses commonly available collection devices in a new and
innovative manner to permanently collect and record ground truth data. The
system consists of a GPS receiver, video camera, video recorder, and a text
overlay system to superimpose GPS positional data directly onto the videotape in
real time (Figure 8). The system is mounted in an automobile and driven over
pre-determined driving transects across the study area. The system is entirely
contained in the automobile and runs off the vehicle’s electrical system ensuring
continuous data collection. This method of collecting ground truth data could

potentially be used anywhere accessible to automobiles or off road vehicles.

 

 

 

 

 

 

 

GPS
Receiver
Text Video I Video
Overlay Recorder

 

 

 

 

 

 

I

Video Camera

 

 

 

 

 

 

 

Figure 8. Ground Truth Data Collection Flow Chart.

52

A Garmin GPS 53 was used to collect positional data for the driving
transect. The Garmin GPS 53 utilizes the WAAS (Wide Area Augmentation
System) technology to provide a real time positional accuracy of 3.2 meters.
More than sufficient to locate and identify agricultural fields within Landsat 7

imagery which consists of 30m X 30m pixels (Wilson, 2001).

The Garmin GPS 53 was coupled with a Sony digital video camera (Sony
Model DCR-PC9), a Sea-Trak GPS video overlay unit and a Panasonic S-VHS
video recorder (Model AG-7400). Together this system creates video imagery
with the real-time GPS positions recorded directly onto a S-VHS videotape. The
system was mounted in a motor vehicle and utilized throughout the driving
transect survey to accurately and permanently record ground truth of field

conditions and position.

In addition to continuous recording of the driving transects, utilizing the
GPS video overlay system, waypoints were also regularly collected. Fields of
interest, large agricultural targets with clearly discernable crops and cropping
methods, were routinely recorded by collecting a GPS waypoint. These
waypoints were downloaded from the GPS unit (UTM, WGSB4) and imported into
ArcView for analysis and display. The waypoints were converted to a point

shapefile and overlaid directly onto the P21R31 Landsat 7 ETM+ imagery. The

53

point shapefile was then used to create a polygon shapefile of the field of interest

corresponding to the appropriate waypoint by on-screen digitizing.

Four hundred and eighty (480) fields were logged as waypoints during the
transect. These 480 fields were used primarily as classification accuracy
assessments but also provided known ground truth points to collect spectral
signatures from the Landsat 7 imagery to be used as input in the supervised
classification routines. Only a small subset of the 480 logged fields, thirty eight,
were used to collect spectral signatures, ensuring that a significant number of

logged fields were available for use as classification accuracy checks.

Traditional driving transects performed to collect data on crops and
cropping methods only provide rough estimates of total field size. The area of
fields are estimated through observation and are not directly measured. Data
obtained from classified imagery on the other hand not only reflects the
occurrence of a land use/land cover, but also provides data regarding the total
land area of each class. Corn and soybean fields tend to be much larger spatially
than either winter wheat or alfalfa, a factor that may not be accurately accounted

for with traditional driving transect data collection methods.

Over the recent past, since no till machinery has become widely available,

it has tended to be the larger farm operators who have embraced no till

techniques rather than the smaller family farms. This is due, in large part, to the

54

large capitol outlay required to obtain a specialized conservation tillage seed drill
which is often cited by small operators as a financial barrier to adopting
conservation tillage techniques. A related reason is economy of scale. The time
required to recoup the costs of a conservation tillage seed drill is significantly
extended on a small farm of limited acreage. There also exists a
social/behavioral component with many small family operated farms that values
time honored techniques and traditions that are passed down from generation to

generation.

These examples illustrate the clear advantage of utilizing remotely sensed
data for agricultural crop, tillage, and acreage assessment over relying solely on
driving transect data. Driving transects only cover a small portion of an area, the
results of the driving transect are then scaled up to estimate the total area for
each crop/cropping method in the study area. Some driving transects make an
attempt to quantify field acreage during the driving transect, but they fail to
account for variation of field sizes throughout the study area. Remote sensing
imagery, with its overhead vantage point, is able to capture large areas
essentially in a single moment in time. This ensures that errors often made from
ground level, including inaccurate estimation of field size and limited distance of

sight, are avoided.

Another shortcoming of traditional driving transects is that they are not

well suited for monitoring yearly temporal change associated with crop rotations.

55

Driving transects are not regularly performed on a yearly basis due to budget and
time constraints. Agricultural extension agents, as well as the vast majority of
people involved in agricultural production, are extremely busy with planting and
field preparation activities during the time driving transects would have to be
performed to characterize tillage methods. The lack of annual data fails to
accurately capture the temporal change involved with agricultural crop rotations

among fields.

Remotely sensed data, on the other hand, are well suited for monitoring
temporal change. Satellite imagery is routinely collected at the beginning of each
growing season. Coupling remotely sensed data with a driving transect to record
ground truth capitalizes on the inherent advantages of both methods. In the case
of Landsat 7 data there isn’t even a need to pre-order imagery for specific areas
since they are collected automatically on a set schedule and can be obtained

well after the data are collected.

3.9 Temporal Considerations

Remote sensing imagery acquisition creates a data set unique in time and
space, a process analogous to a photographic snapshot. The temporal resolution
of the remote sensing platform must be of sufficient resolution to capture the
target of interest. Agricultural row crops also have a temporal component,
production follows a seasonal cycle and the timing of data acquisition must be

taken into account when examining the data. Field preparation takes place prior

56

as

to

we
me
pla
din
tim
the
is c
57
Get
COT

tilla

to planting for conventionally tilled crops and the timing of data acquisition must
fall within a specific temporal window to effectively capture the phenomenon of

interest.

The temporal resolution of the remote sensing platform and the temporal
window of the target of interest must be examined together to determine the best
period of opportunity for capturing data. The remotely sensed imagery must be
collected at a time when conventionally tilled crops have been plowed, but not so
late in the growing season that the crop canopy has developed to such an extent
as to cover the soil surface when viewed from above. Field conditions vary year
to year, and farmers must wait until fields are sufficiently dry to support the
weight of the tractor and implements. Farmers utilizing conventional tillage
methods will begin plowing as soon as possible in the spring to prepare fields for
planting. With plowing completed, planting will begin once the chance of frost has
diminished to an acceptable level of risk. Following planting, there is a period of
time before plants break the surface, usually 2 - 3 weeks. Once the crop breaks
the surface and begins growing, there is another 3-4 week period when the plant
is only 1 to 4 inches in height with no discernable leaf development. During these
5-7 week periods imagery must be obtained, after plowing, but before crop
development. During this time, plowed conventionally cropped fields will be
comprised of smooth bare soil with little to no crop stubble, while conservation

tillage fields will be covered in crop residue from the previous years crop.

57

Discussions with agricultural field agents and farm managers confirm that
all spring tillage activities were completed prior to imagery/data acquisition.
Plowing, if it were to take place, would have been completed in advance of the
date of data acquisition, May 27th 2002. Farmers at this point would have been
well into, if not past, planting crops for the growing season. This was also
confirmed during driving transects. Planting activities were observed during the
driving transects, but no plowing was observed during this portion of this study.
This indicates that the timing of data acquisition was well placed in terms of

capturing the phenomenon of interest.

3.10 Imagery Acquisition

Landsat 7 ETM+ imagery was selected as the primary data source for a
variety of reasons. Given the wide variety of sensors available for data collection
the ETM+ sensor offers the best combination of temporal, spectral, spatial and
radiometric resolution currently available for this study of any widely available
and accessible remote sensing data set. Agricultural practices vary over a broad
range of spatial scales and temporal frequencies; the data source must be of
sufficient spatial and temporal resolution to capture these phenomenon and

agricultural practices.
The ETM+ sensor onboard the Landsat 7 satellite offers a temporal

resolution well suited to agricultural remote sensing applications with a repeat

cycle of 16 days. The 16-day repeat cycle is sufficient to detect and monitor

58

agricultural practices, as well as the phenological development of agricultural
crops which take place over an entire growing season. The repeat cycle is also
sufficiently frequent, that if a data collection opportunity is missed due to
unfavorable atmospheric conditions, common in the spring months of the
Midwest, the next pass (16 days) later will provide another opportunity for data

collection.

ETM+ data are acquired every 16 days over the P21 R31 nominal center,
as controlled by the Landsat 7 satellite orbit. Imagery need not be ordered in
advance, the scene is collected and processed irrespective of standing orders.
From an operational standpoint, Landsat 7 data are reliably collected as
scheduled. The primary source of uncertainty associated with collecting the

imagery is cloud conditions.

The spatial resolution of ETM+ (30 meters) data is an optimal resolution
for monitoring agriculture fields across the study area which vary in size from a
few acres to hundreds of acres. Thirty-meter resolution is adequate to monitor
smaller fields while still permitting the capture of a relatively large swath width of

180 X 180 km.

The radiometric resolution of ETM+ data is also well suited to a study
investigating spectrally similar targets. The ETM+ sensor provides data in 8 bit

(28) radiometric resolution, an important consideration when examining spectrally

59

similar targets. Eight bit radiometric resolution provides 256 discrete distinctions
of energy levels received at the sensor. When reflected energy is received at the
Optical apparatus of the sensor and then transferred to the focal plane, charge
coupled devices (CCD’S), through an analog to digital converter, assign a digital
number of 0-255 to each pixel corresponding to the electrical charge generated
by the analog to digital converter. The ability to discriminate between 256 energy
levels means that small differences in the reflected energy received at the sensor
are distinguishable from one another. If a system were used with less radiometric
resolution such as a four bit system (24) only 16 discrete energy levels would be
available. Lower radiometric resolution results in a diminished ability to

discriminate spectrally similar targets.

3.11 Imagery Accessibility Considerations

The sensor characteristics that make Landsat 7 well suited for this study
also play an important role in making this research available and repeatable to a
wide variety of users. Landsat 7 imagery does not need to be ordered prior to its
acquisition, the scenes are collected for every path and row every 16 days over
North America according to the Long Term Acquisition Plan (LPSD, 1998). Other
sensors only collect data over a specific area if the data are ordered in advance.
Additionally, Landsat 7 data are archived by the EROS data center and are
available for order at a later date. The temporal resolution of 16 days is
sufficiently narrow to ensure 2 to 3 overpasses of an area at a time where the

findings of this study may be applied to other areas. The cost of Landsat data is

60

also an important aspect to ensure that this research is repeatable by others
across the Midwestern United States. Each Landsat 7 scene can be obtained for
approximately $600, a cost far below the expense of time and mileage to
complete a limited driving transect. These are important considerations if the
results of this study are to be expanded to other areas. The widespread
availability, coverage, and cost are such that it will be possible for others to

duplicate the results of this study over other areas.

3.12 Study Area Imagery

Landsat 7 ETM+ imagery for path 21 row 31 was obtained for May 27‘“,
2002 (Figure 9). There were four opportunities for Landsat 7 to acquire imagery
of nominal center P21 R31 in the time window during 2002: April 25, May 11,
May 27 and June 12. Imagery was delivered in Level 1G format, geometrically
and radiometrically calibrated. The ETM+ imagery contains 8 spectral bands
(Table 2), six of which were utilized in this study. The thermal band (band 6) and
the panchromatic band (band 8) were not included in the analysis portion of this
study. These two bands are of differing spatial resolution, and do not provide

additional information useful in this study.

61

 

Figure 9. Landsat 7 P21 R31 Band 5 May 27‘", 2002.

Enhanced Thematic

 

 

Mapper (ETM+)
Band Wavelength (um) Resolution (m)
1 0450-0515 (blue) 30
2 0525-0605 (green) 30
3 0630-0690 (red) 30
4 0750-0900 (near-IR) 30
5 1.55-1.75 (mid-IR) 30
6 10.4-12.5 (thermal-IR) 60
7 2.08-2.35 (mid-IR) 30
PAN 0.52-0.90 15

 

 

 

Table 2. ETM+ Sensor Characteristics.

62

3.13 Imagery Processing

Imagery pre-processing includes several steps to prepare raw satellite
imagery for analysis. The conversion of raw digital number (DNs) to radiance and
reflectance values as well as atmospheric correction and geometric registration
techniques are the main pre-processing preparations. Digital sensors record a
DN value for each pixel for each band. The assigned DN depends on the
intensity of the electromagnetic radiation from each spot viewed on the surface of
the earth (a pixel). The range of DNs that can be assigned to a particular pixel
depends on the radiometric resolution of the digital imaging system. Radiometric
resolution of imaging systems is commonly referred to as the bit value, for
example, 8 bit radiometric resolution data (28) provides 256 gray scale levels

(DNs) for each band.

3.14 Radiance

Digital numbers or brightness values assigned to each pixel are scene
dependent; that is, the same ground feature could exhibit different DN values
from scene to scene. The variance of the assigned DN from scene to scene is
due to a variety of factors including, sun-target-sensor geometry and atmospheric
conditions. Since DN values are scene dependent, a method to normalize the
DN5 to some standardized unit is required. The DN values assigned to pixels are

proportional to upwelling electromagnetic radiation (radiance, watts/m2/

63

steradian/pm. Converting DNs to radiance values is a rather straightforward
procedure:

Radiance = (LMAX - LMIN)!255 * DN +LMIN
The values of LMAX and LMIN, in the case of ETM+ data, can be obtained from

the header file included with the data.

3.15 Atmospheric Correction

Atmospheric effects must also be accounted for to obtain accurate
spectral reflectance measurements of the earth’s surface. Solar radiation striking
the surface of the earth and received at the sensor must pass through the earths
atmosphere twice, once on the way down from the sun to the earth and again as
the radiation is reflected from the earth’s surface back to the sensor. It should be
noted that not all of the solar radiation that strikes the earth’s surface is reflected
back to the sensor. The electromagnetic energy can be absorbed, transmitted or
reflected depending on the properties of the object the radiation encounters. The
only portion that can be directly observed in remotely sensed data is the portion
that is reflected from objects. The reflected electromagnetic energy is not
interacting exclusively with the object it strikes, it also interacts with the

atmosphere.

The atmosphere plays a significant role in transmission, reflection and

absorption, and these atmospheric affects must be taken into account to

generate an accurate dataset. Atmospheric constituents such as gasses and

64

suspended particles interact with the electromagnetic radiation passing through
the atmosphere and can significantly alter the amount of reflected energy
received at the sensor due to atmospheric scattering. Atmospheric correction
protocols provide a method to significantly reduce the effects of atmospheric

scattering resulting in a more accurate data set.

Atmospheric correction procedures are important for not only producing
accurate absolute reflectance values but are also an essential tool anytime data
taken at different dates or geographic areas are directly compared. The
atmosphere is in a constant state of flux, thus the atmospheric profile and
constituents will vary at any point in time. In order to compare reflectance
measurements from one scene to another, the atmospheric effects in both
scenes must be corrected. Otherwise, inaccurate conclusions based on
uncorrected reflectance values from one scene to another may be the end result.
Digital number values are image specific, that is a particular land use may exhibit
differing digital numbers from one to another. The digital number that is assigned
to a specific pixel is a function of a variety of factors, including viewing geometry,
location of the sun, and atmospheric conditions. Atmospheric correction can be
thought of as a normalization technique, accounting for target, sensor, sun
geometry and atmospheric conditions. Each scene is atmospherically corrected
based on a set of atmospheric correction values and target, sensor, sun positions

appropriate for each scene. Once atmospheric correction procedures have been

65

carried out, the absolute reflectance values from one scene to another may be

more accurately compared.

Visibility conditions of the atmosphere during data collection can lead to
inaccurate reflectance values if atmospheric conditions are not taken into
account. Top of atmosphere (TOA) reflectance is not only affected by visibility
conditions, but also by the surface reflectance properties of the object as well.
Atmospheric effects are both wavelength dependent and target (brightness)
dependent. Due to the fact that atmospheric transmission properties are
wavelength dependent, the atmospheric correction procedures must be applied
to each spectral band with appropriate wavelength dependent correction values.
The brightness or target dependencies are also important considerations and the
atmospheric correction procedures must accurately model target brightness as
well. Correction factors must be derived for a variety of brightness values,

including bright and dark objects, across all spectral bands.

These considerations are critical when examining targets of different
brightness conditions in a single scene. If atmospheric correction techniques are
not performed properly, the results found in one study may not be directly
comparable to other research. The ability to compare findings between data

acquired at different times and even other sensors is essential.

66

The following equation illustrates atmospheric and brightness conditions

and their relative influence on the signal received by the sensor:

TOA Reflectance = Path-radiance (Atmospheric reflectance) + (Surface
reflectance)(Atmospheric transmittance)

If surface reflectance is set to zero, ((Surface-reflectance)(Atmospheric
transmittance) = 0), then the entire signal received at the sensor is due to the first
component of the equation (Path-radiance (Atmospheric reflectance» which
consists of path radiance and atmospheric reflectance. When an object with zero
reflectance is imaged in a satellite scene, it follows that the reflectance values
observed over this object should also be zero. This is, however, not the case; the
non-reflective object will usually have reflectance values associated with it.
These reflectance values are due to the interaction of path radiance. It can be
inferred that the signal received at the sensor for dark objects (little reflectance)

will be artificially inflated due to path radiance and atmospheric reflectance.

If, on the other hand, a very bright target is contained in the view of the
sensor (Surface-reflectance = 1) the signal received at the sensor (TOA
Reflectance) will consist of: Path-radiance (atmospheric reflectance) +
transmittance of atmosphere. In this case, the reduction in transmittance due to
poor atmospheric conditions exceeds the contribution from path radiance and

thus lowers the overall reflectance signal received at the sensor.

For dark objects the atmospheric effect of path radiance and atmospheric

reflectance have the effect of brightening the reflectance observed over these

67

re

bri
of t
a s
set
T0

the

red

Old

the

Vall

dark objects. For bright objects (high reflectance), the transmittance of

atmosphere term effectively reduces the observed TOA reflectance.

3.16 Atmospheric Correction - 55 Method

The 55 (Simulation of the satellite signal in the solar spectrum)
atmospheric correction procedure incorporates atmospheric correction algorithms
based on several parameters and accounts for brightness and wavelength as
well. The 55 algorithm uses both an atmospheric and gaseous model based on
atmospheric conditions at the time of data acquisition and uses a linear
regression method to develop correction values based on both wavelength and
brightness. The 55 method, a radiative transfer model, computes the attenuation
of solar radiation and therefore the attenuation of the surface reflectance through
a series of internal models. The 55 technique generates TOA reflectance for a
set of surface reflectance values and establishes a linear relationship between
TOA reflectance and the input surface reflectance values. Linear regression is

then used to compute the surface reflectance values in the image.

The difference in relative brightness between the targets is significantly
reduced if atmospheric correction techniques are not applied to the data set. In
order to maximize the data content, and thus improve classification accuracy,
atmospheric correction is imperative. Without atmospheric correction techniques,
the brightness recorded for bare soil will be artificially inflated and the brightness

values of crop stubble will be reduced, significantly hindering classification

68

efforts. Atmospheric correction procedures ensure that the full range of

brightness values are present in the data, leading to increased overall

classification rates.

Atmospheric correction of Landsat 7 imagery was accomplished using the

55 radiative transfer modeling technique. The 55 atmospheric correction method,

and radiative transfer methods in general, have been cited in the literature as

robust and accurate for dealing with atmospheric correction. An overview of the

atmospheric correction procedure is provided in Figure 10.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Level 1G Convert DN’s Radiance to
ETM+ Data ’ To Radiance '—’ TOA Reflectance
Site Sun Latitude -
Elevation Angle Longitude
55 Model
Simulation
Aerosol Visibility Atmospheric ‘
Model Model
Atmospherically Linear Regression
Corrected Data ‘— Correction Factors

 

 

 

 

 

 

 

 

 

 

Figure 10. 55 Atmospheric Correction Flowchart.

Landsat 7 Data shipped in level 1G format are corrected for scan line

direction and band alignment errors and have also had geometric and

69

radiometric corrections applied as well. During level 1G processing, image pixels
are converted to units of absolute radiance using 32 bit floating point values,
these pixel values are then scaled to byte values or DNs (digital numbers). To
convert the pixel values back to radiance (watts/(meter
squared*steradian*micrometer) values a simple mathematical procedure is
employed, provided by the Landsat 7 data team:

Radiance = (LMAX — LMIN)/255 * DN + LMIN

Computed radiance values can then be converted to top of atmosphere

(TOA) reflectance using the following equation:

 

,0 _ rir-L/l-ai2
p Emmi-00365

where:

pp - Unitless planetary reflectance

l4, Spectral radiance at the sensor's aperture
d Earth—Sun distance in astronomical units
Esun;~ = Mean solar exoatmospheric irradiances
6?, = Solar Zenith Angle

The TOA reflectance (pp) relates the measured radiance (LA) to the solar
irradiance incident at the top of the atmosphere and ranges in value from zero to
one. Computed reflectance values are then used as input into the atmospheric

correction procedure.

The 55 model calculates correction factors individually for each ETM+

band based on model input parameters. Model input parameters include visibility,

70

 

rC

elevation, location, sun angle, atmospheric model and an aerosol model to
compute correction factors. The model then generates a set of apparent
reflectance values for a set of input reflectance values. Input and apparent
reflectance values are then linearly regressed against one another to efficiently
compute any correction value from zero to one. Linear regression parameters

generated for this study are provided in Table 3.

 

Band 1 y = 0.8957x + 0.0486
Band 2 y = 0.8683x + 0.0231
Band 3 y = 0.8981x + 0.0128
Band 4 y = 0.8897x + 0.0053
Band 5 y = 0.8841x + 0.0007
Band 7 y = 0.9072x + 0.0002

 

 

 

Table 3. 55 Derived Correction Parameters for ETM+ data.

The linear regression parameters are then used as input to create a
computer script, executed using commonly available built in scripting capabilities
of standard remote sensing software packages such as Imagine and ENVI. The
script opens the TOA reflectance file and applies the computed correction factors

for each pixel for all six bands and saves the results to a new file.

3.17 Classification Methodology
The processed ETM+ imagery data will be used as input into classification
routines to identify agricultural tillage methods. Several classification routines

exist to aid in information extraction and these methods will be utilized and

71

COITIF

for Ih

comrr
Super
serve

targets
identifi
inform;
land us

site is t

I'dentlfy
USEI de-
CIassil‘ic

classmc

be USed

and 00m;

compared with one another to determine the most suitable classification method

for this particular set of data and information extraction requirements.

As noted in section 2.5, two primary methods of classification are
commonly in use, supervised and unsupervised classification methods.
Supervised classification methods rely upon known ground truth information to
serve as training sites for the classification routine. That is, known ground truth
targets, in this case ground truthed fields with a known cropping method, are
identified in the imagery. These sites are then used to collect statistical
information regarding the set of identified pixels. The pixel information for each
land use is termed a training site and the statistical information for each training

site is then used to classify the rest of the image.

Unsupervised classification methods rely upon clustering algorithms to
identify statistically different sets of pixels and group them together using a set of
user defined rules. If ground truth information is available supervised
classification routines are commonly used and preferable over unsupervised

classification routines.

The ground truth data obtained during the driving transect of this study will
be used to establish training sites for input into a supervised classification
routine. The results of several supervised classification routines will be examined

and compared to determine the most suitable classification routine for

72

 

differentiation between conventional tillage and conservation tillage cropping
methods. Classification Routines to be examined include:

- Spectral Angle Mapping

- Maximum Likelihood

- Cosine of the Angle Concept

- Minimum Distance

- Mahalanobis Distance
3.18 Classification Accuracy Assessment

Classification accuracy results will be analyzed in an error matrix format to

examine and display errors of commission and omission, producers accuracy,
users accuracy, and overall accuracy. An error matrix provides an easily
interpreted data matrix with the main diagonal comprised of correctly classified
pixels. Pixels off the main diagonal are either errors of commission or omission
and can be readily identified as such in the error matrix. The overall classification
accuracy is measured by dividing the sum of all values contained in the main
diagonal by the total number of pixels in the error matrix. The producer’s
accuracy is a measure of omission error and is determined by the total number of
correct pixels in a category divided by the total number of pixels in that category.
The users accuracy is determined by the total number of correct pixels in a
category divided by the total number of pixels that were actually classified in that
category, a measurement of commission error. Both the users accuracy and

producers accuracy will be presented along with the error matrices for each

classification method in the results.

73

Kappa statistics were generated to provide an additional measure of
classification assessment. The kappa statistic is a discrete multivariate technique
which incorporates total overall classification accuracy and off diagonal elements
(omission and commission errors) in a single statistic. Kappa statistics are

derived using the following formula: Kappa =

r r
Nini—ZXi+X+i
i=1 i=1

r
2
N —ZXi+X+i
i=1

Kappa values range from zero to one with a value of zero indicating no

 

agreement and a value of zero indicating perfect agreement. Montserud and
Leamans (1992) evaluated kappa statistics and classification methodologies and
propose that a kappa value of .75 or greater indicates very good to excellent
classification performance. Kappa statistics were also used to evaluate the
statistical difference between the possible classification pairings. A standardized
Z-test incorporating the overall kappa score and kappa variance were used to
determine if a pairing of classification algorithms resulted in statistically

significant results.

Ground truth data collected during the driving transect portion of the study
will also serve as a classification accuracy check and validation. The known

ground truth data, obtained during the driving transect, will be compared to the

74

classified image to check for agreement between the classified image and
ground truth data. Accurate ground truth information is an important step in
classification accuracy assessment. Rather than relying on image processing
software algorithms that assess classification accuracy based on statistical
conditions, ground truth data provides a more reliable method of assessing
classification accuracy or lack there of. The validation of classified imagery
through the use of driving transect data provides additional confirmation over and

beyond error matrix analysis.

75

CHAPTER 4 RESULTS

4.1 Spectral Properties of Tillage Practices

To differentiate the two primary cropping methods, conventional and
conservation tillage, using remotely sensed data, the spectral reflectance
properties of each cropping method were examined. Bare soils (conventional
tillage) and soils with crop residues (conservation tillage) interact with
electromagnetic radiation in unique patterns. These differences in reflected
radiation between the two surfaces are relied upon during classification to
differentiate the two cropping methods. The results of spectral reflectance
patterns, obtained through laboratory, field, and satellite based measurements
are provided in the following section. Classification results based on spectral

reflectance patterns immediately follow the spectral reflectance results.

76

0.80

 

0.70

0.60 ,

0.50

0.40 7*

Reflectance

0.30

0.20

0.10 .-

 

 

 

 

0.00 .
300 800 1300 1800 2300

Wavelength (nanometers) -2- -_
' -—-Bare Soil
, -O-100% Crop Residue
‘ +60% Crop Residue

Figure 11. Laboratory Reflectance Readings.

Laboratory measurements provide data regarding the spectral response
patterns of crop residues and soils individually. That is, laboratory measurements
provide fundamental data regarding the spectral reflectance of crop residues and
soils independently. However, they do not accurately represent the complex

interaction of standing crop residues overlying soils found under field conditions.

This study confirms findings published by Gausman and Gerbermann
(1975) regarding the spectral reflectance of crop residues under laboratory

conditions. They found that crop residue exhibits a higher reflectance than bare

77

soils under laboratory lighting and sample sizes. The findings are also in
agreement with other studies examining crop residues under laboratory
conditions (Stoner and Baumgardner 1980). Spectral reflectance measurements
taken under laboratory conditions are presented in Figure 11. The spectral data
reveals that crop residues (corn and soybean) exhibit greater reflectance than
bare soil. Laboratory measurements require the use of small sample sizes and a
correspondingly small field of view (FOV) with respect to the radiometer. The
field of view used with laboratory equipment must be small enough to only record
reflectance measurements from the sample itself and not surrounding materials.
The combination of small sample size and small FOV (8 degree) results in limited
measurements that do not accurately represent field conditions. The
investigation of laboratory based reflectance measurements was made to confirm
the findings of earlier research and to illustrate the differences obtained in

reflectance measurements made under laboratory and field conditions.

lntuitively, based on laboratory measurements alone, the conclusion that
fields with crop residues (conservation tillage) should exhibit higher reflectance
than fields comprised of bare soil (conventional tillage) would be a reasonable
expectation. These previous studies however, and laboratory studies in general,
do not examine standing crop residues as found in agricultural field settings.
Standing crop residues in field conditions were found to exhibit distinctly different
spectral properties than those found under laboratory conditions. The results of

laboratory measurements and those provided in the next section, field based

78

measurements, illustrates the importance of field based measurements when

working with remotely sensed data.

4.1.2 Field Based Spectral Properties

The Field Spec Pro was also utilized for collecting field based spectral
measurements. Field measurements were collected at the Kellogg Biological
Station’s Long Term Ecological Research site, located in Hickory Corners,

Michigan as well as research fields located on Michigan State University’s East

Lansing campus.

The data reveal that the FOV used to record the reflectance
measurements has a large influence on observed reflectance readings (12).
Using the 8 degree fore optic at a height of 2 meters results in measurements
similar to those made under laboratory conditions, crop residues exhibit greater
reflectance than bare soils. However, using the 24 degree fore optic at a height
of 2 meters markedly changes the results. Crop residue reflectance is now found

to be less than that of bare soil, similar to reflectance response patterns

observed from ETM+ data.

79

 

0.7

0.6 a. _ ._ _ _ _z _ .- __ .2 ,2 ,

0.4 .- ,- a a a

 

0.3 --

Reflectance

I ,3
I, iv.
. V/
0 2 —- —— +7 .—r r .a, — ------~ —-
o ' r,
2 I
. I

 

0.1 -

 

 

 

0 1000 2000 3000 4000
Wavelength (nanometers)

 

-0— 8° Crop Stubble
-- 8° and 24° Bare Soil
-‘- 24° Crop Stubble

 

 

 

Figure 12. Field Based Spectral Reflectance Measurements,
Spectral reflectance curves of soils and crop stubble taken with
two FOV’s.

Reflectance measurements taken under field conditions at a scale where
texture, shadowing, and multiple scattering are significant reveal that fields with
crop residues exhibit lower overall reflectance than fields comprised of bare soil
alone. Field conditions consist of a complex assemblage of crop residues
overlying the soil surface. Stalks and stems rarely lie flat against the soil surface
under field conditions (Figure 13). Crop residues may lie upon one another and

are often times found still attached to the main stalk, not completely cutoff, but

suspended 10 to 12” above the soil surface at the stalk end with the other end

80

(what was once the top of the plant) resting against the soil (Figures 14 and 15).
The complex assemblage and pattern of crop residues creates a scale-
dependent surface that is overall less reflective than bare soils with larger FOV’s

for a number of reasons.

 

Figure 13. Crop Residue — Photograph of Com Residue
Vertical overhead perspective showing shadowing and texture.

 

Figure 14. Crop Residue, Photograph of Com Residues.
Oblique view showing overall field condition.

81

 

Figure 15. Crop Residue, Photograph of Soy Bean Residues.
Oblique view showing overall field condition.

Fields containing crop residues are less reflective than fields comprised of bare
soil by virtue of pattern, texture, and roughness. The rough texture associated
with crop residues results in multiple scattering of electromagnetic energy. The
end result of which is that less of the incoming radiation is reflected back to the
sensor or collection device. Conventionally tilled fields comprised of relatively
smooth soil, on the other hand, present a surface which acts more as a specular
reflective surface. Even though bare soil is less reflective than crop residues
independently, the agglomeration and texture of crop residues overlying soil

creates an overall less reflective surface under field conditions.

Another contributing factor to the observed reflectance pattern of
conservation tillage fields is shadowing. The complex pattern and texture
associated with standing crop residues creates shadowing. Shadowing from crop

residues also acts to reduce the overall observed reflectance associated with

conservation tillage fields.

82

Under field conditions, handheld radiometric measurements indicate that
standing crop residues, measured at a scale where texture and shadow come
into play, exhibit lower overall reflectance than bare soils. Texture and shadowing
by crop residues significantly reduces the amount of electromagnetic radiation
reflected (Figure 16). Henderson (1985) notes that parameters and processes

important at one scale are frequently not important or predictive at another scale.

 

 

 

 

 

0.5
0.4 - -_ , U . i o
o ,r “I.
8 0.3 - z r
:3 El
8
a: 0.2 - z
o
0:
0.1
O 7 .
350 850 1350 1850
Wavelength (nanometers) .2--- -,
0 Bare Soil
, A Corn Stubble
.— SovBsan Styeble *

Figure 16. Field Based Measurements of Reflectance.
Recorded at MSU East Lansing Agriculture Fields.

83

4.1.3 Satellite Spectral Properties

Ground truth data were used to identify known land use/land covers in the
ETM+ data set. The spectral reflectance of 38 representative fields were used to
derive a characteristic spectral reflectance signature based on 30 m resolution
ETM+ data for each land use/land cover class (Figure 17). The spectral
reflectance signatures based on the ETM+ data reveal that the spectral
reflectance of conservation tillage acreage, at this scale, is less than that of

conventionally tilled fields comprised of bare soil.

 

 

 

 

 

 

70 w
60 -
3 50 *
c I
g 40 .
o
5 30
1
5° 20 ~
10
O I ‘ 9 a ‘
0.40 0.90 1.40 1.90 2.40
Wavelength (micrometers) +Eenv."Ti ‘
I 0-~Cons.T. :
I, —r:}—Alfalfa
I -°—Water ‘
t +35%? .

Figure 17. Spectral Reflectance derived from ETM+ Data

84

The effects of texture, shadowing, and crop residue lying directly on soil

collectively create a spectral reflectance signature darker than that of a bare soil
These results are important considerations if this work is to be applied using

different remote sensing platforms. The characteristic reflectance curve of

conservation tillage may be significantly altered based on the FOV of the sensor.

4.1 .4 Spectral Characteristics at Differing Spatial Scales

Spatial scale plays an important role in determining the appropriate
spectral reflectance signatures to be used in analysis of remotely sensed
imagery. Readings taken at one scale may not necessarily be used to assess or
design a classification scheme for data taken at a different sampling scale. The
scale of the data must be taken into account. Accounting for scale can be done
operationally through the use of ground truth information. If known ground targets
are present in the image, the spectral reflectance of the ground truth target can
be obtained directly from the remotely sensed data source. The spectral
reflectance of the known ground truth target will accurately reflect the spectral
reflectance properties of the target, assuming the data was properly acquired and
processed. Spectral reflectance measurements taken at differing scales than the
remotely sensed data, to be used as input into classification routines, should be
carefully examined to determine if the reflectance from the object or land
use/land cover exhibits scale dependency. Confounding factors such as texture
and shadowing play a significant role in the reflectance properties of surface

objects or land use/land covers. If spectral data obtained through laboratory or

85

field measurements are used directly as input into a classification routine, the

results should be carefully evaluated to ensure classification accuracy.

Many man made phenomena, when examined using remotely sensed
data, are not subject to texture and shadowing effects. Concrete and asphalt are
two such surfaces. They both exhibit relatively uniform reflectance over small and
large scales and are not affected by texture and shadowing at any useful
mapping scale. However, the vast majority of natural targets are subject to
texture and shadowing effects, including vegetation, both vigorous and
senescent. Stalks, stems, and leaves create a complex texture and pattern,
creating shadows and presenting a very complicated structure as viewed from
the sensor. If spectral reflectance measurements are made at a scale too small

to account for texture and shadowing of the phenomenon of interest, they fail to

adequately characterize the spectral reflectance at larger scales.

The spatial resolution of ETM+ data is 30 meters, meaning that any object

that contributes to the brightness value (DN) of a specific pixel that is smaller
than 30 meters is a mixed pixel. For example, a pier, entirely contained in an

agricultural field consisting of plowed bare soil with a smooth uniform texture

throughout would be termed a pure pixel. The pixel falls entirely within the field
and the representative reflectance is comprised entirely of smooth bare soil, a

pure pixel. Compare and contrast this pixel to a pixel obtained from a

conservation tillage field. The reflectance value of this pixel, which also lies

86

entirely in the field, is comprised of a contribution from crop residues, shadows,
and soil, a mixed pixel. The terminology of pure and mixed pixels is scale
dependent and can have a variety of meanings. Common usage of the term
mixed pixel is one in which two land use/land covers are contained in a single

pixel. A road bordering an agricultural field in which a single 30-meter pixel

contains a portion of the road and a portion of the field is an example of the most
common meaning of the term.

The spectral reflectance properties of crop residues under laboratory

conditions show markedly higher overall reflectance values than the spectral
reflectance of bare soil. It only follows that adding a more reflective material on
top of bare soil should result in an overall increase in brightness for the pixel, but
this not the case, in fact just the opposite is observed. A pixel with bare soil and
standing crop residue when viewed from space, in 30m pixels, is darker or less
reflective than bare soil alone. The underlying mechanisms responsible for the

variation of reflectance observed for field and laboratory conditions are scale and
texture.

Crop residues under agricultural field conditions are not smooth and
homogenous, they can be thought of as a regular repeating pattern across the
field at 30m scales, but they are still far from a smooth surface. Crop residues do
not entirely lay down flat against the surface of the field but rather are a complex

assemblage of stems, stalks, and leaves. A great deal of crop residue that may

87

at first be regarded as lying down even with the surface of the field, when
examined more closely, may in fact be suspended above the surface. Crop
residues may be suspended at one end still attached to the main stalk, merely
knocked over, and can also be found standing straight in the air to the level the
combine cut them the previous year (usually a height of 8-12 inches). Residues
that were cut with harvesting equipment may be completely suspended
horizontally on top of other stalks. The suspended and standing residues

produce a significant amount of shadowing, shadowing that significantly lowers

the overall observed reflectance.

Another factor that comes into play is multiple scattering. As noted in the

laboratory spectral reflectance measurements of crop residues section,
reflectance is generally high for crop residues, assuming a flat and homogenous
crop residue surface which neglects shadowing. The high reflectivity of crop
residues results in multiple scattering of light photons. Photons striking a stalk
may be bounced at irregular angles either once off the stalk and then into
another direction (not received by the sensor) or may strike multiple stalks before
being reflected into another direction. The end result being the same: the
reflected energy that was radiant onto the crop stubble surface is never received
back at the sensor, resulting in lower recorded reflectance values. Contrast this
with reflectance from a field of recently plowed smooth soil. Even though the
reflectance of soil is less than the reflectance of crop residues individually,

because there is less scattering, and shadowing, more of the electromagnetic

88

energy that strikes the bare soil areas is reflected back to the sensor producing

higher reflectance measurements at the sensor relative to fields with crop

residues.

4.2 Classification Training Sites

With ground truth data processed, it is now possible to extract spectral
profiles from the atmospherically corrected ETM+ imagery to examine the
spectral profiles of conventional and conservation tillage fields. Training site
selection entailed examining available GPS logged fields in the ETM+ imagery
and selecting training areas from the known land cover/land use located in the
imagery. Thirty eight (38) of the four hundred and eighty (480) data points
collected during the driving transect portion of this study were used to identify
areas in the ETM+ to collect training pixels for each class. Each class is
represented by a minimum of 900 pixels. Eleven thousand four hundred eighty
pixels (11,481) were used to collect spectral information from the ETM+ scene as
input for signature training:

Pixels Class

908 Conventional Tillage
1481 Conservation Tillage

942 Alfalfa
1038 Winter Wheat
1 163 Water

5948 Forested

4.3 Classification Methodology Comparison

The classification portion of this study was performed using the same

training site areas and ground truth data points for each classification routine

89

examined. This ensures that the results of each classification are based and
evaluated only on the classification algorithm itself and that differences in
classification accuracy are not attributable to differences in training site selection
or ground truth data. The classification routines examined for this study include:

Minimum Distance

Mahalanobis Classifier

Maximum Likelihood

Spectral Angle Mapping (SAM)
Cosine of the Angle Concept (CTAC)

 

Error matrices and ground truth data were examined for each
classification method to assess the suitability of the classification scheme for this

study. Classification accuracy results and error matrices are reported tables 4

through 8.

90

Maximum Likelihood

 

 

 

 

1 2 3 4 5 6 Total
1 55 33 0 0 0 0 88
2 34 171 0 0 0 0 205
3 0 0 36 20 0 0 56
4 0 0 0 22 0 10 32
5 0 0 0 0 27 0 27
6 0 0 0 0 0 34 34
Total 89 204 36 42 27 44 442

 

1 - Conventional Tillage, 2 - Reduced Tillage, 3 - Alfalfa,
4 - Winter Wheat, 5 - Water, 6 - Forested

Reference Classified Number Producers

Users

 

 

 

 

 

Class Totals Totals Correct Accuracy Accuracy
1 89 88 55 61.80% 62.50%
2 204 205 171 83.82% 83.41%
3 36 56 36 100.00% 64.29%
4 42 32 22 52.38% 68.75%
5 27 27 27 100.00% 100.00%
6 44 34 34 77.27% 100.00%

Total 442 442 345 l

Overall Classification Accuracy = 78.05%

Kappa Statistics = 0.6940

Table 4. Maximum Likelihood Classification Error Matrix.

91

 

Mahalanobis Distance

 

 

 

 

 

1 2 3 4 5 6 Total
1 55 33 0 0 0 0 88
2 34 171 0 0 0 0 205
3 0 0 36 20 O 0 56
4 0 0 0 22 0 1 1 33
5 0 0 0 0 27 0 27
6 0 0 0 0 0 33 33
Total 89 204 36 42 27 44 442
1 - Conventional Tillage, 2 - Reduced Tillage, 3 - Alfalfa,
4 - Winter Wheat, 5 - Water, 6 - Forested
Reference Classified Number Producers Users

 

 

 

 

 

Class Totals Totals Correct Accuracy Accuracy
1 89 88 55 61.80% 62.50%
2 204 205 171 83.82% 83.41 %
3 36 56 36 100.00% 64.29%
4 42 33 22 52.38% 66.67%
5 27 27 27 100.00% 100.00%
6 44 33 33 75.00% 100.00%

Total 442 442 344 I

Overall Classification Accuracy = 77.83%

Kappa Statistics = 0.6909

Table 5. Mahalanobis Distance Classification Error Matrix.

92

 

Minimum Distance

thN-fi

6
Total

 

 

 

 

1 2 3 4 5 6 Total
52 13 0 0 0 0 65
37 190 0 0 0 0 227

0 0 36 1 3 0 0 49

0 0 0 28 0 6 34

0 0 0 0 27 0 27

0 1 0 1 0 38 40
89 204 36 42 27 44 442

 

1 - Conventional Tillage, 2 - Reduced Tillage, 3 - Alfalfa,
4 - Winter Wheat, 5 - Water, 6 - Forested

Class
1

(Vi-FUN

Total

 

 

 

 

 

Reference Classified Number Producers Users

Totals Totals Correct Accuracy Accuracy
89 65 52 58.43% 80.00%
204 227 190 93.14% 83.70%
36 49 36 100.00% 73.47%
42 34 28 66.67% 82.35%
27 27 27 100.00% 100.00%
44 40 38 86.36% 95.00%
442 442 371 l

Overall Classification Accuracy = 83.94%

Kappa Statistics = 0.7719

Table 6. Minimum Distance Classification Error Matrix.

93

 

Spectral Angle Mapping

 

 

 

 

 

1 2 3 4 5 6 Total
1 87 8 0 0 0 0 95
2 2 196 0 0 0 0 198
3 0 0 35 5 0 0 40
4 0 0 0 15 0 9 24
5 0 0 0 0 27 0 27
6 0 0 1 22 0 35 58
Total 89 204 36 42 27 44 442

 

 

1 - Conventional Tillage, 2 - Reduced Tillage, 3 - Alfalfa,
4 - Winter Wheat, 5 - Water, 6 - Forested

Reference Classified Number Producers Users

 

 

 

Class Totals Totals Correct Accuracy Accuracy
1 89 95 87 97.75% 91 .58%
2 204 198 196 96.08% 98.99%
3 36 40 35 97.22% 87.50%
4 42 24 15 35.71% 62.50%
5 27 27 27 100.00% 100.00%
6 44 58 35 79. 55% 60.34%

Total 442 442 395 I

 

 

Overall Classification Accuracy = 89.37%

Kappa Statistics = 0.8524

Table 7. Spectral Angle Mapping Classification Error Matrix.

94

Cosine of the Angle Concept

 

 

 

 

1 2 3 4 5 6 Total
1 88 6 0 0 0 0 94
2 1 198 0 0 0 0 199
3 0 0 36 5 0 0 41
4 0 0 0 37 0 1 38
5 0 0 0 0 27 0 27
6 O 0 0 0 0 43 43
Total 89 204 36 42 27 44 442

 

 

1 - Conventional Tillage, 2 - Reduced Tillage, 3 - Alfalfa,
4 - Winter Wheat, 5 - Water, 6 - Forested

Reference Classified Number Producers Users

 

 

 

 

 

Class Totals Totals Correct Accuracy Accuracy
1 89 94 88 98.88% 93.62%
2 204 199 198 97.06% 99.50%
3 36 41 36 100.00% 87.80%
4 42 38 37 88.10% 97.37%
5 27 27 27 100.00% 100.00%
6 44 43 43 97.73% 100.00%
Total 442 442 429 |
Overall Classification Accuracy = 97.06%

Kappa Statistics = 0.9592

Table 8. Cosine of the Angle Concept Classification Error Matrix.

95

Of the five classification routines examined only two provided overall
classification accuracy results of greater than 85%: spectral angle mapping
(SAM), and the cosine of the angle concept (CTAC). The overall classification

accuracies obtained for all six land use/land cover classes are as follows:

 

Overall Classification Accuracy Kappa Statistic
Mahalanobis Classifier 77.83% .6909 f
Maximum Likelihood 78.05% .6940 ,
Minimum Distance 83.94% .7719 _1"
Spectral Angle Mapping 89.37% .8524 l
Cosine of the Angle Concept 97.06% .9592

Table 9. Classification Results.

When examining the overall classification accuracies of all six land
use/land cover types, particular attention must be paid to the individual
contributions of the six land use/land covers to the overall classification results.
The water class was correctly identified by all five classification routines with

100% accuracy, thus raising the overall classification accuracy results.

An examination of the individual land use/land cover type classification
accuracies also indicates that the spectral angle mapping and cosine of the angle
classification routines were the only classification routines to achieve 85%
classification accuracy for conventional tillage and conservation tillage land
use/land covers. The remaining classification routines fell below the 85%
classification accuracy level for both conventional tillage and conservation tillage

land use/land cover classes.

96

Statistical analysis utilizing Kappa statistics and measures of variance
derived from the classification error matrix were examined to determine if the
results of the classification methods are statistically different from one another at
a 95% confidence interval. The error matrices were examined and analyzed
using a method outlined by Congalton (1983). The method involves computing
the variance of the error matrices as well as row and column residuals to
calculate kappa variance. Statistical analysis indicates that of all the possible
pairings of the classification methods only the pairing of maximum likelihood and
mahalanobis distance classifiers were not significantly different examined at a
95% confidence level (Z-Score < 1.96). All other classification method results

were statistically different using a 95% confidence interval (Z-Score > 1.96).

 

Classification Pairing Z-Score
Max. Like. vs. Mah. Dist. 0.08
Min. Dist. vs. Max. Like. 2.08
Min. Dist. vs. Mah. Dist. 2.16
SAM vs. Min. Dist. 2.53
SAM vs. Max. Like. 4.65
SAM vs. Mah. Dist. 4.73
COTA vs. SAM 5.10
COTA vs. Min. Dist. 7.15
COTA vs. Max 9.13
COTA vs. Mah. Dist. 9.22

 

 

 

Table 10. Results of Classification Statistical Significance Pairings.

97

4.4 Classified Imagery Results

Results of total land use/land cover for the study area, using the cosine of
the angle classified image, the classified data with the highest classification
accuracy results, are presented in Table 11 and an image of the classified data is
presented in Figure 18. The results indicate that overall corn and soybean
production under both conservation and conventional tillage practices account for
a total of 2,956,309 acres or 47% of the total land use/land cover. If only
agricultural land use/land cover classes are examined, then corn and soybean
production accounts for 82.28% of all agricultural land use/land cover throughout
the study area. Of the 2,956,309 acres of land use/land cover used for corn and
soybean production, 63.63% (1,881,241 acres) are cropped using conservation
tillage techniques, while the remaining 36.37% (1,075,067 acres) are produced

utilizing conventional tillage techniques (Figure 19).

98

 

 

E Conventiona
1:] Conservation Tillage
|]|]][|l|]]||||]| Alfalfa

\Mnter Wheat

- Water

Forested

 

Figure 18. Classified Data Based on the COTA Classification Routine.

99

Total LanduselLandoover

 

Conventional Tillage
Conservation Tillage
Winter Wheat
Alfalfa

Water

Forested

Cloud, Urban

Total

 

 

 

 

Square Meters Acres %
4,352,500,690.73 1,075,067.67 12.49%
7,616,362,599.87 1,881,241.56 21.86%

600,593,79934 148,346.67 1.72%
1,977,373,212.72 488,411.18 5.67%
638,604,589.74 157,735.33 1.83%
11,623,696,504.34 2,871,053.04 33.36%
8,035,462,95875 1,984,759.35 23.06%
34,844,594,355.49 8,606,614.81 100.00%

 

Total LanduselLancover Area used in Analysis - Urban/Cloud Masked

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

LanduselLandoover Square Meters Acres %
Conventional Tillage 4,352,500,690.73 1,075,067.67 16.24%
Conservation Tillage 7,616,362,599.87 1,881,241.56 28.41%
Winter Wheat 600,593,79934 148,346.67 2.24%
Alfalfa 1,977,373,212.72 488,411.18 7.38%
Water 638,604,589.74 157,735.33 2.38%
Forested 11,623,696,504.34 2,871,053.04 43.36%
Total 26,809,131 ,396.74 6,621,855.45 100.00%
Total LanduselLancover - Agriculture

Square Meters Acres Total %
Conventional Tillage 4,352,500,690.73 1,075,067.67 29.92%
Conservation Tillage 7,616,362,599.87 1,881,241.56 52.36%
Winter Wheat 600,593,79934 148,346.67 4.13%
Alfalfa 1 ,977,373,212.72 488,411.18 13.59%
Total 14,546,830,302.66 3,593,067.08 100.00%
Total Major Row Crop

Square Meters Acres Total %
Conventional Tillage 4,352,500,690.73 1,075,067.67 36.37%
Conservation Tillage 7,616,362,599.87 1,881,241.56 63.63%

11,968,863,290.59 2,956,309.23 100.00%

 

Table 11. Study Area Land Use/Land Cover.

100

 

 

 

 

The results reveal the extent to which corn and soybean acreage dominates
agricultural land use/land cover throughout the study area (Figure 19). Winter
wheat and alfalfa only account for 9.62% of the total land use/land cover
throughout the study area. If the results are examined only in terms of agricultural
land use/land cover, winter wheat and alfalfa account for 17.72% of agricultural

land use/land covers.

 

 

 

 

 

 

3,000,000.00 __
21500100000 m * at")? rioaa‘refiaar
Tillage , ,
2,000,000.00 .W2 A - git-39 lConservatlon a
"’ I}: militia eWh tI
Fvw-i? In er ea
g 1,500,000.00 22;, ' l
< lAlfalfa
1,000,000.00 - -7 . mWater
500,000.00 \ f} » ElForested
0.00 - \

 

 

 

 

 

 

 

 

 

Figure 19. Study Area Land Use/Land Cover in Acres.

4.5 Cosine of the Angle Concept - Spectral Angle Mapping

Two relatively new classification routines were examined to assess their
suitability for successful discrimination of cropping types using remotely sensed
data. Spectral angle mapping is a technique that uses angular measures to
match image spectra to reference spectra in n-dimensions using a physically

based spectral classification method. The reference end member spectra used in

101

a spectral angle mapping classification are no different than the end member
spectra used in any other classification routine and can imported or derived from
a variety of sources including ASCII text files, spectral libraries, statistics files, or
can be extracted directly from the image (as ROI or AOl average spectra). SAM
compares the angle between the end member spectrum (considered as a n-
dimensional vector, where n is the number of bands) and each pixel vector in n-
dimensional space. Smaller angles represent closer matches to the reference
spectrum. This technique, when used on calibrated data, is relatively insensitive
to illumination and albedo effects. (ENVI 3.5 Online Documentation). Sohn et al.
(1999) has also developed a classification method based on spectral angle
mapping known as the cosine of the angle concept (CTAC). Their technique is
similar to the spectral angle mapping algorithm available in ENVI with the notable
difference that the cosine of the angle concept uses the cosine of the angle as
the decision rule rather than just the angle. While spectral angle mapping uses
the angle between the reference spectra and the pixel to be classified in n
dimensional space (n being the number of bands). The CTAC technique takes
the procedure one step further and bases the classification rule on the cosine of
the angle formed between the reference signature and the pixel being classified.
The cosine of the angle concept is reported to be less sensitive to illumination
and linear scaling (Sohn et al., 1999), an important attribute when trying to

differentiate spectrally similar targets.

102

As illustrated by the classification results, both the spectral angle mapping
and cosine of the angle concept classification routines produced overall
classification results in excess of 85%. More importantly, both classification
routines were successful in discrimination between conventional and
conservation tillage cropping methods, surpassing 85% classification accuracy
for both tillage methods. Both methods claim to be insensitive to illumination and
albedo effects and successfully discriminate between spectrally similar targets, a I“
task where other classification methods failed to produce classification

accuracies of greater than 85%.

With such spectrally similar targets (bare soil vs. bare soil with senescent
plant material) traditional classification methods such as geometric or probability
decision rules are simply not robust enough to successfully classify the
phenomena of interest. Geometric and probability based classification routines
do not provide a decision rule or surface sufficiently discriminatory to successfully

classify spectrally similar targets.

4.6 Angular vs. Geometric and Probability Based Classification
Geometric classification routines, based on the nearest neighbor principle,
where the unclassified pixel is assigned class membership based on the
minimum distance to a class mean provides a simple and time efficient method of
classification. However, minimum distance to means classification routines, due

to their inherent simplicity, often lead to inaccurate classification results when

103

trying to differentiate class membership of spectrally similar targets, as is the
case in this study. The mean value of conservation tilled and conventionally tilled
classes are relatively similar to one another, and the classification decision rule
based on minimum distance can often lead to inaccurate classification results

when working with such spectrally similar classes.

Probability based classification decision rules such as maximum
likelihood, which incorporate variance and covariance data into the decision rule,
are generally expected to outperform distance based classification decision rules.
Contrary to this commonly held belief, this study indicates that minimum distance
classification decision rules outperformed probability based classification
decision rules. Whereas the minimum distance classification decision rule
produced overall classification accuracy results of 83.94%, the mahalanobis
distance and maximum likelihood classification decision rules resulted in overall

classification accuracies of 77.83% and 78.05% respectively.

The differences observed between distance based and probability based
classification decision rules are a result of data normality, or more appropriately,
the lack thereof. Probability based decision rules such as maximum likelihood
operate under the assumption of data normality, if the data do not follow a normal
distribution the calculation of variance data is fundamentally flawed from the
beginning. The lack of normality in the data sets resulted in lower overall

classification accuracies for classification decision rules working under normality

104

 

assumptions. Distance based classification decision rules outperformed these

classification routines for the same reason.

4.7 Validation

A total of 480 GPS data points were created during the ground truth
driving transect portion of the study. The locations of fields were mapped within
an accuracy of 4 meters to identify and locate these fields in the ETM+ data set.
Of the 480 data points six main classes were identified from ground level
observations:
Conventionally tilled fields comprised of bare soil
Conservation tilled fields comprised of soil and crop residue
Winter wheat
Alfalfa

Forested
Water

93.01.6935)?

These six feature classes represent the great majority of land use/land cover
throughout the ETM+ scene. Of the 480 GPS points, 38 were used to identify and
develop classification training sites. For validation purposes only the remaining

442 data points were used.

In addition to these six primary classes, several other land use/land cover
types occur in the imagery but were not included in this analysis for a variety of
reasons. Urban land use/land cover types such as residential, commercial and
industrial sites were masked out/ignored due to the fact that virtually no
agricultural production occurs in urban areas. Several rural land use/land cover

types were also not included as classification types due to the small spatial

105

 

extent of these areas. Barnyards and pastures were not separated from alfalfa
acreage for due to the fact that these open grass areas do not cover much
acreage spatially, both the driving transect portion of this study and the ETM+

imagery confirm this.

Winter wheat is a row crop that differs from corn and soybeans in several
aspects. The most evident difference is crop phenologY; winter wheat is planted
in the late fall/early winter and harvested in late spring/early summer. At the time
of ETM+ imagery acquisition, most winter wheat was still in the field in the final
stages of plant development. Winter wheat fields during this time period are a
deep green color with a very tightly spaced plant pattern offering an almost

uniformly closed and tight crop canopy.

Alfalfa is grown throughout the Midwestern United States as a forage crop.
Alfalfa cropping is significantly different than any other crop grown in the region.
Alfalfa is an annual crop that is planted once every three to four years. Once
planted, the crop will reestablish itself for several years without replanting. The
temporal cycle of harvesting is also unique to alfalfa, the alfalfa is cut and
harvested several times a year. The number of cuttings for each growing season
is dictated primarily by temperature and precipitation; in the study area, three to
four cuttings per growing season is typical. Because alfalfa is harvested
numerous times throughout the growing season, remotely sensed imagery

collected during the growing season may capture any particular alfalfa field in a

106

‘
ni

§/.-__

continuum of plant development from recently cut to full growth. Alfalfa fields
were recorded during the driving transect in a variety of conditions across the
study area, including recently harvested fields as well as fields in peak
development. During this time (late May to early June), most alfalfa fields were
found in the late stages of peak development, just prior to the first cutting of the
year. However, some fields had just been recently harvested. Alfalfa also has a
very tight and closed canopy during the final stages of plant development, but

differs from winter wheat in color and overall canopy texture. Alfalfa is generally a

t

lighter green color than winter wheat and the plant itself is leafier and more

spread out.

The temporal aspects of alfalfa and winter wheat, which are green during
the time of imagery acquisition, lend themselves well to agricultural classification.
Throughout the Midwest, and the study area, any conventional or conservation
tillage field will be used for corn or soybean production with few exceptions.
During the driving transect portion of this study both conventionally tilled and
conservation tillage fields were recorded, of which either corn or soybeans will be
produced. While differentiation between corn and soybeans in not possible at this
time, however, a second image acquired later in the growing season would allow

for classification between the two crops.

Forested areas across the study area were recorded during the driving

transect as well. Forested areas were found in a leaf on stage of plant

107

development, with deciduous trees exhibiting full canopies. Water bodies such as
lakes and ponds were also logged during the driving transect, although no effort
was made to explicitly include water bodies in the driving transect development.
Both forested areas and water bodies are easily distinguished in remotely sensed
imagery, therefore most of the driving transect data points were focused on

collecting data regarding agricultural crops and cropping methods, the primary

. "!

focus of this study. An effort was made to collect ground truth data for forested

areas, water bodies, and secondary agricultural crops such as winter wheat and

w .

alfalfa to ensure that the classified thematic map is as accurate as possible. It
should be recalled, however, that the primary goal of this study is the

differentiation of cropping methods utilizing remotely sensed data.

All six major classes were documented using GPS data logging methods
with an accuracy of 4 meters or less. These GPS data points, waypoints, were
then overlaid onto the ETM+ imagery to locate the ground truthed fields in the
imagery. The waypoints were converted to an ArcView point shapefile to
represent the GPS waypoints in both a GIS vector layer that can be used
independently as well as serve as a vector overlay onto the ETM+ imagery. The
point shapefile was used for both accuracy assessment and training site

selection for the supervised classification portion of this study.

The GPS data was exported from the GPS unit (Garmin GPS V) in a .dbf

format which provides the GPS information in a tabular format providing such

108

information as x,y coordinates, date and time, elevation, heading and other
information. The DBF file was then imported into ESRI’s ArcView GIS software
and converted to a point shapefile. The GPS data was recorded in UTM WGSB4
coordinates, the same projection properties as the ETM+ data, so no reprojection
of either the GPS or ETM+ data was required. Attribute information was added to

the point shapefile from notes taken during the driving transect.

During the driving transects, when GPS data points were recorded, a
notebook was used to record the waypoint number and class. For example, when
waypoint 10 was recorded with the GPS unit, an entry was made into the
notebook with the waypoint number and a description of the field associated with
waypoint 10. Recall that the entire driving transect data collection process was
also recorded on digital videotape with GPS position overlay to further ensure
accuracy. The GPS data recorded as waypoints and the GPS overlaid video
recording both contain UTM coordinates and a time stamp. Positional data along
with a time stamp ensures data collected as waypoints can be identified and

correlated with data contained in the continuous video recordings.

The field description information was then entered into the point shapefile
attribute table providing descriptive information for each point. Each point was
manually adjusted spatially to account for the fact that the GPS data was
re corded from a road and not the actual field itself. The appropriate offset was

determined from the directional heading information provided by the GPS data in

109

 

the attribute table in conjunction with the field notes to ensure that each point
location was spatially repositioned in the appropriate field or representative

location. The 442 GPS data points are comprised of:

89 Conventional Tillage

204 Conservation Tillage

36 Alfalfa

42 Winter Wheat

44 Forested

27 Water

During GPS data collection an effort was made to concentrate primarily on

agricultural targets. Forested areas and water are easily discernable in ETM+
imagery and so in the interest of time and effort many forested areas and water
bodies were not logged with GPS positional data. However, throughout the
driving transect every effort was made to log large (greater than approximately
80 acres) agricultural fields encountered during the course of the driving transect.
Larger agricultural fields were used for several reasons, including spatial
accuracy and pixel purity. Because these points provide ground truth data, the
spatial accuracy of the GPS points and the associated fields must be made with
a great deal of accuracy. Focusing on large fields during GPS data acquisition
ensures that the data points will be located in the appropriate field. Large fields
also provide better targets for training sites and classification accuracy
assessments due to the fact that large fields will be comprised of pixels entirely

contained in the field. Small agricultural fields have a much higher chance of

being represented by mixed pixels in 30 meter ETM+ imagery.

110

The results of validation, comparing the classified image to known ground
truthed data points collected during the driving transects, are reported in Table

12.

 

 

Ground Truth
Data Points Class Classified Correctly
89 Conventional Tillage 88
204 Conservation Tillage 198
36 Alfalfa 36
42 Winter Wheat 37
44 Forested 43
27 Water 27

 

 

 

Table 12. Validation Results of the Cosine of the Angle Classification Routine.

4.8 Potential Sources of Classification Error
4.8.1 Soil Background Effect
Analysis of the classified data indicates that soil background has little if

any affect on the accuracy of the classified imagery. The study area contains 35

soil types based on MLRA types. Of the 35 soil types located in the study area,

30 are suitable for agricultural production. The remaining soils are either too wet

(mucky) or predominately glacial tills (low in organic matter, high in sand and

gravel). A GIS analysis was performed to assess the classification rates based
on MLRA soil types, and the classification accuracy was found to be the same
over all agricultural soils (Figure 20). The MLRA identified agricultural soils may
vary to the soil scientist, but when viewed based on their spectral properties, the

soils are very similar spectrally. The spectral similarity of the agricultural soil

111

 

types works well for this study, as the background effect of soils is essentially the
same throughout the scene. If the findings of this study are carried out in other
areas of the Midwestern United States, the spectral properties of the background
soils should be assessed to ensure that soil types do not affect classification

results.

   

       
      

 

 

      
   
  

RN. \31 ‘ ‘ I.
“.55 /%//§\ \
it? ‘/\

 

/
/ s. /é\
'\

\k/
4% $‘/

 

 

§/%/>//§\k\§\\ lsrll.
/

 

   
        

 

 

 

    

     
 
 

   
 

 

 

 

 

 

 

 
 

.. ///K II II "I
xl'". It... ”Illa-551..., 1..
s“ % II" III §II A! “$1 I
A MN ’ l "lull" ° _ I

Figure 20. Soil Types and Ground Truth Collection Locations
Shadings correspond to STATSGO soil types.

Weather across the study area was also examined to determine if soil

moisture across the area was uniform. Data from eight weather stations across

112

the study area indicate that there was a uniform light rain event three days prior
to image acquisition amounting to .05” across the study area. Prior to this rainfall
event, no rain was reported across the study area through May 20‘“, 7 days prior
to imagery acquisition. Based on these findings, the assumption that soil
moisture was relatively uniform for each soil type found throughout the study

area is being relied upon. While some soils may retain moisture more efficiently

1!
3L

than other soil types, the classification results indicate that soils and/or soil

moisture had no impact on overall classification accuracies across the soil types.

or”:

4.8.2 Mixed Pixels

Mixed pixels are another potential source of classification error. Mixed
pixels are pixels that contain a spectral contribution from more than one land
use/land cover. During the development of classification training sites, areas are
chosen that completely lie within a field. Each pixel, and the associated
reflectance values, is comprised of a single land use/land cover. If a pixel to be
classified lies on the boundary of a field, half the pixel contained in a
conservation tilled field and the other half contained in an alfalfa field, the
associated reflectance values will be a combination of the two targets. When a
classification routine is presented with such a target or spectral reflectance
pattern, it will not be readily identified as belonging to a particular class

membership and may be misclassified.

113

CHAPTER 5 CONCLUSIONS I DISCUSSION

A significant gap in the literature exists regarding the spectral
characteristics of tillage methods, as well as the application of classification
routines to accurately identify and map tillage methods. While many articles have
been published on the environmental impacts and consequences of cropping
methods, they lack fundamental spatial data regarding the total area and the
spatial distribution of cropping methods. Without fundamental spatial data, the
results of research investigating the environmental impacts of cropping methods
from field based experiments cannot be scaled up to larger watershed or regional

scales.

The classification of conservation and conventional cropping methods was
achieved using spectral angle classification techniques. The two spectral angle
techniques examined were spectral angle mapping and the cosine of the angle
concept. Of the classification routines investigated, only the two methods based
on spectral angles satisfied the requirement of 85% classification accuracy, a
commonly accepted guideline set forth by the USGS standards for interpretation

of remotely sensed data.

The underlying mechanism and explanation of the success of spectral

angle classification techniques over geometric and statistical based classification

routines, lies in the ability of spectral angle mapping techniques to account for

114

illumination and brightness effects. The reflectance properties of individual land
use/land cover types or objects on the surface of the earth interact with
electromagnetic radiation across the electromagnetic spectrum in a unique way.
However, these unique reflectance patterns associated with the interaction of
electromagnetic radiation and the surface of the object can be perturbed a
number of ways. The most common of which is a linear scaling of the spectral
reflectance pattern above and below the mean reflectance value due to

illumination and brightness effects.

Individual land use/land cover types captured in remotely sensed imagery
exhibit a range of reflectance values throughout the remotely sensed data
product. These variations are caused by lighting and illumination effects such as
shadowing and are also exacerbated by the atmosphere. While the overall
reflectance for any particular land use/land cover may be higher or lower than the
mean reflectance values across the electromagnetic spectrum, the reflectance
pattern is a linearly scaled version of the mean overall reflectance pattern. While
individual land use/land covers exhibit a range of linearly scaled possibilities, this

knowledge can be used advantageously within the classification decision rule.

Traditional geometric and statistically based classification methods rely on
classification decision rules that do not account for the linear scaling of overall
reflectance patterns. The linear scaling of overall reflectance patterns are often

times mistakenly classified as altogether different land use/land cover classes

115

 

using geometric or statistically based decision rules for classification. Spectral
angle mapping techniques incorporate linearly scaled reflectance patterns into
the decision rule. This avoids the problem of misclassifying land use/land covers,

which are linearly scaled versions of a particular reflectance pattern.

Spectral angle mapping techniques more accurately account for
illumination and brightness effects than standard statistical or geometric
classification decision rules. Spectral angle classification routines are able to
account for brightness and illumination effects by virtue of the decision rule, one
based on an angle formed between the reference spectral reflectance pattern
and an unknown spectral reflectance pattern. Rather than partitioning the
decision space based on geometric or statistical measures, the unknown
reflectance signature is compared to a reference signature based on the angle
formed between the two. Comparing reflectance patterns using these angular
methods better accounts for spectral variations due to brightness and illumination

effects.

As reported by Sohn and Rebello (2002), the cosine of the angle approach
is well suited for discrimination of linearly scaled reflectance patterns. This is due
to the fact that linearly scaled reflectance patterns, when compared to the
reference spectrum, result in a cosine of the angle value of zero or very close to
it. Atmospheric and topographic effects act to linearly scale reflectance patterns.

That is, atmospheric and topographic effects may brighten or darken the overall

116

observed reflectance, but the overall reflectance pattern is, in fact, a linearly

scaled version of the reference reflectance pattern.

Traditional geometric and statistical classification routines base
classification decision rules on statistical measures, whereas spectral angle
mapping techniques rely on spectral pattern shape. The reliance on spectral 1
F1 1

angles, rather than probabilities, offers other advantages as well. A primary

advantage of using angular measures is that assumptions regarding data

 

normality, which are often violated with remotely sensed data sets, can be
avoided all together. “Spectral angle classifiers do not require the data to be
normally distributed and they are insensitive to data variance and the size of the

training data set” (Sohn and Rebello, 2002).

Results indicate that the discrimination of cropping methods through the
use of remotely sensed data is not only possible, but can be performed using
widely available data sources. These results provide a method to quantify carbon
sequestration potentials based on agricultural management practices over a
broad geographic area. The findings of this study can be further applied to other
Landsat scenes across the Midwestern United States to more accurately quantify
the carbon sequestration potential of agricultural lands. The findings of scaled up
research can then be used as input into carbon modeling algorithms to determine
the impact of agricultural cropping methods across the landscape as they relate

to the overall carbon budget.

117

Additionally, the findings can be used for a wide variety of environmental
studies where widespread agricultural production occurs. These studies may
include the assessment of water quality and soil erosion over entire watersheds,
providing new insight into the relationship of row crop agriculture and the

environment.

5.1 Carbon Cycle Implications

The processes and dynamics controlling the global carbon cycle have
been the subject of a great deal of research over the past decades. Research
has provided a better understanding of the carbon cycle, however, fundamental
data to provide a full account of the carbon cycle and budget has eluded
researchers. Additional data regarding the spatial extent and areal distribution of
carbon storage and release are required to arrive at more accurate estimates of
carbon sinks and sources. The United States Global Change Research Program
(USGCRP) notes that a need exists for “comprehensive, unbiased scientific
understanding of sources and sinks of carbon dioxide on continental and regional

scales...” (USGCRP, 2003).

Carbon is the basis of all life on earth and is the primary building block of
plant life. Plant growth is driven by photosynthesis, the process by which plants
convert sunlight to chemical forms of energy which drive biological systems.

Plants, through the process of photosynthesis, take in carbon dioxide from the

118

 

atmosphere and store it as plant material (carbohydrates) as illustrated by the
following equation:

6 C02 + 6 H20 —» CGH1205 + 6 02
Extraction of atmospheric carbon dioxide through photosynthetic action by plants
is only one of the many processes and interactions that transforms and cycles

carbon through the biosphere.

The carbon cycle is comprised of a series of interactions among large
carbon pools consisting of atmospheric, terrestrial and oceanic components.
Carbon cycles through these large pools or reservoirs much like the hydrologic
cycle and carbon is regularly exchanged between all three major carbon pools
through a complex series of interactions. The carbon cycle was once primarily
controlled by nature and natural processes governed flows between and among
carbon reservoirs. The actions of man have significantly altered the natural flow
of carbon and, now more than ever, have a large impact on the carbon cycle.
Land use conversions from natural ecosystems to managed production and fossil
fuel combustion are two examples of man's impact and influence on the carbon
cycle. Conversion of native ecosystems to managed production has resulted in a
redistribution of carbon from plants and soils to the atmosphere. The extent to
which anthropomorphic changes have on natural processes is crucial to a better

understanding of mans role and influence on the carbon cycle.

Recent observations of atmospheric carbon dioxide levels taken at the

Mauna Loa observatory in Hawaii since 1958 show a rapidly increasing trend in

119

atmospheric levels of carbon dioxide. The primary drivers behind the rise of
atmospheric carbon dioxide levels are thought to be due to land use/land cover
change and the combustion of fossil fuels. Land use/land cover change releases
carbon in large part due to deforestation activities including slash and burn
agriculture. Human land use/cover change has transformed one third to one half
of the earth’s ice free surface (Vitousek, 1994). This land cover change, taking
place over such great spatial extents, has resulted in a large release of carbon
dioxide to the atmosphere. Additionally, the mining of old carbon in the form of
fossil fuels and the ensuing combustion of these fuels has also been a primary

source of atmospheric carbon dioxide.

A significant scientific effort has been underway over the past decades to
more thoroughly understand the carbon cycle. These efforts have provided a
great deal of knowledge and understanding regarding the carbon cycle but
several questions remain unanswered. The carbon cycle can be represented by
the following equation:
A = F + B - O — [3 Units: Pg or 1015 g/year

A = Atmosphere, F = Fossil Fuels, B = Terrestrial Biota, O = Oceans, 13 = Error
Term or Missing Sink

Substituting values based on current research (Schimel et al., 1995) the carbon
cycle equation can be restated:
3.0 = 5.2 +1.8-2.0— 13

3.0 = 5.0 — I3
13 = 2.0 Pg of Carbon, the missing carbon sink

120

 

The error term in the preceding equation, the missing carbon sink, is an
amount of carbon that has not been readily accounted for. While the values
associated with the atmospheric, oceanic and fossil fuel components of the
carbon cycle are believed to be reliable, an accurate estimate of the terrestrial
component has thus far eluded researchers. The influence of widespread land
use/land cover change and the ensuing human management of these lands
(forestry, agriculture) has not been accounted for accurately enough to reach a
reliable estimate of mans influence on the terrestrial component of the carbon
cycle (Follett, 2001). The missing sink, many believe, may be a part of the
terrestrial biota term. The missing sink may be accounted for by enhanced
carbon sequestration rates in soils due to past and present management
practices, lengthening of the growing season, especially in the northern
hemisphere due to warming, CO; fertilization through elevated CO; levels in the

atmosphere, and sequestration of carbon in soils and sediments.

More detailed and accurate information regarding the carbon
consequences of land use/land cover change and subsequent land use
management practices are required to more accurately estimate the carbon
implications of such land use/land cover change. Conversion from forested land
covers to agriculture resulted in a large amount of carbon released to the
atmosphere. The carbon that was released during land use/land cover
conversion, by soils in particular, now have the potential to sequester carbon due

to changes in agricultural production methods. Prior to settlement and

121

cultivation, the long term balance of carbon stored in soils was dramatically
disrupted. The advent of large scale cultivation resulted in a large scale release

of soil organic carbon to the atmosphere (Follett, 2001).

Additional research has focused on what is termed the greening of the
north. Models by Dai and Fung (1993) suggest that the carbon sink is
predominately located in the northern middle latitudes (20-60° N). Dr. Fan and
other researchers have concluded that the missing sink is caused primarily by
increased terrestrial biosphere carbon storage (Fan et al., 1998). This has been
corroborated in part by satellite observations, which indicate that the growing
season for the northern mid latitudes has increased by several weeks (Myneni,
1996). Snow cover melts earlier in the spring and does not form until later in the

fall, a consequence of overall higher than average temperatures.

Feedback is another process suspected in increased carbon storage by
plants. Plants rely upon atmospheric carbon dioxide during photosynthesis;
elevated levels of carbon dioxide in the atmosphere due to anthropomorphic
activities may be responsible for increased plant growth. Increasing ambient
levels of carbon dioxide in laboratory studies has been shown to increase overall
plant growth. Increased plant growth leads to more carbon dioxide taken out of
the atmosphere and sequestered in plant materials. Research examining this
phenomenon indicates that vegetation has, in fact, experienced increased growth

in the northern latitudes due to increases in atmospheric carbon dioxide (Keeling

122

 

et al., 1996). These findings have been further substantiated by Tucker, who
examined the relationship between atmospheric carbon dioxide variations and

satellite derived vegetation indices (Tucker, 1986).

Houghton has noted “the rate at which carbon is accumulating in
terrestrial ecosystems in the United States is uncertain, as are the mechanisms

responsible for the current sink” (Houghton, 1999). Houghton based his research

 

on historic land use change, and related this data to carbon changes which are
direct results of land use change. The research considered the conversion of
natural ecosystems to cultivated cropland and pastures but did not take into
account the cropping method used for the agricultural production. Houghton
states, “the evidence is more compelling that carbon is accumulating in
agricultural soils as a result of changes in management but our book keeping

model did not consider these changes” (Houghton, 1999).

5.2 Regional Carbon Accounting

Conversion of natural ecosystems to managed agricultural production
during the 17‘“, 18th and 19th centuries has resulted in a dramatic loss of soil
organic matter due to mineralization and erosion. Since land use conversion
occurred, both vegetation and soils have undergone extensive changes (Lal and
Follett, 1999). Clearing, filling, and draining of these soils for long term
agricultural production has released large amounts of carbon dioxide to the

atmosphere from the soils’ fertile soil organic matter (Figure 21) (Lal and Follett,

123

1999). The soil organic carbon in topsoil often was depleted by up to half of its
original content (Cambardella, 1992). Lal and Follett (1999) estimated the loss of
soil organic content from US. cropland soils at about 5 Pg. Therefore, U.S.
cropland has a potential to sequester approximately 5 Pg of carbon through

improved soil management practices (Lal, 2000).

 

 

 

 

 

 

 

 

I» Effect on
amospherlc 1‘ ~ l.
g c0. ‘ . A ‘ ’
agriculture :
E \
E Soil C Soil c ‘ Soil C
3; Decrease Stable Increase
1 Inltlal Hananement
cultivation change

Figure 21. Agriculture and Carbon.
From Follett, 2001

Regional estimates of the carbon sequestration potential of conservation
tillage cropping practices are crucial if policy makers are to plan future land uses
to reduce national CO; emissions (Falloon, 1998). Questions related to the
distribution and spatio-temporal dynamics of the terrestrial carbon fluxes are at
the core of current scientific and policy debates (Cihlar, 2000):

Improved knowledge of the carbon cycle, its variability, and its likely future

state is essential. There are large uncertainties in the magnitudes and

locations of carbon fluxes between the land, atmosphere, and oceans. We

currently lack the understanding and observations needed to close the
annual carbon budget. It is not (currently) possible to unambiguously

124

determine the spatial distribution and longevity of carbon sinks at regional
to landscape scales, and previous attempts to do so have suffered from
inadequate data.
Atmospheric concentrations of carbon dioxide may be lowered by reducing
emissions or by taking carbon dioxide from the atmosphere via photosynthesis

and sequestering it in different components of terrestrial, oceanic, and freshwater

aquatic ecosystems (Bruce et al., 1999).

Soil organic matter represents a major pool of carbon within the
biosphere. It is estimated to contain 1550 Pg globally, which is roughly twice that
found in the atmosphere (Follett, 2001). Soil can act as both a source and a sink
for carbon and nutrients. Changes in agricultural land use and climate can lead to
changes in the amount of carbon held in soils, thus affecting the fluxes of 002
to and from the atmosphere. Carbon levels in soils are determined by the
balance of carbon inputs and outputs. Carbon inputs such as crop residues add
carbon to the soil balance. Over time, if carbon inputs are continual, soil carbon
stocks will build to a point where they outpace carbon loss, leading to net carbon
sequestration in soils. Conservation tillage methods have been shown to
sequester carbon in the soil over time, whereas conventionally tilled fields are

considered a net source of C02.

The management of crop residue plays a significant role in the fate of the

carbon stored in these residues. Under conventional tillage 60 - 80% of the

carbon stored in crop residues is released to the atmosphere as carbon dioxide

125

‘3;

Kin»
II ‘ . ,-
2

due to microbial soil respiration (Cavigelli, 2000). Conservation tillage practices
significantly slow decomposition rates, resulting in a net sequestration of carbon

in soils versus the net carbon loss found under conventional tillage practices.

5.3 Carbon and Tillage

The increase of atmospheric 002 is partly the result of increasing the
oxidation of the two main sources of organic carbon stored in the land - fossil
fuel (old organic carbon) and conventional tillage (recent organic carbon). The
studies by Robertson et al. (2000) at the Kellogg Biological Station (KBS) Long
Term Ecological Research site (LTER) has shown that fields that have never
been in agricultural production contain 40% to 50% more carbon by weight than
active agricultural fields. Agricultural fields under no-till conservation tillage
cropping methods were found to sequester 300 kg carbon per hectare per year,
whereas conventionally tilled crops exhibit no annual carbon sequestration
(Robertson et al., 2000). Other estimates of carbon sequestration rates for
conservation tillage range from 500 kg/Clha/yr to 600 kg/C/halyr (Lal and Follett,
1999). The United States Department of Energy also estimates the conversion
from conventional tillage to conservation tillage methods to potentially sequester

300 kg C/ha/yr.

Conservation tillage cropping methods result in improved carbon retention

due to the fact that less organic matter is lost to oxidation from mixing of the soil,

and soil temperatures tend to be lower, which slows oxidation rates (Uri, 1999).

126

“ii

Uri also notes conservation tillage cropping methods require less energy inputs
by function of fewer field activities, thereby reducing carbon emissions due to

decreased fuel use (Uri, 1999).

Robertson et al. (2000) state that agriculture plays a major role in the
global fluxes of greenhouse gasses. They note that radiative forcing of Earth’s

atmosphere is increasing at unprecedented rates, largely due to increases of

I

greenhouse gases. Carbon dioxide is produced anthropogenicly primarily t3
through the burning of fossil fuels, deforestation and conventional plowing.

Decaying plant material also produces carbon dioxide through the decomposition

process. Conservation tillage practices delay decomposition, which results in

increased soil carbon storage. Increased soil carbon leads to improved overall

soil quality and fertility. In the United States, half the organic matter in soils has

been oxidized by more than 100 years of conventional tillage (Falloon, 1998).

The depletion of soil carbon from agricultural lands has many farmers concerned

and is another associated benefit of conservation tillage practices that are

drawing farmers to the technique.

Two types of processes affect terrestrial carbon fluxes: relatively slow
accumulation over large areas in vegetation and soils, and the relatively rapid
release over small areas through disturbances (Cihlar et al., 1998). There are
numerous studies in the literature focusing on the rapid release of carbon over

small areas due to human disturbances such as deforestation, fire, and slash and

127

burn agriculture (Foody, 1997, Houghton, 1999). Regarding the relatively slow
accumulation over large areas in vegetation and soils, there are several articles
focusing on forestry, the “greening of the north”, and the impacts of agriculture
and carbon. However, while there are many articles regarding the potential of
agricultural soils to sequester carbon, none of these articles are able to
accurately quantify these estimates due to the fundamental lack of knowledge
regarding the spatial distribution and total acreage of conservation tillage

methods.

 

5.4 Study Area Total Carbon Sequestration

The carbon sequestration potential of the study area based on the cosine
of the angle concept classification, the most suitable classification method
identified in this study, is estimated currently at 228,490,877,995.98 grams

(7,616,362,599.87 square meters * 3Og/m2) or 228,490.87 metric tons.

The classified data indicate that conservation tillage cropping methods are
utilized on 1,881,241.56 acres (64% of major row crop agriculture) within the
study area. This estimate is based on the total acreage in the classified data and
the findings published by Robertson et al. (2000) of carbon sequestration rates of

30 g/mzlyear utilizing no-till conservation tillage techniques.

These estimates can be taken a step further to estimate the total carbon

sequestration rate for the study area if all corn and soybean acreage were to be

128

raised using conservation tillage cropping methods. For this estimate, the total
acreage of conventional and conservation tillage acreage were aggregated to
arrive at a total acreage of the study area and then multiplied by Robertson’s
estimates of carbon sequestration rates to arrive at a figure of potential carbon
sequestration of 359,065,898,717.7 grams (359,0658 Metric Tons). This
represents the overall carbon sequestration potential of agricultural land use/land

ha

cover across the study area.

 

 

 

5.5 Regional Estimates of Carbon Sequestration Based on Tillage
Agricultural production occurs on over 200 million acres of land use/land
cover across the United States, yet little data exists regarding the methods used
to produce crops and the associated benefits and costs related with each
cropping method. The methods and data presented here may prove useful for
analyzing the spatial distribution and extent of cropping methods and, by

extension, environmental and ecological impacts.

If the results found throughout the study area are representative of the
cropping methods across the Midwestern United States, then the findings here
may provide a useful starting point for assessing and possibly revising carbon
sequestration estimates currently available for agricultural soils. If the study area
is indeed representative of the larger Midwestern, then the 146 million acres
used in the production of corn and soybeans across the Midwestern United

States has the potential to sequester 17.725 million metric tons of carbon on a

129

yearly basis using a sequestration rate of 300 kg/hectare. The ranges of potential

yearly carbon sequestration rates are shown in Table 13.

 

. Potential C
Carbon SRequestration Sequestration (Metric
ate
Tons)
300 kg/C/ha/yr 17,725,23103
500 kg/C/ha/yr 29,542,051.72
600 kg/C/ha/yr 35,450,46206 r1

 

 

 

Table 13. Yearly Carbon Sequestration Potential

 

The carbon sequestration rates are based on published figures in the
literature, 300 kg/C/ha/yr (Robertson, 2000 and United States Department of
Energy), 500 and 600 kg/C/ha/yr (Lal, 1999) for conservation tillage practices as
compared to a net carbon sequestration rate of zero (0) for conventional tillage
practices. As Follett (2002) notes, even a small annual percent change in the
amount of carbon stored or released from large terrestrial carbon stocks could

easily affect the net change in atmospheric carbon dioxide.

5.6 Soil Erosion and Tillage

Soil erosion is a major environmental problem in the US and worldwide
(Nyakatawa et al., 2001). Soil erosion exhibits four primary effects on cropland:
nutrient loss, decreased water storage capacity, crop damage and decreased
yields, and ultimately if severe, field abandonment (Kohl, 1998). Soil loss due to
erosion under conventional tillage practices can occur at astounding rates of 21

tons/acre/year (Aber, Melillo, 2001). Conservation tillage cropping methods often

130

reduce the soil erosion rate to 0.8 tons/acre/year (Aber, Melillo, 2001). The
overall annual soil loss due to agriculture practices in the United States is
estimated by the USDA at 1.25 billion tons per year (USDA, 1991). With such
enormous erosional rates, it is not surprising that soil erosion is the largest

source of non point source pollution in the United States (USEPA, 1995).

Runoff from agriculture fields chokes streams with excessive sediment

loads, altering stream ecology and the streambed itself. Intensive row crop

 

agriculture has resulted in elevated amounts of suspended sediments, fertilizers,
and pesticides (Baker, 1993). When pesticides and herbicides are applied to
fields, these components are also washed away in the eroding soil, polluting
fresh water streams and lakes. Runoff containing fertilizers, when washed into
streams and lakes, can provide the nutrients necessary for algae blooms. The
growing algae deplete the water of oxygen creating a eutrophic environment
where little can live but algae and bacteria. Uri (1999) reports 70% less herbicide

runoff for conservation tillage versus conventionally tilled fields.

A United States Geological Survey (USGS) study investigating soil erosion
and tillage methods found a correlation between agricultural cropping methods
and suspended sediment loads in water bodies (USGS, 2000). The USGS
performed the study on the Maumee river basin, contained in the southwestern
portion of this study area, encompassing portions of Ohio, Michigan, and Indiana.

They found that sediment loads in water bodies are closely related to agricultural

131

cropping methods. They also noted that as conservation tillage has increased in
some areas the corresponding watershed units showed lesser amounts of
sediment loads in individual watershed units. Areas identified in the USGS study
as contributing the highest amounts of soil erosion to water bodies were areas

characterized by the lowest rates of conservation tillage.

Soil erosion associated with agricultural production represents a
tremendous loss of a valuable natural resource. Conventional tillage practices
create a soil surface highly susceptible to erosional processes. Loss of topsoil
due to erosional action may eventually lead to loss of production capabilities if
left unchecked. In addition to loss of productivity, the eroded soil significantly
impacts overall water quality in water bodies across the United States. The

eroded soils also play a role in the carbon budget as well.

The suspended sediment load carried through the nations watenivays is
deposited in the waterway itself or ultimately to the oceans. The redistribution of
carbon through erosion and deposition, which effectively sequesters carbon, is
yet another portion of the carbon cycle that has yet to be fully accounted for. The
spatial distribution and the total area of conservation tillage and conventional
cropping methods by watershed may provide data to more accurately quantify

the erosional potential of agricultural lands.

132

 

5.7 Scaling Up - Regional Measurements and Estimates

In order to fully take advantage of the findings of this study, the results
should be scaled up to a regional level encompassing the Midwestern states that
account for the nations vast majority of row crop agriculture. Scaling the results
to larger geographic areas would provide for additional data regarding the spatial
distribution of cropping methods including both total area and the spatial

distribution of cropping methods.

Data regarding the total area of each cropping method could then be used
to more accurately derive carbon sequestration potentials of agricultural lands.
Data regarding the spatial distribution of cropping methods could be used in a
wide variety of studies investigating soil erosion and, in turn, water quality.
Perhaps watersheds with higher percentages of conservation tillage acreage can
be correlated with water quality measurements. Knowledge of the spatial
distribution and total acreage under each cropping method may also lead to more
accurate estimates of soil erosion estimates. An enhanced understanding of the
total extent and spatial distribution of cropping methods at larger regional scales

holds a great deal of promise for a variety of ecosystem studies.

Data sources and analytical procedures used in this study were based not
only on best methodology, but also ensured that the study could be repeated by
other researchers with a minimum amount of specialized software and at

reasonable cost. Landsat 7 ETM+ data are arguably the most widely available

133

high resolution and cost efficient satellite based source of remotely sensed data
currently available. Additionally, the analytical procedures used in this study rely
upon industry standard software and hardware configurations widely available in

research, industry, and governmental settings.

The findings of this research indicate that the mapping of agricultural
crops and cropping methods is possible over entire Landsat ETM+ scenes with
only two ETM+ images per season. The first image, taken after field preparation
activities in the spring, as illustrated by this study, can be used to differentiate
cropping methods. Adding a temporal component, a second image acquired later
in the growing season towards peak plant development, in conjunction with the

first image, allows for differentiation between common agricultural crops.

The first image obtained early in the season, after field preparation
activities, provides data regarding cropping method: conventional tillage vs.
conservation tillage. The first image also contains data regarding other
secondary crops. While it is too early in the season to distinguish corn acreage
from soybean acreage in this early image, the data contains information
regarding winter wheat and alfalfa acreage. Winter wheat acreage is in peak
growth at this time, and exhibits a very strong healthy green vegetation

reflectance pattern. The only other field crop in the region that may be in
L/egetative growth at this point in time is alfalfa, which can be differentiated from

\Jmmt wheat based on spectral reflectance properties. The first image provides

134

data regarding cropping method for corn/soybean acreage as well as acreage for
winter wheat, and alfalfa acreage. The second image, obtained later in the
growing season, could then be used in conjunction with the earlier image to
determine corn and soybean acreage as well as alfalfa and winter wheat (Table
14). Utilizing crop phenological development with a temporal data collection
scheme allows for complete characterization of crop and cropping method for the

major agricultural crops commonly found throughout the Midwestern United

 

 

States.

Image 1 Image 2
Corn Bare soil/Crop Residue Peak Growth
Soybean Bare soil/Crop Residue Peak Growth
Winter Wheat Peak Growth Senescent/Bare Soil
Alfalfa Early Growth Variable

 

Table 14. Temporal Image Acquisition for Crop Identification.

5.8 Further Research

The findings and conclusions of this research have resulted in the
identification of suitable classification methods to differentiate between
conventionally tilled crops and those raised using conservation tillage methods.
However, the topic has not yet been exhausted, and research potential still exists

for improvement and further knowledge.

The differentiation between conventional tillage and conservation tillage

cropping practices is possible using commonly available data sources, software,

135

 

and analytical methods. To successfully apply these findings to other areas
throughout the Midwestern United States several considerations concerning

scale and timing must be carefully examined to ensure a successful outcome.

The spatial resolution of the remote sensing instrument must be carefully
considered to ensure that the phenomena of interest is accurately captured. The
findings presented here indicate that 30 meter spatial resolution is sufficiently
detailed to capture the phenomena of interest (conventional vs. conservation
tillage cropping methods) while sufficiently large to enable data acquisition over
large spatial areas. Additionally, the data source (Landsat 7 ETM+ data) is widely
available to the research community and the public at a reasonable cost. The
possibility of using other data sources with differing spatial resolutions, either
lesser or greater, has not been specifically addressed, and these scaling issues
should be examined if these techniques and methods are to be applied using

different data sources.

The use of other sources of remotely sensed data may yield improved
classification accuracies due to improved spatial, spectral, or radiometric
resolutions of newer sensors. Much of the data from new experimental satellites
and remote sensing instruments are not yet available to the research community.
Advanced remote sensing instruments with improved resolutions (spectral,
radiometric, temporal) may provide more information regarding the target of

interest, thus leading to improved classification results.

136

Additional studies regarding scaling issues are warranted, based on the
conclusions of hand held and satellite based sensor readings. The interaction of
electromagnetic energy with the target of interest and the scale at which the
reflected electromagnetic energy is measured shows a dependent relationship. A
study utilizing a wide variety of spatial resolutions (4, 10, 30, 60, 120 meters)
from a variety of satellite based sensors would provide basic data regarding ’“

scaling effects on the complex interaction of reflection, scale, and texture.

 

Further studies utilizing these techniques should also consider the broad
range of moisture conditions that could be encountered in future research. While
spectral angle mapping and the cosine of the angle concepts take into account
variations in illumination due to a broad range of conditions, including in theory
moisture, additional research to ensure tillage methods are differentiatable under
a wide range of moisture conditions is warranted. The possibility exists that the
soil and crop residue moisture conditions at the time of imagery collection for this
study were ideal. Data acquired under differing soil and crop residue moisture
conditions may have an impact on classification results. For example a rainfall
event just prior to imagery collection may result in wet or moist soils and
relatively dry crop residues. While the classification methods utilizing spectral
angle concepts should theoretically account for such variations further
investigation and caution is warranted before relying upon these methods and

techniques under differing conditions.

137

The possibility also exists for examining historical change over time.
Landsat data are available from 1972, providing a rich historical record of land
use/land cover change. An assessment to determine the suitability of using
earlier Landsat remote sensing systems to study cropping methods is warranted.
If earlier Landsat based systems are found suitable for the differentiation of
cropping methods it would provide valuable data regarding the past spatial
distribution and extent of cropping methods and their associated environmental
implications. Archived data regarding atmospheric carbon dioxide concentrations
or stream flow sediment loads could be linked to past land use/land cover
change to more accurately assess the environmental impacts of intensive row

crop agriculture production methods.

138

 

 

BIBLIOGRAPHY

Aase, J.K., Tanaka, D.L., 1991, Reflectance from four wheat residue densities as
influenced by three soil backgrounds, Agronomy Journal, vol. 83, 753-757.

Aber, J.D., Melillo, J.M., 2001, Terrestrial Ecosystems, Harcort Academic Press,
New York.

Albert, Dennis A. 1995. Regional landscape ecosystems of Michigan, Minnesota,
and Wisconsin: a working map and classification. Gen. Tech. Rep. NC—178.
St. Paul, MN: US. Department of Agriculture, Forest Service.

Allmaras, R.R., Schomberg, H.H., Douglas, C.L., Dao, TH, 1999, Conservation

tillage’s unforeseen advantage. Research Engineering Technology, vol. 6, pp 7-
8.

Appenzeller, C., Stocker, T.F., Anklin, M., 1998, North Atlantic oscillation
dynamics recorded in Greenland ice cores, Science, vol. 282, 446-448.

Atkinson, P.M., Curran, P.J., 1997, Choosing an appropriate spatial resolution for
remote sensing investigations. Photogrammetric Engineering and Remote
Sensing, 63, 1345-1351.

Atkinson, P.M. Tatnall, A.R.L., 1997, Neural networks in remote sensing.
International Journal of remote sensing, vol. 18, no. 4, 699-709.

Baker, DB, 1993, Lake Erie agroecosystem program — Water quality
assessments: Agriculture, Ecosystems, and Environment, v. 46, p. 197-215.

Baker, D.B., Calhoun, F.G., Matisoff, G., 1998, LEASEQ, the Lake Erie
agricultural systems for environmental quality project: Fact Sheet, 2 p.

Baker, D.B., Calhoun, F.G., Matisoff, G., 1998, LEASEQ, the Lake Erie
agricultural systems for environmental quality project.

Batjes, N.H., 1996, Total C and N in soils of the world, European Journal Soil
Science, vol. 47 pp 151-163.

Batjes, N. H., 1998, Mitigation of atmospheric carbon dioxide concentrations by
increased carbon sequestration in the soil, Biological Fertility and Soils, vol. 27,
230-235.

Boyd, R., Uri, ND, 2001, A note on the use of conservation practice in US.
agriculture, Environmental monitoring and assessment, vol. 72, 141-178.

139

 

Bricklemyer, R., Lawrence, R., Miller, P., 2002, Documenting no-till and
conventional till practices using Landsat ETM+ imagery and logistic regression,
Journal of Soil and Water Conservation, vol. 57, no. 5, 267-271.

Bruce, J.P., Frome, M., Haites, E., Janze, H., Lal, R., Paustian, K., 1999, Carbon
sequestration in soils, Journal of Soil Water Conservation, vol. 54, pp 382-389.

Cavigelli, M.A., Mutch, DR, 2002, Cover Crops and Carbon, Michigan State
University Extension, Info-Sheet.

Campbell, J., 1987, Introduction to Remote Sensing, New York: Guilford Press.

Cihlar, J., Xiao, Q., Chen, J., Beaubien, J., Fung, K., Latifovic, R., 1998,
Classification by progressive generalization: a new automated methodology for

remote sensing multichannel data. International Journal of Remote Sensing, vol.
19, no. 14, 2685-2704.

Cihlar, J., 2000, Land cover mapping of large areas from satellites: status and
research priorities, International Journal of Remote Sensing, vol. 21, no 6 &7,
1093-1 1 14.

Cipra, J.E., Baumgardner, M.F., Stoner, E.R., MacDonald, RB, 1971, Measuring
radiance characteristics of soil with a field spectroradiometer. Soil Science
Society of America Proceedings, 35, 1014-1017.

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

Cortijo, F., and Perez de la Blanca, N., 1998, Improving classical contextual
classifications. International Journal of Remote Sensing, vol. 19, no. 8, 1591-
1613.

Coulson, KL, Reynolds, D.W., 1971, The spectral reflectance of natural
surfaces. Journal of Applied Meteorology, vol. 10, 1285-1295.

Dai, A., F ung, I.Y., 1993, Can climate variability contribute to the missing carbon
dioxide sink?, Global Biogeochemical cycles, vol. 7 599-609.

Daughtry, C.S.T., J.E. McMurtrey J.E., E.W. Chappelle, W.J. Hunter and J.L.
Steiner, 1996, Measuring Crop Residue Cover Using Remote Sensing
Techniques, Theoretical and Applied Climatology, 54, 17-26.

Daughtry, C.S.T., McMurtrey, J.E., Dulaney, W.P., Irons, J.R., Sattervvhite, MB,

1995, Potential for discriminating crop residues from soil by reflectance and
fluorescence, Agroclimatology, Agronomy Journal, 87:165-171.

140

‘2

Elvidge, DD, 1990, Visible and near infrared reflectance characteristics of dry
plant materials. International Journal of Remote Sensing, 11, 1775-1795.

Erol, H., Akdeniz, F ., 1998, A new supervised classification method for
quantitative analysis of remotely-sensed multi-spectral data. International Journal
of Remote Sensing, vol. 19, no 4, 775-782.

Falloon, PD, 1998, Regional estimates of carbon sequestration potential: linking
the Rothamsted carbon model to GIS databases. Biological Fertility of Soils, vol.
27, 236-241.

Fan, 8., M. Gloor, J. Mahlman, S. Pacala, J. Sarmineto, T. Takahashi, P. Tans,
1998, A large terrestrial carbon sink in North America implied by atmospheric and
oceanic carbon dioxide data and models. Science, vol. 28, 442-446.

Field, C.B., Fung, I.Y., 1999, The not so big U.S. carbon sink, Science, vol. 285,
544-545.

Foody, GM, 1997, An evaluation of some factors affecting the accuracy of
classification by an artificial neural network. International Journal of Remote
Sensing, vol. 18, no 4, 799-810.

Foody, G.M., Palubinskas, G., Lucas, R.M., Curran, P.J., Honzak, M., 1996,
Identifying terrestrial carbon sinks: classification of successional stages in
regenerating tropical forest from Iandsat TM data, Remote Sensing of the
Environment, vol. 55, 205-216.

Follett, R.F., Paul, E.A., Leavitt, S.W., Halvaorson, A.D., Lyon, D., Peterson,
GA, 1997, The determination of the soil organic matter pool sizes and
dynamics. Soil Science Society of America Journal, vol. 61, 1068-1077.

Follett, R.F., McConkey, B., 2000, The role of cropland agriculture for carbon
sequestration in the Great Plains. Proceedings of the conference on Great Plains
soil fertility, vol. 8, 1 ~15.

Follett, R.F., Samson-Liebig, S., Kimble, J.M., Pruessner, E.G., Waltman, S.W.,
2001, Carbon sequestration under the CRP in the historic grassland soils in the
USA, Soil Carbon Sequestration and the Greenhouse Effect, SSSA Special
Publication, No. 57, Madison, WI.

Follett, R.F., 2001, Soil management concepts and carbon sequestration in
cropland soils, Soil and Tillage Research, vol. 61, 77-92.

141

 

Gausman, H.W., Gerbermann, A.H., Wiegand, C.L., Leamer, R.W., Rodriguez,
RR, 1975, Reflectance Differences Between Crop Residues and Bare Soils,
Soil Science Society of America Proceedings, vol. 39, 752 - 755.

Goward, S., Masek, J., Williams, 0., Irons, J., Thompson, R., 2001, The Landsat
7 mission Terrestrial research and applications for the 21St century. Remote
Sensing of the Environment, vol. 78, 3-12.

Houghton, R.A., Davidson, E.A., Woodwell, GM, 1998, Missing sinks,
feedbacks, and understanding the role of terrestrial ecosystems in the global
carbon balance. Global Biogeochemical Cycles, vol. 12, No. 1, 25-34.

Houghton, R.A., Hackler, J.L., Lawrence, K.T., 1999, The US. Carbon budget:
Contributions from land-use change, Science, vol. 285, 574-578.

Henderson-Sellers, A., Wilson, M.F., Thomas, G., 1985, The effect of
spatial resolution on archives of land cover type, Climatic Change, vol.
7, 391-402.

Jahne, B., 1997, Practical handbook on image processing for scientific
applications, CRC Press, Boca Raton, NY.

Janzen, H.H., Campbell, CA, lzaurralde, R.C., Ellert, B.H., Juma, N., McGilI,
W8 and Zentner, RP, 1998. Management effects on soil C storage on the
Canadian prairies. Soil Till. Res. 47, pp. 181—195.

Kaminsky, E., Barad, H., Brown, W., 1997, Textural neural network and version
space classifiers for remote sensing. International Journal of Remote Sensung,
vol. 18, no. 4, 741-762.

Kanellopoulos, l., and Wilkinson, G., 1997, Strategies and best practice for
neural network image classification. International Journal of Remote Sensing,
vol. 18, no. 4, 711-725.

Kartikeyan, B., Sarkar, A., Majumder, K., 1998, A segmentation approach to
classification of remote sensing imagery. International Journal of Remote
Sensing, vol. 19, no. 9, 1695-1709.

Keeling, R.F., Piper, SC. and Heimann, M., 1996, Global and atmospheric C02
sinks deduced from changes in atmospheric 02 concentration, Nature, vol. 381,
218-221.

Keeling, C.D., Chin, J.F.S., Whorf, T.P., 1996, Increased activity of northern

vegetation inferred from atmospheric carbon dioxide measurements, Nature, vol.
382, 146-149.

142

sz‘flh‘

“W. T.‘ ' L

 

 

Kimberlin, L.W., Moldenhauer, WC, 1977, Predicting soil erosion, Proceedings
of the national symposium on soil erosion and sedimentation by water: Chicago,
IL., Dec 12 —13, 1977, American Society of Agricultural Engineers, Publ. 4 - 77.

Kohl, Kris, 1998, Conservation Tillage - Effects on soil erosion, Department of
agricultural and biosystems engineering, Iowa State University, Extension
Publication # AE-3050.

Lal, R., Follett, R.F., 1999, Managing U.S. cropland to sequester carbon in soil.
Journal of Soil and Water Conservation, First Quarter, 374 — 381.

Lal, R., Kimble, J.M., 2000, Soil carbon sink in US. cropland.

LPSO (Landsat Project Science Office), 1998, Landsat 7 Science Data Users
Handbook,http://ltpwww.gsfc.nasa.govllAS/handbooklhandbook_htmls/chapter5/
chapter5.html#section5.4.

Magnussen, S., 1997, Calibrating photo interpreted forest cover types and
relative species compositions to their ground expectations, Canadian Journal of
Forest Research, vol. 27, 491-500.

Montserud, R.A., Leamans, R., 1992, Comparing global vegetation maps with
the kappa statistic. Ecological Modeling, vol. 62, 275-293.

Moulin, S., 1998, Combining agricultural crop models and satellite observations:
from field to regional scales. International Journal of Remote Sensing, vol. 19,
no. 6, 1021-1036.

Mural, H., and Omatu, S., 1997, Remote sensing image analysis using a neural
network and knowledge-based processing. International Journal of Remote
Sensing, 1997, vol. 18, no. 4, 811-828.

Myers, D.N., Metzker, K.D., Davis, 8., 2000, Status and trends in suspended
sediment discharges, soil erosion and conservation tillage in the Maumee river
basin — Ohio, Michigan and Indiana, US. Department of Agriculture, Natural
Resources Conservation Service, Lima, Ohio.

Myneni, R.B., Keeling, C.D., Tucker, C.J., Asrar, G., Nemani, RR, 1997,
Increased plant growth in the northern high latitudes from 1981 to 1991, Nature,
vol. 386, 698-702.

Myneni, R.B., Tucker, C.J., Asrar, G., Keeling, CD, 1998, lnterannual variations
in satellite-sensed vegetation index data from 1981 to 1991, Journal of
Geophysical Research, vol. 103, 6145-6160.

143

Nagler, P.L., Daughtry, C.S.T., Goward, SN, 2000, Plant Litter and Soil
Reflectance. Remote Sensing of the Environment, vol. 71, 207-215.

Nyakatawa, E.Z., Reddy, K.C., Lemunyon, J.L., 2001, Predicting soil erosion in
conservation tillage cotton production systems using the revised universal soil
loss equation. Soil and Tillage Research, vol. 57, 213—224.

Parsch, L.D., Keisling, T.C., Sauer, P.A., Oliver, L.R., Crabtree, NS, 2001,
Economic analysis of conservation and conventional tillage cropping systems on
clayey soil in eastern Arkansas, Agronomy Journal, vol. 93, 1296-1304.

Post, W.M., Peng, TH, 1998, The Global Carbon Cycle. American Scientist, vol.
78, 310-326.

Reicosky, DC. 1996. Tillage as a mechanism for CO; emission from soils.
Proceedings of International Symposium on Soil--Source and Sink of
Greenhouse Gases, 14-20 September, 1995. Nanjing, China.

Richards, J.A., Jia, X., 1999, Remote Sensing Digital Image Analysis,
Springer, New York.

Robertson, G.P, Eldor, P., HanNood, R., 2000, Greenhouse Gases in intensive
Agriculture: Contributions of Individual Gases to the Radiative Forcing of the
Atmosphere. Science, vol. 289, 1922 — 1925.

Schimel, D., Alves, D., Enting, l., Heimann, M., 1995, Carbon dioxide and the
carbon cycle, Climate Change.

Sohn, Y., Moran, E, Gurri, F., 1999, Deforestation in north-central Yucatan
(1985-1995): Mapping secondary succession of forest and agricultural land use
in Sotuta using the cosine of the angle concept. Photogrammetric Engineering
and Remote Sensing, vol. 65, No. 8, 947-958.

Sohn, Y., Rebello, NS, 2002, Supervised and unsupervised spectral angle
classifiers. Photogrammetric Engineering and Remote Sensing, vol. 68, No. 12,
pp 1271 - 1280.

Stoner, E.R., Baumgardner, M.F., 1980, Extension of laboratory measured soil
spectral to field conditions. Soil Science Society of America Journal, vol. 44, 572-
574.

Stoner, E.R., Baumgardner, M.F., 1981, Characteristic variations in reflectance of
surface soils. Soil Science Society of America Journal, vol. 45, 1161-1165.

Swain, P.H., Davis, SM, 1978, Remote Sensing: The Quantitative Approach,
N.Y., McGraw-Hill.

144

 

Tans, P.P., Fung, Y., Takahashi, T., 1990, Observational constraints on the
global atmospheric carbon dioxide budget. Science, vol. 247, 1431 - 1438.

Tucker, D.J., Fung, I.Y., Keeling, C.D., Gammon, RH, 1986, Relationship
between atmospheric carbon dioxide variations and a satellite derived vegetation
index, Nature, vol. 319, 195-199.

Uri, ND, 1999, The environmental benefits and costs of conservation tillage.
Environmental Geology, Vol. 38, no. 2.

Uri, ND, 2001, Conservation practices in US. Agriculture and their impact on
carbon sequestration. Environmental monitoring and assessment, vol. 70, 323-
344.

USDA, 1991, Agricultural Statistics, National Agricultural Statistics Service, US
Government Printing Office, Washington, DC.

USDA 1991, Yearbook of Agriculture, Agriculture and the environment, US
Government Printing Office, Washington DC.

US. Department of Agriculture, 1991, State soil geographic data base
(STATSGO): Soil Conservation Service Miscellaneous Publication 1492.

USDA, 1993, Erosion and sedimentation dynamics of the Maumee River Basin
and their impact on Toledo Harbor: Columbus, Ohio, Soil conservation service,
publication 1492.

USDA, 2000, Agricultural Statistics, National Agricultural Statistics Service, US
Government Printing Office, Washington, DC.

USDA soil taxonomy
http://soils.usda.gov/classification/taxonomy/maps.pdf

USEPA, 1995, Non Point Source Agricultural Pollution, USEPA, Washington,
DC.

USGS, 1998, Regional landscape ecosystems of Michigan, Minnesota, and
Wisconsin, Northern Prairie Wildlife Research Center.

USGCRP, 2003, www.usgcrp.org

Vitousek, PM, 1994, Beyond Global Warming: Ecology and Global Change.
Ecology, vol. 75(7), 1861-1878.

145

 

 

Wilson, D.L., 2001, GPS WAAS Accuracy,
http:/lusers.erols.com/dlwilson/gpswaas.htm.

146

 

IIIIIIIIIIIIIIIIIIIIIIIIIIIIIII

Illlljjlljljllljllllljlll