NCEA-C-1698
April 2006
Classification of High Spatial
Resolution, Hyperspectral Remote
Sensing Imagery of the Little Miami
River Watershed in
Southwest Ohio, USA
National Center for Environmental Assessment
Office of Research and Development
U.S. Environmental Protection Agency
Cincinnati, OH 45268
-------
NOTICE
The U.S. Environmental Protection Agency through its Office of Research and
Development funded and managed the research described here under contract no.
68-C-02-060 to Forest One, Inc. and 3C-R337-NTSA to Hyperspectral Data
International, Inc. It has been subjected to the Agency's peer and administrative review
and has been approved for publication as an EPA document. Mention of trade names
or commercial products does not constitute endorsement or recommendation for use.
ABSTRACT
This report summarizes a collaborative project led by the U.S. Environmental
Protection Agency to create a high spatial resolution land use/land cover (LULC)
dataset for the entire Little Miami River watershed in southeastern Ohio, USA from
remotely sensed imagery. The LULC classification was derived from 82 flight lines of
Compact Airborne Spectrographic lmager(CAS\) hyperspectral imagery acquired from
July 24 through August 9, 2002 via fixed wing aircraft. Categories within this
classification include: water (both lentic and lotic), forest, corn, soybean, wheat, dry
herbaceous vegetation, grass, urban barren, rural barren, urban/built, and unclassified.
A hierarchical classification approach was used involving object image segmentation in
eCognition (Definiens Imaging GmbH., 2003), and spectral angle mapper (SAM) in the
ENvironment for Visualizing Images (ENVI) (Research Systems Inc., 2003). A final
classification was completed after an extensive Quality Assurance and Quality Control
(QA/QC) phase which included manual editing. The final product includes classification
results at the original data spatial resolution of 4m x 4m.
Preferred citation:
Troyer, M.E., J. Heo and H. Ripley. 2006. Classification of High Spatial Resolution, Hyperspectral
Remote Sensing Imagery of the Little Miami River Watershed in Southwest Ohio, USA. U.S.
Environmental Protection Agency, Office of Research and Development, National Center for
Environmental Assessment, Cincinnati, OH.
-------
TABLE OF CONTENTS
LIST OF TABLES iv
LIST OF FIGURES v
LIST OF ABBREVIATIONS vi
AUTHORS, CONTRIBUTORS AND REVIEWERS vii
ACKNOWLEDGEMENTS ix
PREFACE x
1. PROJECT DESCRIPTION 1
2. THE REMOTE SENSING EFFORT 3
2.1. INTRODUCTION 3
2.2. RADIOMETRIC CORRECTION 3
2.3. GEOMETRIC CORRECTION 8
2.4. ATMOSPHERIC CORRECTION 8
3. IMAGE CLASSIFICATION 9
3.1. INTRODUCTION 9
3.2. THE CLASSIFICATION PROCESS 9
3.2.1. Quality Review of the CASI Data for Purposes of Classification 9
3.2.2. Ground Truth 13
3.2.3. Classification Method 14
3.2.4. Consistency of Classification Across Flightlines 23
4. POST-PROCESSING AND QA/QC 24
4.1. MANUAL EDITING 24
4.2. MAP GENERALIZATION 25
4.3. ACCURACY ASSESSMENT 25
5. RESULTS AND DISCUSSION 29
6. REFERENCES 31
APPENDIX A: FLIGHT LOGS OF THE CASI DATA COLLECTION IN THE
LITTLE MIAMI RIVER WATERSHED A-1
APPENDIX B: PROTOCOL FOR COLLECTING GROUND TRUTH LAND
COVERS FOR THE LITTLE MIAMI RIVER WATERSHED 2002 B-1
APPENDIX C: GENERAL RULES FOR OBTAINING TRAINING SET AND
SUPPLEMENTAL GROUND TRUTH DATA C-1
APPENDIX D: METADATA D-1
-------
LIST OF TABLES
No. Title
1 Spectral Bands Remotely Sensed in the Little Miami River Project 5
2 Projection for the Little Miami River Land Use/Land Cover 22
3 Classification Error Matrix 26
4 Summary Statistics of the Accuracy Assessment 27
IV
-------
LIST OF FIGURES
No. Title
1 Orientation Map of the Little Miami River Watershed in Southwestern
Ohio, USA 2
2 Remote Sensing Instrumentation Used in this Project 4
3 Flight Lines and Acquisition Dates of CASI Data in the Little Miami
River Watershed 7
4 An Unclassified Gap Between Flight Lines 13a and 13b, Acquired on
July 25, 2002 10
5 Cross-track Illumination Artifacts from Flight Line 38 Acquired on
July 24, 2002 12
6 Cross-track Illumination Artifacts from Flight Line 10 Acquired on
July 25, 2002 12
7 Image Object Segmentation versus Pixel-Based Classification of a
Forest Area in the Little Miami River Watershed 16
8 Flowchart of the Classification Method 18
9 Schematic of the Hierarchical Classification and Derived Classes 19
10 Legend for the Final 4m x 4m LULC Product 20
11 Overview of the 4m x 4m LULC Classification for the Little Miami
River watershed 21
-------
LIST OF ABBREVIATIONS
ACORN Atmospheric CORection Now
BIP Band Interleaved-by-Pixel
CART Classification and Regression Tree
CASI Compact Airborne Spectrographic Imager
DN Digital Number
ENVI ENvironment for Visualizing Images
FGDC Federal Geographic Data Committee
GMT Greenwich Mean Time
GPS Global Positioning System
HDI Hyperspectral Data International
IFOV Instantaneous Field of View
LULC Land Use/Land Cover
MLC Maximum Likelihood Classification
NLCD National Land Cover Dataset
PPS Pulse-Per-Second
QA/QC Quality Assurance and Quality Control
RMS Root-Mean-Square
SAM Spectral Angle Mapper (classification in ENVI)
VI
-------
AUTHORS, CONTRIBUTORS AND REVIEWERS
AUTHORS
Michael E. Troyer, Ph.D.
National Center for Environmental Assessment
Office of Research and Development
U.S. Environmental Protection Agency
Cincinnati, OH 45268
Joon Heo, Ph.D.
Forest One, Inc.
Itasca, IL
Herbert Ripley, FRSPSoc
Hyperspectral Data International, Inc.
Halifax, Nova Scotia
CONTRIBUTORS
Christopher Bolton
Earth Satellite Corp.
Rockville, MD
Michael Diller
Earth Satellite Corp.
Rockville, MD
Christopher Jengo
Earth Satellite Corp.
Rockville, MD
William Jones
Hyperspectral Data International, Inc.
Halifax, Nova Scotia
Sitansu Pattnaik
Forest One, Inc.
Itasca, IL
Laura Roy
Hyperspectral Data International, Inc.
Halifax, Nova Scotia
Francois Smith
Earth Satellite Corp.
Rockville, MD
Michelle Warr
Hyperspectral Data International, Inc.
Halifax, Nova Scotia
INTERNAL PEER REVIEWERS
F. Bernard Daniel, Ph.D.
Senior Environmental Scientist
National Exposure Research Laboratory
Office of Research and Development
U.S. Environmental Protection Agency
VII
-------
AUTHORS, CONTRIBUTORS AND REVIEWERS cont.
NoelW. Kohl, M.P.H.
EPA CIS Working Group
Office of Information Services
U.S. Environmental Protection Agency
Region 5
Chicago, IL
Cheryl G. Itkin
EPA CIS Working Group
Information Management
National Center for Environmental Assessment
Office of Research and Development
U.S. Environmental Protection Agency
Washington, DC
Tara A. Maddock, Ph.D.
Geographer
National Center for Environmental Assessment
Office of Research and Development
U.S. Environmental Protection Agency
Mary L. White, Ph.D.
EPA CIS Working Group
Water Quality Branch, Water Division
U.S. Environmental Protection Agency
Region 5
Chicago, IL
EXTERNAL PEER REVIEWERS
Nina Lam, Ph.D.
R.J. Russell Professor of Geography
Louisiana State University
Baton Rouge, LA
Victor Mesev, Ph.D.
Associate Professor of Geography
Florida State University
Tallahassee, FL
R. Douglas Ramsey, Ph.D.
Associate Professor and Director of the Remote Sensing and CIS Laboratory
Department of Forest, Range, and Wildlife Sciences
Utah State University
Logan, UT
VIM
-------
AUTHORS, CONTRIBUTORS AND REVIEWERS cont.
ACKNOWLEDGMENTS
Many people provided their time and talents to ensure the success of this project.
Essential fieldwork and ground-truth data were provided by the following scientists from
the U.S. Environmental Protection Agency's research facilities in Cincinnati, Ohio: Drs.
F. Bernard Daniel, Michael B. Griffith, James M. Lazorchak, Tara Maddock, Matthew A.
Morrison, Bruce W. Peirano, Joseph P. Schubauer-Berigan, Christopher Schultz and
William Shuster. Peter Mueller and Drew Comley (Dr. Troyer's summer interns) also
assisted in these field efforts as participants of the "Research Apprenticeship Program"
sponsored by the University of Cincinnati and the U.S. EPA. Additional aerial imagery
used in this project was generously provided by Dr. Lawrence Spencer from the Ohio
State University - Center for Mapping. Early discussions on remote sensing,
classification and choosing the proper spectral bands and widths for this project were
substantively informed by conversations with Dr. Prasad S. Thenkabail of the Center for
Earth Observation at Yale University, and David J. Williams of the U.S. EPA's
Environmental Services Division in Reston, Virginia. Appreciation also goes to Bette
Zwayer (U.S. EPA National Center for Environmental Assessment), Lana Wood
(Intellitech Systems, Inc.) and Elizabeth Fryer (ECflex) who helped with preparing
tables, and formatting and editing the document into its final form.
IX
-------
PREFACE
Geographic products such as the land use/land cover (LULC) dataset presented
herein enable a wide array of studies important for sustaining both society and nature.
For example, opportunities and risks associated with various human-ecological
interactions can be identified, as well as a greater elucidation of other spatial
relationships and processes affecting the Earth's surface and our environment.
In 2001, an interdisciplinary group of scientists based at the U.S. Environmental
Protection Agency's research facilities in Cincinnati, Ohio formed a collaborative effort
to study the Little Miami River (LMR) and its watershed. This work was performed in
support of the U.S. EPA's watershed (or geographic) approach for protecting
designated uses of America's rivers and streams. In particular, development of this
LULC dataset provides a basis for studying: (1) social drivers of land use change, (2)
how land use change affects the hydrology, sediments and nutrients of streams, and (3)
aquatic biological responses as a result of changes in the landscape. This work also
assisted in exploring and suggesting refinements to spatially-explicit criteria intended to
prevent habitat alteration, excess nutrients, suspended and bedded sediments,
pathogens, toxic chemicals, and other stressors affecting the Nation's waters.
A primary goal of this project is to help enhance the use of geographic and
spatial analytic tools in risk assessments at U.S. EPA, and to improve the scientific
basis for risk management decisions. This is important because environmental
problems are inherently spatial. For example, many pollutants originate from multiple
non-point sources in the landscape and spread to other areas within a particular
watershed, "airshed," or across ecological and political boundaries. From a
technological point of view, applications of this product will represent an improvement
from more readily available data with coarser spatial and spectral resolution. The
hierarchical classification approach used incorporating both object-based pattern
recognition and spectral techniques may be beneficial and transferable to other settings.
This dataset may also be useful to others outside the Agency; particularly, those
interested in studying anthropogenic and natural processes occurring at watershed or
smaller spatial scales.
-------
1. PROJECT DESCRIPTION
The goal of this project was to create a higher spatial resolution land use/land
cover (LULC) dataset for the entire Little Miami River watershed in southeast Ohio for a
number of studies being conducted by the U.S. Environmental Protection Agency (U.S.
EPA). The Little Miami River has a drainage area of 175.5 square miles and stretches
in a southwesterly direction for 105.5 miles originating from near South Charleston,
Ohio to its confluence with the Ohio River east of Cincinnati, Ohio (Figure 1). It is one
of the oldest river groups in the state and became Ohio's first State and National Scenic
River (Sanders, 2002). Prior to this project, existing LULC datasets based on Landsat
imagery had spatial resolutions of 30 to 60 meters. Hyperspectral Data International
(HDI) collected 4 meter spatial resolution, hyperspectral imagery of the watershed from
July 24 through August 9, 2002. Forest One Incorporated (Earth Satellite Corporation)
subsequently classified the resulting 82 flight-lines of Compact Airborne Spectrographic
Imager (CASI) data into the 11 classes of land cover type presented here.
-------
Springfield-]
South Charleston
Xenia
Wilmington
Little Miami River
East Fork Little Miami River
Little Mia mi River
Watershed
Ohio. USA
A
60 Mies
FIGURE 1
Orientation Map of the Little Miami River Watershed
in Southwestern Ohio, USA
-------
2. THE REMOTE SENSING EFFORT
2.1. INTRODUCTION
The remote sensing effort consisted of planning and implementing all the
logistics and details necessary for the collection of CAS I data from a fixed-wing aircraft.
The CASI instrument used in this project was a rack-mounted, fully programmable, high
resolution (12 bit resampled to 16 bit) pushbroom imaging spectrometer system with the
capability to collect data in 19 channels between 400-950 nm at intervals ranging from
as small as 2 nm in width to tens of nanometers. This system was fully georeferenced
using data from an aircraft mounted gyroscope, an aircraft mounted Global Positioning
System (GPS), base station GPS data, and Digital Elevation Model (DEM) data (see
Figure 2). As a result, all recorded data for this project was corrected for variations in
aircraft altitude and attitude during flight. Spectral wavelengths and widths chosen for
this project evolved from conversations with Herb Ripley of HDI, Inc. (based upon his
many years of experience with this particular instrument), Dr. Prasad Thenkabail from
the Center for Earth Observation at Yale University (also see Thenkabail et al., 1994,
2000), and David Williams from the U.S. EPA's Environmental Services Division in
Reston, Virginia. The resulting 19 spectral bands chosen were selected in the hope of
best discerning both the urban and agricultural landscapes expected in this particular
watershed (Table 1). Flight line acquisition dates for this mission are shown in Figure 3.
A detailed flight log is also provided in Appendix A.
The remote sensing effort also involved a series of pre-processing corrections
(radiometric, geometric, and atmospheric) required to make the CASI data suitable for
input and analysis within a geographic information system (CIS) as well as the next
step, classification of this data into a land use/land cover product (Section 3).
2.2. RADIOMETRIC CORRECTION
Two files derived from data collected during the over-flights were used as a basis
for all image processing steps in this project. One was an ASCII file which contained
aircraft CASI pitch, roll, and GPS data. The other was a raw image file in band
interleaved-by-pixel (BIP) format. Remote sensing instruments on the aircraft (i.e., the
imager, gyroscope, GPS, computer and hard drive storage, etc.) were linked by a
Pulse-Per-Second (PPS) signal and time stamp based on Greenwich Mean Time
(GMT).
-------
FIGURE 2
Remote Sensing Instrumentation Used in this Project
-------
TABLE 1
Spectral Bands Remotely Sensed in the Little Miami River Project
Band
1
2
3
4
5
6
7
8
9
10
11
Spectral
Region
Blue
Blue
Green
Green
Green
Green-Red
Red
Red
Red
Red
Red-edge
Band
Center
Nm
449.6
490.4
520.2
550.2
574.6
600
619.8
659.6
674.8
691
700.5
Band Width
(+/-)
Nm
15.0
15.0
9.5
9.5
7.7
8.6
7.7
7.7
7.8
4.9
4.9
Band
Range
Nm
30.0
30.0
19.0
19.0
15.4
17.2
15.4
15.4
15.6
9.8
9.8
Comments
Crop to soil reflectance ratio
minima.
1 st order derivative. Positive
change in reflectance per unit
change in wavelength is
maximized. Pigment content.
Green band peak. Related to
total chlorophyll.
1 st order derivative. Negative
change in reflectance per unit
change in wavelength is
maximized. Pigment content.
Chlorophyll absorption pre-
maxima.
Chlorophyll absorption
maxima. Greatest crop-soil
contrast. Related to
chlorophyll a and b.
Chlorophyll absorption post-
maxima.
-------
TABLE 1 cont.
Band
12
13
14
15
16
17
18
19
Spectral
Region
Red-edge
Red-NIR
NIR
NIR
NIR
Shoulder
NIR Peakl
NIR Peak2
NIR-
Moisture
Sensitive
Band
Center
Nm
719.6
750.1
799.9
820.1
845.1
899.9
920.2
937.5
Band Width
(+/-)
Nm
6.8
10.7
10.7
9.7
9.8
10.7
9.8
7.9
Band
Range
Nm
13.6
21.4
21.4
19.4
19.6
21.4
19.6
15.8
Comments
1 st order derivative. Maximum
change in slope of reflectance
spectra per unit change in
wavelength. Vegetative stress.
Nitrogen status of plants.
Red edge/vegetative stress.
Atmospheric water
absorption/correction
Center of near-infrared (NIR)
shoulder. Broad or narrow
band may provide same result.
Related to total chlorophyll.
Crop growth, stress or
senescense. Useful for crop
moisture sensitive index.
Crop growth, stress or
senescense.
Atmospheric water
absorption/correction.
-------
Acquisition Dates
Dats_
7/25.'2Q03
7 1-31, '2003
mama
S 9 2003
FIGURES
Flight Line Acquisition Dates of CASI Data
in the Little Miami River Watershed
7
-------
Radiometric correction was the first and most important step in the image
processing stream. In short, this correction calibrates pixel digital numbers to radiance
units. This is necessary because imagery from any given sensor may be influenced by
factors such as changes in scene illumination, atmospheric conditions, viewing
geometry variations, and instrument response characteristics (Lillesand and Kiefer,
1994). The CASI instrument used in this project normally undergoes at least one and
sometimes more system calibrations per year. The system manufacturer conducts this
calibration in the laboratory using an U.S. National Institute of Standards and
Technology (NIST) radiance standard. Full frame uniformity and absolute data at each
aperture stop are obtained to determine Radiant Sensitivity Coefficients (RSCs) for all
pixels in the CCD array. The RSC files are used to convert scene data into units of
spectral radiance. The calibration procedures compensate for variations in optical
transmission and CCD responsivity. Signal contributions arising from electronic offset,
dark current, frame shift smear and scattered light are also removed. CASI image scan
lines were then correlated and interpolated with the internal navigation data records,
i.e., position (X, Y) and CASI gyro-based attitude (pitch and roll) measurements, to
create a single file suitable for geocorrection.
2.3. GEOMETRIC CORRECTION
Each flightline was geometrically corrected and geographically registered using
the CASI manufacturer's software (ITRES Research Limited, 2006). This project
generated CASI data with a nominal accuracy of 3 pixels RMS or 4 meter RMS
accuracies which fell within the published accuracies of the CASI instrument. PCI
Geomatica geocorrection software was employed subsequent to the ITRES procedures
to further refine the geographic accuracy. Geometric correction also involved the use of
road vectors (Wessex Inc., 1997) to ensure accurate map registration.
2.4. ATMOSPHERIC CORRECTION
Each geocorrected flight line was then atmospherically corrected using ACORN
(Atmospheric CORection Now) software (Analytical Imaging and Geophysics LLC.,
2002). In brief, ACORN assesses, models, and compensates for the atmosphere to
convert input radiance spectra to apparent surface reflectance. After atmospheric
correction, the spectral absorption features inherent to surface materials are revealed.
This software uses the MODTRAN4 algorithm for atmospheric correction of calibrated
hyperspectal and multispectral data in the 350 to 2500 nm spectral range.
8
-------
3. IMAGE CLASSIFICATION
3.1. INTRODUCTION
The end product objective for this project was to develop a 4-meter spatial
resolution classification of the entire Little Miami River watershed, and to do so with an
accuracy of 80% or better for each class. Particular emphasis was placed on discerning
landscapes thought to potentially contribute nutrients to the Little Miami River and its
tributaries (i.e., various chemical species of nitrogen or phosphorus). A variety of
classification methods were considered during the preparation of a quality assurance
project plan for this project. The final classification approach chosen (a hierarchical,
two-step scheme incorporating the strengths of both object-image segmentation and
pixel-based approaches) is detailed in the next section.
3.2. THE CLASSIFICATION PROCESS
Key steps for deriving the Little Miami River Watershed Land Use/Land Cover
Classification presented in this report included:
• Reviewing the quality of the CASI data for purposes of classification (Section
3.2.1);
• Additional ground-truth work through the collection and interpretation of aerial
images from 2002 and 2003 (Section 3.2.2);
• Generating classification results (Section 3.2.3); and
• Ensuring consistency of the classification across flightlines (Section 3.2.4).
3.2.1. Quality Review of the CASI Data for Purposes of Classification. Considering
the climate conditions during the period of collection, the quality of the CASI data was
deemed sufficient for image classification purposes. However, two artifacts present in
the data did require some additional work. The first type of artifact was the presence of
clouds and corresponding cloud shadow in some of the imagery. Overall, this type of
artifact was minor except for a 342 acre gap between flight lines 13a and 13b which
remained unclassified (Figure 4).
The second type of artifact was cross-track illumination effects. Cross-track
illumination artifacts routinely occur in airborne hyperspectral data and are a result of
sun-sensor-target-geometry, atmospheric conditions, differential path length (across the
instantaneous field of view or IFOV) and spectral band selection. The CASI dataset of
the Little Miami River watershed is variably affected by cross-track illumination artifacts
-------
•1 U-v B*wd ttouk 1173.1 I 2
KBSflOO "2M 16,800 22.4M
FIGURE 4
An Unclassified Gap Between Flight Lines 13a and 13b, July 25, 2002
10
-------
with some flight lines more affected than others. The procedure used to evaluate the
data was to examine each flight line both visually and statistically. Figures 5 and 6
provide a graphical representation of cross-track illumination. These figures show a
section of uncorrected CASI data for two different flight lines acquired on July 24 and
July 25 respectively. A distinct brightening on the right side of the images is apparent
and visible in the cross sections on the right side of the graphics. These cross sections
were generated by averaging the Digital Numbers (DNs) under a graphic mask in each
band in the cross-track direction of the flightline. A polynomial curve has been fitted to
the averaged profile to demonstrate the increased DN values at the edges of the flight
line relative to the center. There is variability in the cover type under the mask as can
be seen in the spectral profiles, however, the polynomial curve represents an averaged
measure and is generally representative of overall trends. The elevated DNs on the left
of the flight line are not overly apparent in the image on the left.
The results of image classification using data with cross-track illumination effects
can be variable and is dependent on the spectral properties of the individual classes
relative to the scene. In general, the ability to extract any particular class from a dataset
is based on the presence of a unique statistical signature that represents a subset of the
full scene variance. In the case of multiple classes, the individual classes must also be
unique relative to each other. Class confusion occurs when there is statistical overlap
between two distinct classes. In images with cross-track illumination effects, the spectral
signature of ground cover types is different at the edges of the image than at the center.
Consequently, training areas selected at the edges of the image will have a different
spectral signature, or statistical representation, than at the center of the image.
Two approaches were considered to mitigate the effects of these cross-track
illumination artifacts: (1) correction of the individual flight lines to remove the cross-track
illumination effects, or (2) adapt to the cross-track illumination effects during the
classification phase of the project. It was determined that the first approach was
beyond the scope and available budget of the project since it would require substantial
reprocessing of the data including performing corrections on all 82 flightlines and a
reapplication of geometric and atmospheric corrections to each flight line. As such the
second approach of adapting to the cross-track illumination effects in the classification
procedure presented a more economical approach for this project and may in fact be
superior to the first approach given the paucity and nature of correction algorithm tools
available in most image processing software. Moreover, the second approach can
effectively deal with misclassification due to cloud and shadows. The chosen approach
11
-------
FIGURES
Cross-Track Illumination Artifacts from Flight Line 38 Acquired on July 24, 2002
FIGURES
Cross-Track Illumination Artifacts from Flight Line 10 Acquired on July 25, 2002
12
-------
basically amounts to increasing the number and strategic distribution of training areas in
the affected flight lines such that an accurate representation of each class is acquired.
3.2.2. Ground Truth. In addition to 390 ground truth points collected by U.S. EPA
personnel during the image acquisition in 2002, supplementary ground training samples
were collected using two different sources of high resolution aerial images. The first
was the 2003 metropolitan Cincinnati digital orthophoto quads (Aerials Express Inc.,
2003) used to sample urban classes. The second dataset consisted of un-rectified
2002 aerial images from the Center for Mapping at the Ohio State University used to
increase the number of sample locations in agricultural fields. Overall, 3613 and 351
reference polygons were obtained from each data set, respectively. Reference
polygons from aerial images were selected from homogeneous areas of 40m x 40m (or
0.4 acre) minimum. Some of this data was used for training the two supervised
classifications: image object segmentation and spectral angle mapper (SAM). Other
portions of it were used for accuracy assessment following classification (Section 4.3).
Obtaining adequate ground truth data was a key component of this project and a
critical aspect to acquiring ground truth data from secondary sources was in setting
business rules or guidelines which precisely specified the process to be followed. The
key guidelines followed included:
1. Location of sample site
It is not uncommon to collect information on the wrong location because of
inadequate procedures. This project used orthophotos of high spatial accuracy
(<2m) as base layers for the ground truthing process.
2. Data collection and entry error
Data collection errors occur when measurements are done incorrectly and
variables of the classification scheme are misidentified (i.e., crop type). Ground
truth data given by EPA, particularly on specific variables such as crop type, was
the reference for the ground truthing process. Errors were monitored and
removed from the data set when found.
3. Data collection consistency
Training and the development of objective data collection procedures ensure
data collection consistency. It is important to ensure that everyone identifying
ground truth points follow the same process. Appendix B contains the protocol
used by EPA to collect ground truth data. Appendix C contains the supplemental
ground truth data collection protocol for the project.
13
-------
4. Date of reference data
If change in land use/land class occurs between the date of imagery capture and
the date of ground truth data, the results may be impacted. Therefore, ground
truth data were collected as close as possible to the date of the CASI image
acquisition. For example, EPA's ground truth data from the field coincided with
the CASI flyover. Supplemental aerial images used as secondary sources of
agricultural ground truth came also from 2002. The 2002 LMR CASI dataset and
the Aerials Express Inc. (2003) dataset (used as a secondary source of urban
ground truth) are only a year or less apart, and seasonal difference within the
CASI dataset itself was also minimal (ranging only from late July through early
August 2002).
5. Sampling design
The choice and distribution of ground truth samples is an important aspect of
accuracy assessment. The concept of randomness is a central issue to ensure
that the samples are chosen without bias and eventually the accuracy of the map
is statistically sound. For this project, the 82 flight lines were used as an
independent spatial framework to guarantee that the required minimum number
of ground truth data is collected for each class, and that they are statistically
random samples.
6. Number of ground truth data
In the remote sensing community, a general guideline or good "rule of thumb" is
to collect a minimum of 50 samples for each class. The objective was to collect
samples for each class for accuracy assessment and training set samples for
each class on every flight line. Overall, 4354 ground truth samples were
obtained; 902 of these were used for the accuracy assessment.
3.2.3. Classification Method. A number of classification algorithms including
Maximum Likelihood Classification (MLC), Spectral Angle Mapper (SAM), and
Classification And Regression Tree (CART) were tested on the CASI imagery collected
from the Little Miami River watershed. However, no single algorithm alone proved to be
capable of successfully classifying the hyperspectral data. Some of the problems
encountered included the following:
• Sub-classes of forest (e.g., deciduous versus conifer) were not separable,
especially because the areas are difficult to find ground-truth for.
• Other vegetation, such as corn, grass, soybean, etc. gets classified into forest,
and vice versa.
• Pasture, soybeans, corn and grass get confounded, particularly in urban areas.
• Roads are not discernible from urban/built and barren land covers.
• Rock and stone quarries were not discernible either.
14
-------
Nevertheless, a solution was found to resolve the above problems and to obtain
as many classes as possible under the study constraint of 80% accuracy per individual
class. The solution was to use a hierarchical approach using two different classification
algorithms: "Image Object Segmentation" and "Spectral Angle Mapper."
Image object segmentation is an innovative method which uses
homogenous image objects as processing units at a given scale for classification,
instead of pixels. One motivation for the image object approach is to make use
of powerful generalizations of the image to generate level-1 strata of different
contexts. With the creation of these strata (urban, rural and water), further
independent classification (e.g., pixel-based or other algorithms) of each stratum
can be applied and then the results combined later to improve the accuracy of
the final classified dataset.
Figure 7 shows an example result of image object segmentation versus pixel-
based classification of the same forest area in the Little Miami watershed. Forest areas
typically contain areas of dark shadows, bright canopy tops, and medium bright
illumination under the tree canopy, which pixel-based approaches without level-1 strata
classify into all different classes, for example: forest (dark green), corn (gold), and
soybean (light green). Unreasonable classes, such as corn and soybean crops in the
midst of forest land will be removed if a hierarchical approach is used. Also, image
object segmentation could also be used without any negative effect on delineating
riparian features such as a thin forest strip along a river.
On the other hand, the Spectral Angle Mapper (SAM) is a whole pixel, spectral
method for supervised classification developed specifically for hyperspectral data. It is
based on entirely different principles than common multispectral classification
algorithms such as the probability-based Maximum Likelihood Classifier. SAM is
thought of as a similarity classifier and can be used on multispectral data too, often with
improved results in comparing the spectral properties of materials. In brief, it is a
physically-based spectral classification that uses an n-dimensional angle to match the
spectra of imaged (remotely sensed) pixels to reference spectra. The algorithm
determines the spectral similarity between image and reference spectra by calculating
the spectral angle between them, treating them as unit vectors in spectral space with
dimensionality equal to the number of bands (in this particular case, 19 bands or
dimensions). Since only the angle between the vectors matters (the smaller the angle,
the better the match) and not the vector's length, brightness variations such as
topographic and cross-track illumination effects are well accounted for. In other words
SAM's advantage over common or more "traditional" classifiers is its relative
15
-------
FIGURE/
Image Object Segmentation versus Pixel-Based Classification
of a Forest Area in the Little Miami River Watershed
16
-------
insensitivity to illumination and albedo effects inherent with remotely sensed imagery
(Research Systems Inc., 2002; Kruse et al., 1993).
In order to combine the power of both approaches, object image segmentation in
eCognition (Definiens Imaging GmbH., 2003) was applied as a "Level 1" classification of
water, urban, and rural features which required consideration of large-scale factors, as
well as area-based parameters such as adjacency, texture and shape. Next, Spectral
Angle Mapper in ENVI (Research Systems Inc., 2003) was applied as a "Level-2"
classifier to discern the urban and rural areas into more specific classes (barren, built,
grass, corn, soypbeans, wheat, etc.). Wire diagrams of the methodology used are
shown in Figures 8 and 9. Throughout the process, image object segmentation created
a spatially exclusive mask of urban and rural regions and then each region was filled
with the classification result from SAM. As a consequence, the final classification result
remained a uni-scale product with a spatial resolution of 4 meters throughout the entire
watershed.
Training samples were chosen along all flight lines and selected in order to
capture intra-class variation. The same classification rules were applied to each
flightline. After the first round of classification, results were fine-tuned by adding training
sets to accurately define inter-class boundaries. The classification results for each
individual flight line were assessed for accuracy and accepted if they did not show any
overall discrepancies with respect to the aerial images. Classification results were also
re-examined after joining the classified flight lines together in a mosaic of the
watershed.
Any problem areas noted during this stage were addressed during the QA/QC
phase (Section 3.2.4 and Section 4).
Overall, 11 land cover classes were mapped. A twelfth class included pixels that
could not be classified, for example, due to cloud cover. The 11 land cover classes
consisted of:
Lentic: Open water associated with still water systems, such as lakes,
reservoirs, potholes, and stockponds. Such bodies typically do not have a
defined channel or associated floodplain.
Lotic: Open water associated with running water systems, such as rivers or
streams. Such waterways typically have a defined channel and an associated
floodplain.
Forest: Contains either or both deciduous and coniferous trees in any degree of
mixture. Single stemmed, woody vegetation with canopy spanning greater than
4 meters and tree canopy accounting for 25-100% of the cover
Corn: Area under cultivation of food and fiber, where corn is the primary crop.
17
-------
Flight line-
1
Flight line
-82
Ground
Truth
- Special care was placed on
riparian features here
- Water was categorized into
lotic and lentic
- Removal of cross track
lumination effect
FIGURES
Flowchart of the Classification Method
18
-------
I Level 2 Classification
i ENVI Spectral Angle
Mapper
Special care was (
placed on riparian i
i features '
L
FIGURE 9
Schematic of the Hierarchical Classification and Derived Classes (shadowed)
19
-------
Soybean: Area under cultivation of food and fiber, where soybean is the primary
crop.
Wheat: Area under cultivation of food and fiber, where wheat is the primary crop.
Dry Herbaceous: Dominated by dry and/or less vigorous herbaceous vegetation;
herbaceous vegetation accounts for more than 25% of the ground cover. This
class mainly includes naturally occurring and unmanaged herbaceous vegetation,
and dried out, unhealthy, or stressed croplands. Dry herbaceous vegetation
prevailed in croplands, as well as, "Other Agriculture" lands (fallow, hay, pasture,
or natural grassland prairies or fields), due to drought in the Summer of 2002.
Dry herbaceous vegetation had little chlorophyll content and very similar spectral
signatures without regard to vegetative species.
Grass: Dominated by cultivated grasses planted in developed settings for
recreation, erosion control, or aesthetic purposes. Examples include parks,
lawns, golf courses, airport grasses, and industrial site grasses.
Urban Barren: Composed of bare soil, rock, sand, silt, gravel, or other earthen
material with little (less than 25%) or no vegetation within urban areas.
Examples include exposed soil in urban areas and construction sites.
Rural Barren: Composed of bare soil, rock, sand, silt, gravel, or other earthen
material with little (less than 25%) or no vegetation in rural areas. Typically
fallow fields are included in this class too.
Urban/Built: Areas covered by structures and impervious surfaces in urban,
suburban, and rural areas. Typically buildings, parking lots, and paved roads.
Unclassified: This class includes areas of image gaps among flight-lines and
cloud cover where land cover classification was not feasible.
Figure 10 shows the class code, name and corresponding color schemes for all
classification results. Figure 11 provides an overview of the LULC classification.
QassCode I CbssNaroe Color
FIGURE 10
Legend for the Final 4m x 4m LULC Product
20
-------
FIGURE 11
Overview of the 4m x 4m LULC Classification for the Little Miami River Watershed
21
-------
Final products were delivered in the projection and data format provided in
Table 2.
TABLE 2
Projection for the Little Miami River Land Use/Land Cover
Projection:
1st Standard Latitude:
2nd Standard Latitude:
Latitude of Projection's Origin:
Central (Meridian) Longitude
Origin:
Datum (Ellipsoid)/Spheroid
Units:
Orientation:
Pixel Size:
Precision for mosaicked flight
lines:
Data Format:
Naming Convention
Albers Conic Equal-Area
29 degrees, 30 minutes, 00 seconds
45 degrees, 30 minutes, 00 seconds
23 degrees, 00 minutes, 00 seconds
-96 degrees, 00 minutes, 00 seconds
NAD83/GRS80
Meters
North up
4 meters
+ 3 pixels
Erdas Imagine 32 bit signed integer .img file format
Oriainal/unsmoothed LULC product:
little_miami_river_watershed_4m_before_clumping_
signed32bit.img
Smoothed LULC product:
Iittle_miami_river_watershed_4m_signed32bit.img
22
-------
3.2.4. Consistency of Classification Across Flightlines. Since there were 82
independent flight lines of imagery, it was important to manage consistency of
classification across the flight lines. In order to do this, the classification team classified
and conducted an extensive accuracy assessment on 8 selected flight lines (10% of
total number of images). As the selected flight lines were classified to an acceptable
level of accuracy, they were used to aid in ensuring consistency across all the
neighboring images. When an initial classification was completed for any given flight
line, it was compared to all of its neighbors and any distinct classification differences
along the boundaries or overlap between flight lines were addressed. This process
helped to mitigate categorical edge-matching errors when the 82 individual classified
flight lines were finally stitched together into the watershed mosaic. Discontinuities
along image boundaries were removed using polynomial based rubber-sheeting using
ground control points and corresponding cut-lines. The final mosaicked product has
three pixel RMS errors on 4 meter spatial resolution data.
23
-------
4. POST-PROCESSING AND QA/QC
Image post-processing and QA/QC following the Level 1 and 2 classifications
included manual editing (Section 4.1) and map generalization to create a second LULC
product (Section 4.2). Follow-up steps used to ensure the accuracy and proper
interpretation of the end product included:
• The completion of an accuracy assessment including an error matrix and
computations of overall, producer, and user accuracy (Section 4.3); and
• Generating metadata for the final product compatible with the Federal
Geographic Data Committee (FGDC) standard (Appendix D).
4.1. MANUAL EDITING
Manual editing was used as a final QA/QC step. Final image edits included:
1) Differentiation of water bodies into lotic and lentic. The discrimination of
lotic versus lentic waters was completed using the U.S. EPA's National
Hydrography Dataset or "NHD."
2) Misclassifications due to clouds, shadows, or haze. Hyperspectral images
were systematically examined along with the classification results to
identify any misclassified pixels due to the presence of clouds, shadows,
or haze. These pixels were then manually included in a mask to convert
them to the "Unclassified" class.
3) Removal of boundary effects. Cross-track illumination affected the SAM
classification results along the boundary of each flight line. As such,
manual editing was necessary to correctly assign the classes. The other
task was to remove any inconsistent classifications in overlapping flight
lines due to the temporal gap in image acquisition. Atmospheric
conditions varied day-by-day, and a 16-day gap in image acquisition
throughout the watershed is long enough to expect some changes in
vegetation condition. The mosaicking phase was prolonged due to cross-
track illumination effects and the temporal gap in image acquisition. All
the image boundaries had to be manually inspected and the correct class
in the overlap area determined. The percentage of "dry herbaceous" may
have increased in the southern region due to the later dates of imagery
acquisition. Wheat class was scarce in the southern region as the fields
had been harvested, or had senesced. Small portions of some soybean
and corn fields had also senesced and, in such cases, were not assigned
to the correct class. Such fields were manually interpreted and assigned
to the correct class.
4) Other discernible misclassifications. Dark shadow areas and black
asphalt pavement were easily confused with water bodies. In such cases,
these areas were manually interpreted and assigned to the correct class.
24
-------
4.2. MAP GENERALIZATION
A "clump-sieve-and-fill" technique was used to eliminate single pixels or
groupings of pixels that were smaller than the minimum target mapping unit (e.g.,
random pixels of "Forest" denoting scattered trees in an otherwise homogeneous 40-
acre plot of "Corn"). As a result, a second LULC mosaic product was produced which
eliminated the "salt and pepper" effect common in classifications of smaller pixel or
"finer" spatial resolution imagery. Although smoothed images are generally more
"pleasing" to the eye, U.S. EPA was concerned about any smoothing process which
might potentially wipe out important small features, or thin, linear features such as
riparian forest. On the other hand, available literature also suggested that whereas
unsmoothed or "salt and pepper" classification results may be more realistic (e.g., cases
where random pixels of "Forest" or scattered trees really do exist within a "Corn" field),
smoothed classifications may be more meaningful in terms of deriving land use
statistics important for interpreting ecological processes across the landscape (e.g.,
Burnett and Blaschke, 2003; Dorren et al., 2003). As a result, two LULC products were
created for the Little Miami River watershed, one smoothed and one not. For the
smoothed product, the minimum mapping unit of the classification result is about 0.04 of
an acre as represented by 10 pixel clusters (at 4m x 4m spatial resolution), or linear
chains of minimally four contiguous pixels in any direction. The original, unsmoothed
product remains at a 4m x 4m spatial resolution.
4.3. ACCURACY ASSESSMENT
The accuracy assessment was based on whether the majority of classed pixels
in a 3 x 3 pixel window, centered on a ground truth site, agreed or not. Thus, if five or
more pixels were classified as corn, and ground truth indicated corn, then the majority
criterion was satisfied and "corn class" would be considered correct for that site. A
standard error matrix was used in reporting classification accuracies (Table 3). This
matrix reports the number of pixels assigned to a particular category in a classification
relative to the number of pixels assigned to a particular category in a reference
classification. In this case, the classified data, represented by rows in Table 3 and
Table 4, are the land cover classifications derived for this project, and the reference
data are represented by the columns in the tables. A total of 902 independent ground
truth sites were used for the accuracy assessment, including primary data (i.e., data
collected by U.S. EPA scientists in the field at the time of the overflights), and
secondary data from 2002 and 2003 aerial images of the watershed as explained in
Section 3.2.2.
25
-------
TABLE 3
Classification Error Matrix
(units are the number of reference points)
Classified
Lentic
Lotic
Forest
Corn
Soybean
Wheat
Dry Herbaceous
Grass
Urban Barren
Rural Barren
Urban/Built
Column Total
Reference (Known Cover Types)
Lentic
85
3
0
0
0
0
2
1
0
6
1
98
Lotic
1
17
2
0
0
0
0
0
0
1
3
24
Forest
0
0
95
1
0
0
3
1
0
0
0
100
Corn
0
0
0
106
3
0
9
0
0
2
0
120
Soybean
0
0
0
3
107
0
12
0
0
0
0
122
Wheat
0
0
0
0
2
18
8
0
0
10
0
38
Dry
Herbaceous
0
0
3
1
5
0
82
8
0
1
0
100
Grass
0
0
4
0
1
0
16
75
1
3
0
100
Urban
Barren
0
0
0
0
0
0
0
8
57
0
4
69
Rural
Barren
0
0
2
0
0
0
8
0
1
20
0
31
Urban/
Built
0
0
0
0
0
0
0
0
4
1
95
100
Row
Total
86
20
106
111
118
18
140
93
63
44
103
902
26
-------
TABLE 4
Summary Statistics of the Accuracy Assessment
(Overall Accuracy of Classification = 83.92%)
(KHAT statistic = 0.82)
Class Name
Lentic
Lotic
Forest
Corn
Soy
Wheat
Dry Herb
Grass
Urban Barren
Rural Barren
Urban/Built
Reference
Totals
98
24
100
120
122
38
100
100
69
31
100
Classified
Totals
86
20
106
111
118
18
140
93
63
44
103
Number
Correct
85
17
95
106
107
18
82
75
57
20
95
Producers
Accuracy
86.73%
70.83%
95.00%
88.33%
87.70%
47.37%
82.00%
75.00%
82.61%
64.52%
95.00%
Users
Accuracy
98.84%
85.00%
89.62%
95.50%
90.68%
100.00%
58.57%
80.65%
90.48%
45.45%
92.23%
27
-------
Three major descriptive measures were used for accuracy assessment, namely:
"overall accuracy," "producer's accuracy," and "user's accuracy." Overall accuracy is
computed by dividing the total number of correctly classified pixels (the sum of the
elements along the major diagonal in the error matrix) by the total number of reference
points. Producer's accuracy indicates how well reference pixels of the given cover type
are classified. It is computed by dividing the number of correctly classified pixels in
each category (on the major diagonal of the error matrix) by the number of reference
points used for the class (the column total). User's accuracy is a measure of
commission error and indicates the probability that a pixel classified into a given
category actually represents that category on the ground. It is computed by dividing the
number of correctly classified pixels in each category (on the major diagonal of the error
matrix) by the total number of pixels classified into that category (row total) (Lillesand
and Kiefer, 1994).
Tables 3 and 4 present the quantitative summary of the accuracy assessment for
the classification result using 902 reference points. The overall classification accuracy
was 83.92%, which is above the project's target of 80% accuracy. However, the
producer's and user's accuracy of some classes fell short of the target of 80%. The
KHAT statistic, a comparative statistic to use in comparing this classification product
with others, was 0.82. In other words, the classification result here was 82% better than
one resulting from chance. Overall, the strengths of this classification relative to other
existing LULC datasets of this watershed, such as the National Land Cover Dataset or
NLCD (Vogelmann et al., 2001), and the State of Ohio Land Cover (Ohio DNR, 1994)
are:
• Higher spatial resolution (e.g., 4m x 4m rather than 30m x 30m pixel resolution of
previous existing classifications of this watershed),
• Higher spectral resolution (the classification is derived from more spectral bands
of information than provided by Landsat platforms),
• Ground-truth data used as a basis for this classification includes primary sources
of information taken in the field at the same time the remote sensing data was
collected,
• The discernment of corn from soybeans (the two major agricultural crops found
within this watershed), and
• Better spatial and updated temporal discernment of urban built (existing and new
infrastructure) versus urban barren (newly developing areas).
28
-------
5. RESULTS AND DISCUSSION
A basic understanding of the physical (i.e., geological and anthropogenic)
processes at work in the Little Miami River watershed will help the user of this LULC
dataset interpret some of the resulting land use patterns shown in Figure 11. For
example, one should note that the northern half of the watershed is relatively flat and
composed of "Till Plains" with soils that developed from the loamy, limy deposits of the
Wisconsinan glaciation roughly 18,000 years ago. These soils normally have better
natural drainage and fertility than those of the southern half of the watershed (or "Drift
Plain"). The southern half of the watershed has more deeply-leached, acidic, pre-
Wisconsinan till and thin loess, as well as, very poorly-drained soils with fragipans
(clays). The southern half of the watershed also exhibits relatively modest relief, but
with dissected areas and somewhat more complex topography than the north (Omernik,
1987; Woods et al., 1998). As such, the northern and southern parts of the watershed
may be expected to have different types and proportions of certain land uses or land
covers based on the differing soils and micro-climates found in these two distinct
"ecoregions" (the "Till Plain" and the "Drift Plain," refer to Figure 1).
Spatial patterns separating western and eastern portions of the watershed exist
too. Perhaps most notable is the western urban/exurban corridor stretching from
Cincinnati (in the south) to Dayton and Xenia (in the north) and beyond, encompassing
portions of Hamilton, Warren, and Montgomery Counties. These growing urban
landscapes run parallel to and already straddle much of the mainstem of the Little Miami
River which can be observed as a nearly contiguous linear band of riparian forest
running up along the western part of the image. The eastern half of the watershed
tends to be more agricultural in character, particularly in the north. But this appears to
wane in the east-central part of the image near the city of Wilmington (a crossroads or
pole for the primary economic sector in this region, as well as, a major air transportation
hub), and in the south as well, particularly along the East Fork of the Little Miami River
in Clermont County where urban development and human population continue to rapidly
grow.
Other major patterns in the final classification include concentrations of "dry
herbaceous" land cover in the western part of the watershed near the urban/rural fringes
of Warren, Montgomery and Green Counties, as well as along, or at the source of,
many headwater streams to the east (i.e., dry, thin vegetative areas buffering perennial
or low-flow streams from adjacent croplands). Recall that "dry herbaceous" was defined
29
-------
in this project as a variety of "Other Agricultural" components including hay, pasture,
fallow, dried out crops, and natural herbaceous vegetation. As such, concentrations of
dry herbaceous land cover in the west are likely non-irrigated lands, areas practicing
water conservation, or perhaps areas left fallow due to failed crops or in anticipation of
near-term development. The thin lines of dry herbaceous cover bordering perennial or
low-flow streams in the east was likely spectrally distinct, or separable from corn and
soybeans based on aerial photography, yet still dry from drought and "managed" only in
the sense that it was not mowed or turned over by farmers.
The ratio of corn to soybeans is lower in the southern part of the watershed. This
is expected to a certain extent since soybeans are well known to be common and well
adapted to spring soil wetness in the southern half of the Little Miami River watershed
(Woods et al., 1998). This observation may well represent the reality of crop planting
patterns in 2002, and/or failed corn crops susceptible to the early flood then drought
extremes experienced that year (and thus, classifications of "dry herbaceous" rather
than "corn"). The planting of corn was delayed at least three weeks in many areas of
Ohio in 2002 due to heavy spring rains and flooded fields that year. However,
subsequent to seeding, drought conditions ensued and lingered in 2002.
The summer drought of 2002 likely affected other classes in this LULC as well.
For example, lotic or running waters were rarely detected because many drainage ways,
or headwater or perennial streams during the period of late July to early August were
already dry or experiencing low flows. Un-watered grasses in otherwise managed
grassy areas might also have been classified as "dry herbaceous."
Nevertheless, the resulting LULC is an important dataset for a variety of
environmental and geographic studies within the Little Miami River watershed. Even
given the predominance of the "dry herbaceous" class, it remains meaningful in terms of
studying several urban and agricultural patterns or gradients, as well as, anthropogenic
and natural processes within the watershed. Many hydrological, ecological, and
geological applications may be possible too, along with the ability to assist in urban
planning, study urban sprawl, and contribute to measurements and evaluations of
zoning, congestion, pollution and human health.
30
-------
6. REFERENCES
Aerials Express Inc. 2003. Cincinnati Metropolitan Aerial Photography. Tempe, AZ.
Analytical Imaging and Geophysics LLC. 2002. ACORN Quick Start Guide. ACORN
Version 4.10, March, 2002 Edition. Boulder, CO.
Burnett, C. and T. Blaschke. 2003. A multi-scale segmentation/object relationship
modelling methodology for landscape analysis. Ecol. Model. 168:233-249.
Definiens Imaging GmbH. 2003. eCognition 3.0. Munich, Germany.
Dorren, L.K.A. , B. Maier and A.C. Seijmonsbergen. 2003. Improved Landsat-based
forest mapping in steep mountainous terrain using object-based classification. Forest
Ecol. Manag. 183:31-46
ITRES Research Limited. 2006. ITRES: Airborne Hyperspectral Remote Sensing
Systems & Solutions. Calgary, Alberta, Canada. Available at http://www.itres.com.
Kruse, F.A., A.B. Lefkoff, J.B. Boardman, K.B. Heidebrecht, AT. Shapiro, P.J. Barloon,
and A.F.H. Goetz. 1993. The Spectral Image Processing System (SIPS) - Interactive
Visualization and Analysis of Imaging spectrometer Data. Remote Sens. Environ.
44:145-163.
Lillesand, T.M. and R.W. Kiefer. 1994. Remote Sensing and Image Interpretation, 3rd
ed. John Wiley and Sons, Inc., New York, NY.
Ohio DNR (Ohio Department of Natural Resources). 1994. State of Ohio Land Cover.
Division of Real Estate and Land Management. Columbus, OH.
Omernik, J.M. 1987. Ecoregions of the conterminous United States. Map
(1:7,500,000). Ann. Assoc. Am. Geograph. 77(1):118-125.
Research Systems Inc. 2002. Spectral Analysis with ENVI. Boulder, CO: The
Professional Services Group, Research Systems, Inc., A Kodak Company.
Research Systems Inc. 2003. ENVI 3.6 (the Environment for Visualizing Images).
Boulder, CO. Available at http://www.ResearchSystems.com.
Sanders, R.E., Ed. 2002. A guide to Ohio streams. Columbus, Ohio: Streams
Committee, Ohio Chapter of the American Fisheries Society. In partnership with the
Ohio Environmental Education Fund, Ohio Department of Natural Resources, and the
Ohio Environmental Protection Agency.
31
-------
Thenkabail, P.S., A.D. Ward, J.G. Lyon and C.J. Merry. 1994. Thematic Mapper
vegetation indices for determining soybean and corn crop parameters. Photogramm.
Eng. Rem. S. 60:437-442.
Thenkabail, P.S., R.B. Smith and E. De Pauw. 2000. Hyperspectral vegetation indices
for determining agricultural crop characteristics. Remote Sens. Environ. 71:158-182.
Vogelmann, J.E., S.M. Howard, L. Yang, C.R. Larson, B.K. Wylie and N. Van Driel.
2001. Completion of the 1990s national land cover data set for the conterminous United
States from Landsat thematic mapper data and ancillary data sources. Photogramm.
Eng. Rem. S. 67:650-662.
Wessex, Inc. 1997. "First St." software. U.S. Streets CDROM, Release 4.0. Winetka,
IL. Available at www.wessex.com.
Woods, A.J., J.M. Omernik, C.S. Brockman, T.D. Gerber, W.D. Hosteterand S.H.
Azevedo. 1998. Ecoregions of Indiana and Ohio. Interior-Geological Survey, Reston,
VA. Accessed May 19, 2005 at ftp://ftp.epa.gov/wed/ecoregions/oh in/ohin front.pdf.
32
-------
APPENDIX A
FLIGHT LOGS OF THE CASI DATA COLLECTION
IN THE LITTLE MIAMI RIVER WATERSHED
June 24-August 8, 2002
A-1
-------
DATE
July 24, 2002
July 25, 2002
LINE
1
2
3
4
5
6
7
25
26
27
36
37
38
39
8
FILE
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
1
2
START
TIME
12:55:23
13:10:49
13:18:57
13:27:51
13:36:01
13:46:01
13:57:05
14:07:00
14:13:49
14:24:00
14:38:15
14:52:45
15:15:20
15:32:45
15:50:10
16:01:47
12:31:28
12:48:10
STOP
TIME
12:56:30
13:15:50
13:24:26
13:33:01
13:42:40
13:53:45
14:04:00
14:13:20
14:19:30
14:34:54
14:49:20
15:04:12
15:29:40
15:47:11
16:00:10
16:13:10
12:32:40
12:58:02
/STOP
5.6
5.6
5.6
5.6
5.6
4
4
4
4
4
4
4
5.6
5.6
5.6
5.6
4
4
INT
TIME
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
ALTITUDE
(FEET)
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
HEADING
VISIBILITY
(MILES)
SPEED
(MPH)
Dark Data
E
W
E
W
E
W
E
10
10
10
10
10
10
10
124
120
127
121
128
121
128
Dark Data
W
E
W
E
W
E
W
10
10
10
8
8
8
8
127
123
127
124
127
127
125
Dark Data
W
5
118
A-2
-------
DATE
July 31, 2002
LINE
9
10
11
12
13
13
14
15
16
17
18
19
20
21
22
23a
FILE
3
4
5
6
7
9
10
11
12
13
14
15
16
17
18
1
3
4
START
TIME
13:00:45
13:13:15
13:27:20
13:39:55
13:54:20
14:00:10
14:08:40
14:24:10
14:39:05
14:56:42
15:12:03
15:28:30
15:43:00
15:58:20
16:07:12
12:38:19
12:39:50
12:51:23
13:05:30
STOP
TIME
13:09:10
13:24:45
13:37:10
13:51:25
14:00:00
14:05:00
14:20:30
14:35:45
14:52:35
15:09:00
15:25:40
15:40:00
15:55:50
16:06:34
16:08:15
12:39:20
12:40:50
13:02:00
13:19:00
/STOP
4
4
4
4
4
4
4
4
4
4
4
5.6
5.6
5.6
5.6
4.0
5.6
4.0
4.0
INT
TIME
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
31
31
ALTITUDE
(FEET)
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
HEADING
E
W
E
W
E
E
W
E
W
E
W
E
W
E
VISIBILITY
(MILES)
5
5
5
5
5
5
5
5
5
5
5
5
5
5
SPEED
(MPH)
130
119
130
118
132
116
132
116
132
116
132
116
132
116
Dark Data
W
E
8-10
8-10
133
105
A-3
-------
DATE
August 9, 2002
LINE
28
29
30
31
32
33
34
35
24
27
26
25
23b
21
20
57
56
55
54
53
FILE
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
2
3
4
5
6
START
TIME
13:34:30
13:47:01
14:03:10
14:15:10
14:30:14
14:41:45
14:56:25
15:10:29
15:31:15
15:46:48
15:51:56
15:54:45
16:02:17
16:05:10
16:09:35
12:58:56
13:17:48
13:38:26
13:56:41
14:18:34
STOP
TIME
13:44:30
14:00:05
14:12:55
14:28:15
14:39:45
14:54:30
15:05:30
15:25:38
15:41:30
15:50:38
15:53:32
16:00:25
16:03:33
16:08:20
16:13:00
13:15:57
13:36:24
13:54:55
14:14:40
14:34:59
/STOP
4.0
4.0
4.0
4.0
4.0
4.0
4.0
5.6
5.6
5.6
5.6
5.6
5.6
5.6
4
4
4
4
4
INT
TIME
31
31
31
31
31
31
31
32
32
32
32
32
32
32
31
31
31
31
31
ALTITUDE
(FEET)
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
HEADING
W
E
W
E
W
E
W
E
W
E
W
E
E
W
E
W
E
W
E
W
VISIBILITY
(MILES)
8-10
8-10
>10
10
10
10
10
SPEED
(MPH)
130
105
130
105
133
105
142
105
140
105
140
105
105
140
105
130
112
127
115
130
A-4
-------
DATE
August 1 , 2002
LINE
52
51
50b
39b
38b
37b
36b
40
41
42
43a
43b
44
45
46
47
48
FILE
7
8
9
10
11
12
13
1
2
3
4
5
6
7
8
9
10
11
12
START
TIME
14:36:32
14:56:00
15:13:35
15:25:14
15:34:57
15:41:07
15:47:25
12:46:47
12:48:37
13:10:30
13:20:05
13:51:29
14:09:20
14:17:46
14:39:20
14:56:25
15:16:45
15:33:25
15:54:07
STOP
TIME
14:54:00
15:12:01
15:17:45
15:33:25
15:39:51
15:45:48
15:50:38
12:48:00
12:49:55
13:26:40
13:48:55
14:06:10
14:14:50
14:32:35
14:54:10
15:14:50
15:30:55
15:51:20
16:08:30
/STOP
4
5.6
5.6
5.6
5.6
5.6
5.6
5.6
40
40
40
40
40
40
40
5.6
5.6
5.6
5.6
INT
TIME
31
31
31
31
31
31
31
32
32
32
32
32
32
32
32
32
32
32
32
ALTITUDE
(FEET)
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
HEADING
E
W
E
E
W
E
W
W
E
W
E
E
W
E
W
E
W
VISIBILITY
(MILES)
10
10
10
10
10
10
10
8-10
8-10
8-10
8-10
8-10
8-10
8-10
8
8
8
SPEED
(MPH)
120
135
115
115
135
116
135
136
114
138
110
110
137
108
138
110
137
A-5
-------
DATE
August 3, 2002
August 4, 2002
Augusts, 2002
August 7, 2002
LINE
49
7
13
23
1
2
3
4
5
50a
68a
74
75
76
77
78
79
80
FILE
13
14
1
2
3
1
2
3
14
19
1
1
2
3
4
5
6
7
8
START
TIME
16:10:35
16:36:10
13:19:12
13:27:26
13:34:56
12:59:12
13:01:56
13:07:29
13:09:21
13:11:02
13:06:02
13:16:21
13:27:02
13:35:43
13:45:11
13:53:28
14:02:20
14:10:46
14:17:53
STOP
TIME
16:29:20
16:37:10
13:23:28
13:29:35
13:37:23
13:00:28
13:04:55
13:09:18
13:10:29
13:12:31
13:24:15
13:23:34
13:33:38
13:42:24
13:51:42
13:59:52
14:08:09
14:16:18
14:22:07
/STOP
5.6
5.6
4.0
4.0
4.0
4
4
4
4
4
4
4
4
4
4
4
4
4
4
INT
TIME
32
32
31
31
31
31
31
31
31
31
31
31
31
31
31
31
31
31
31
ALTITUDE
(FEET)
10,500
5000
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
HEADING
E
W
W
E
W
W
E
E
E
W
W
E
W
E
W
E
W
E
VISIBILITY
(MILES)
8
3-4
3-4
3-4
<3
<3
<3
<3
<3
6
>10
>10
>10
>10
>10
>10
>10
>10
SPEED
(MPH)
110
112
110
130
117
117
133
133
133
130
120
120
120
120
120
120
120
120
A-6
-------
DATE
August 8, 2002
LINE
81
82
73
72
71
70
62-2
61-2
60
61-1a
62-1
63
64
65
66
67
69
68b
61-1b
FILE
9
10
11
12
13
14
15
16
1
2
3
4
5
6
7
8
9
10
11
START
TIME
14:23:41
14:30:21
14:37:37
14:47:04
14:56:29
15:07:05
15:21:57
15:25:31
13:09:19
13:30:20
13:49:53
14:06:54
14:23:12
14:39:09
14:54:03
15:09:00
15:22:45
15:37:23
15:48:19
STOP
TIME
14:27:42
14:33:08
14:44:19
14:54:24
15:04:30
15:15:44
15:23:00
15:27:26
13:27:22
13:46:49
14:04:36
14:21:54
14:37:35
14:52:30
15:07:00
15:21:00
15:34:56
15:44:25
15:55:02
/STOP
4
4
4
5.6
5.6
5.6
5.6
5.6
4
4
4
4
4
4
5.6
5.6
5.6
5.6
5.6
INT
TIME
31
31
31
31
31
31
31
31
31
31
31
31
31
31
31
31
31
31
31
ALTITUDE
(FEET)
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
10,500
HEADING
W
E
W
E
W
E
E
W
W
E
W
E
W
E
W
E
W
E
E
VISIBILITY
(MILES)
>10
>10
>10
>10
>10
>10
>10
>10
10
10
10
10
10
10
10
10
10
10
10
SPEED
(MPH)
120
120
120
120
120
120
120
120
120
120
120
120
120
120
120
120
120
120
125
A-7
-------
DATE
LINE
59
58
FILE
12
13
START
TIME
15:58:43
16:18:12
STOP
TIME
16:16:05
16:35:44
/STOP
5.6
5.6
INT
TIME
31
31
ALTITUDE
(FEET)
10,500
10,500
HEADING
W
E
VISIBILITY
(MILES)
10
10
SPEED
(MPH)
125
122
A-8
-------
APPENDIX B
Protocol for Collecting Representative Ground Truth Land Covers
for the Little Miami River Watershed 2002
B.1. INTRODUCTION
The purpose of this protocol is to obtain example landcovers within the Little
Miami River Watershed to assist in the classification of remotely-sensed imagery during
the period of July 1 to July 30, 2002.
B.2. EQUIPMENT NEEDED
1. Field sheets (see Attachment A)
2. Highway maps
3. Map of assigned random lat/long starting points
4. GPS (calibrated by GPS Coordinator beforehand)
5. Compass
6. Digital or 35mm camera
7. Range finder (optional)
8. Adequate disk space on digital camera or several rolls of 35mm film
9. New and/or fully charged batteries for GPS unit and camera
10. Copies of official letter for private landowners
11. Pencils
12. Tape measure
B.3. PROCEDURE
On the day before going out to the field, all GPS units and their respective
manuals (if available) need to be temporarily given to the GPS coordinator. The GPS
coordinator will standardize all the units with respect to the datum, spheroid, and the
lat/long format to be used, and return them to the field crew the following day.
Using highway maps and GPS, proceed to first assigned random starting point
(see Attachment B).
If a landcover of interest (see Section B.4 below) exists at the starting point, fill
out your first record here.
From this point, ascertain how to travel and look for additional land covers within
a mile radius of the starting point. This can be done by traveling in four different
directions emanating from this point (e.g., N, S, E, W) or as best you can throughout this
area.
B-1
-------
While traveling throughout the area (i.e., within a mile radius of each starting
point), locate, stop and fill out a record for as many unique landcovers as possible. Use
only one (1) field sheet per landcover found. Once a particular landcover is
recorded around a random starting point do not provide repeated records of this type of
landcover unless it differs significantly (e.g., an agricultural crop at different stages of
growth, or till or no-till, etc.) or unless you're beginning at a new random starting point
within the watershed.
For each type of landcover found, ideally move to a point where you are
surrounded by the landcover for 100m on all sides (approximately the length of a
football field). If the GPS signal is poor (e.g., while located in a forest with dense
canopy or an urban setting with tall buildings) or you are not able to gain access to a
private site, attempt to take the coordinates of a point just outside of the landcover area
and using a compass and range finder (if available) note on the field sheet where these
GPS coordinates are with respect to the landcover being sampled (e.g., northeast
corner of corn field). If a GPS reading would ideally be taken on private property,
receive permission first.
Once in position:
1) set the GPS unit in a stationary position and allow the GPS to acquire a 3D fix
(i.e., on at least 4 satellites) for at least 5 minutes before recording lat/long
coordinates,
2) record each satellite # with a solid black bar fix (e.g., on the Garmin units), or
otherwise the signal strength of each satellite being used by the GPS unit.
3) take GPS measurements at this single point for a minimum of three minutes,
recording all latitude and longitude coordinates displayed by the GPS unit for
up to twelve unique sets of coordinates.
4) take a picture or digital image of the landcover sampled, and using a
compass, note on the field sheet the view and orientation of the picture or
image (e.g., "corn plants 6 inches high taken from sample point in field
looking NW," or "Geist reservoir taken from SE shore looking NE").
5) record how to cross-reference the camera images with the field sheets (note:
if you run out of camera film or file space for a digital camera, please provide
as detailed a description of the landcover and its adjacent surrounding area
as possible).
6) fill out the remaining field sheet in as much detail as possible including
information on any notable landmarks nearby and any characteristics about
the landcover surface which might assist in classifying the remotely-sensed
imagery later (e.g., is the terrain flat or hilly? What is the name/type, color,
height, or texture of the vegetation or other surface sampled?).
B-2
-------
B.4. LAND COVER CLASSES OF INTEREST AND THEIR DEFINITIONS (Modified
from the NLCD Land Cover Classification System Key - Rev. July 20, 1999)
B.4.1. Land Cover Classification System Key.
Water
Developed
Concrete
Asphalt
Other construction material
Low Intensity Residential
High Intensity Residential
Commercial/Industrial/Transportation
Barren
Bare Rock/Sand/Clay
Quarries/Strip Mines/Gravel Pits
Transitional
Forested Upland
Deciduous Forest
Evergreen Forest
Mixed Forest
Shrubland
Non-natural Woody
Orchards/Vineyards/Other
Herbaceous Upland
Grasslands/Herbaceous
Herbaceous Planted/Cultivated
Pasture/Hay
Row Crops
Corn
Soybeans
Small Grains
Wheat
Fallow
Urban/Recreational Grasses
Wetlands
Woody Wetlands
Emergent Herbaceous Wetlands
B-3
-------
B.4.2. Land Cover Class Definitions.
Water. All areas of open water; typically 25% or greater cover of water (per pixel).
Developed. Generally defined as areas of intensive human use with a high
concentration (30% or higher) of constructed materials (e.g., asphalt, concrete,
buildings, glass and metal structures, etc.). If possible, look for and take coordinates in
the center of large areas (100m by 100m) which are uniformly asphalt or concrete, or
composed of other constructed materials. In addition, record points in the center of the
following general areas too:
Low Intensity Residential - Includes areas with a mixture of constructed materials
and vegetation. Construction materials account for 30-80% of area. Vegetation
may account for 20-70% of area. These areas most commonly include single-
family housing units. Population densities will be lower than in high intensity
residential areas.
High Intensity Residential - Includes highly developed areas where people reside
in high numbers. Examples include apartment complexes and row houses.
Vegetation accounts for less than 20% of the cover. Constructed materials
account for 80-100% of the cover.
Commercial/Industrial/Transportation - Includes infrastructure (e.g., roads,
railroads, etc.) and all highly developed areas (e.g., central business district,
shopping or strip mall, warehouses, etc.) not classified as High Intensity
Residential.
Barren area. Areas characterized by bare rock, gravel, sand, silt, clay, or other earthen
material with little or no "green" vegetation present regardless of its inherent ability to
support life. Vegetation, if present, is more widely spaced and scrubby than that in the
"green" vegetated categories; lichen cover may be extensive.
Bare Rock/Sand/Clay - perennially barren areas of bedrock, desert pavement,
scarps, talus, slides, volcanic material, glacial debris, beaches, and other
accumulations of earthen material.
Quarries/Strip Mines/Gravel Pits - Areas of extractive mining activities with
significant surface expression.
Transitional - Areas of sparse vegetative cover (less than 25% of cover) that are
dynamically changing from one land cover to another, often because of land use
activities. Examples include forest clearcuts, a transition phase between forest
and agricultural land, temporary clearing of vegetation, new construction sites,
and changes due to natural causes (e.g., fire, flood, etc.).
Forested Upland. Areas characterized by tree cover (natural or semi-natural woody
vegetation, generally greater than 6 meters tall); tree canopy accounts for 25-100% of
the cover.
B-4
-------
Deciduous Forest - Areas dominated by trees where 75% of more of the tree
species shed foliage simultaneously in response to seasonal change.
Evergreen Forest - Areas dominated by trees where 75% of more of the tree
species maintain their leaves year-round. Canopy is never without green foliage.
Mixed Forest - Areas dominated by trees where neither deciduous or evergreen
species represent more than 75% of the cover present.
Shrubland. Areas characterized by natural or semi-natural woody vegetation with
aerial stems, generally less than 6 meters tall, with individuals or clumps not touching to
interlocking. Both evergreen and deciduous species of true shrubs, young trees, and
trees or shrubs that are small or stunted because of environmental conditions are
included.
Shrubland - Areas dominated by shrubs; shrub canopy accounts for 25-100% of
the cover. Shrub cover is generally greater than 25% when tree cover is less
than 25%. Shrub cover may be less than 25% in cases when the cover of other
life forms (e.g., herbaceous or tree) is less than 25% and shrubs cover exceeds
the cover of the other life forms.
Non-natural Woody. Areas dominated by non-natural woody vegetation; non-natural
woody vegetative canopy accounts for 25-100% of the cover. The non-natural woody
classification is subject to the availability of sufficient ancillary data to differentiate non-
natural woody vegetation from natural woody vegetation.
Orchards/Vineyards/Other - Orchards, vineyards, groves, nurseries, ornamental
horticultural areas or other areas planted or maintained for the production of
fruits, nuts, berries, or ornamentals.
Herbaceous Upland. Upland areas characterized by natural or semi-natural
herbaceous vegetation; herbaceous vegetation accounts for 75-100% of the cover.
Grasslands/Herbaceous - Areas dominated by upland grasses and forbs. In rare
cases, herbaceous cover is less than 25%, but exceeds the combined cover of
woody species present. These areas are not subject to intensive management,
but they are often utilized for grazing.
Planted/Cultivated. Areas characterized by herbaceous vegetation that has been
planted or is intensively managed for the production of food, feed, or iber; or is
maintained in developed settings for specific purposes. Herbaceous vegetation
accounts for 75-100% of the cover.
Pasture/Hay - Areas of grasses, legumes, or grass-legume mixtures planted for
livestock grazing or the production of seed or hay crops. Visual clues to these
areas include grazing areas used by cows or horses, and fields of alfalfa hay in
different stages of cutting.
B-5
-------
Row Crops - Areas used for the production of crops, such as corn, soybeans,
vegetables, and tobacco, or cotton. In the LMR, we are most likely to find corn
and soybeans with some vegetables and tobacco.
Small Grains - Areas used for the production of graminoid crops such as wheat,
barley, oats, or rice. In the LMR, we are most likely to find wheat. Oats may be
a possibility, but this crop is more common in northerly to northeastern counties
of Ohio.
Fallow - Areas used for the production of crops that are temporarily barren or
with sparse vegetative cover as a result of being tilled in a management practice
that incorporates prescribed alternation between cropping and tillage.
Urban/Recreational Grasses - Vegetation (primarily grasses) planted in
developed settings for recreation, erosion control, or aesthetic purposes.
Examples include parks, lawns, golf courses, airport grasses, and industrial site
grasses.
Wetlands. Areas where the soil and substrate is periodically saturated with or covered
with water as defined by Cowardin, et al.
Woody Wetlands - Areas where forest or shrubland vegetation accounts for
25-100% of the cover and the soil or substrate is periodically saturated with or
covered with water.
Emergent Herbaceous Wetlands - Areas where perennial herbaceous vegetation
accounts for 75-100% of the cover and the soil or substrate is periodically
saturated with or covered with water.
B-6
-------
Attachment A: Remote Sensing Classification Field Data Sheet
Little Miami River Watershed 2002
Today's Date:
Investigator(s):
Random Starting Point
Landcover#
Picture#
3D fix? Yes / No Satellite #s:
Displayed Accuracy?
Latitude:
Begin/End Time
Longitude:
Landcover Type (circle one):
Cement
Asphalt
Other construction material
Low-density residential
High-density residential
Commercial/Industrial/Transportation
Other urban or built-up land:
Corn
Soybeans
Wheat
Pasture/Hay (horse? cow? alfalfa hay? )
Fallow (i.e., croplands temporarily out of production)
Grass (residential, golf course, park, airport, erosion
control, industrial, developed, or managed site)
Grasslands (primarily upland, natural or semi-natural,
or not intensively managed)
Orchard, grove, vineyard, nursery, or ornamental
horticultural area
Other crop or agricultural land:
Open Water
Woody Wetland
Emergent Herbaceous Wetland
Barren areas:
BareRock/Sand/Clay
Quarry/Strip Mine/Gravel Pit
Transitional (clearcuts,
clearings, burnt or flood area)
Deciduous Forest
Coniferous Forest
Mixed Forest
Shrubland
B-7
-------
Attachment B
North
Radius from center (GPS location): -100 meters
Detailed description of landcover type. For example, predominate species
present and growth stage or height of vegetation or other structures? Dry or wet
conditions? Dust or puddles present? Flat or hilly topography? Adjacent
landcovers, roads or landmarks?
B-8
-------
APPENDIX C
GENERAL RULES USED FOR OBTAINING TRAINING SET
AND SUPPLEMENTAL GROUND TRUTH DATA
The following guide sets forth the rules to follow while creating polygons of land
areas that are samples of a particular land cover/land use class. These sample areas
should be as homogenous in pattern and color as possible; in short, the "best" or
"cleanest" examples that can be found of that particular land cover/land use class.
General Rules:
1. Polygons should be no smaller than 5x5 pixels or 20m x 20 m, preferably larger.
Exceptions would be narrow linear features such as roads and riparian forest.
See the Key below for "Roads".
2. Scale should be approximately 1:5000. Urban areas will likely require a larger
scale.
3. Do not create a polygon that intersects or crosses a flight-line.
4. There should be at minimum a 10m border between the edge of the polygon and
the outside edge of the land use being sampled. In the case of roads, the border
should be at least 5 meters. The purpose of the border is to leave ample room
for error.
5. Choose only areas that are uniform in color and pattern for a particular class.
C-1
-------
APPENDIX D
METADATA
Metadata ("data about the data") created for this project was passed through the
USGS's metadata-parser software to insure that the metadata files were error free.
Metadata included in this appendix includes information about the image sources,
dates, datums, projections, resampling algorithms, processing steps, file records,
accuracy assessment, and other pertinent information associated with this geographic
product. This metadata complies with Federal Executive Order 12906.
Sample Metadata of the 4-meter Classification of the Little Miami River Watershed
Identificationjnformation:
Citation:
Citationjnformation:
Originator: U.S. Environmental Protection Agency (EPA)
Publication_Date: 2006
Title: Little Miami River Watershed Hyperspectral Classification Dataset
Geospatial_Data_Presentation_Form: Map
Publicationjnformation:
Publication_Place: EPA, Cincinnati, OH
Publisher: EPA, Cincinnati, OH
Online_Linkage: http://www.epa.gov
Larger_Work_Citation:
Citationjnformation:
Originator:
This is a land cover classification based upon Compact Airborne
Spectrographic Imager (CASI) data for the Little Miami River
watershed, an EPA project.
Publication_Date: 2006
Title:
Classification of High Spatial Resolution, Hyperspectral Remote
Sensing Imagery of the Little Miami River Watershed in Southwest
Ohio, USA
Publicationjnformation:
PublicationJ^lace: EPA, Cincinnati, OH
Publisher: EPA
Other_CitationJDetails:
This classification is based on 82 CASI flight lines acquired on
7/24/2002, 7/25/2002, 7/31/2002, 8/1/2002, 8/7/2002, 8/8/2002/,
8/9/2002.
Troyer, M.E., J. Heo and H. Ripley. 2006. Classification of High Spatial
Resolution, Hyperspectral Remote Sensing Imagery of the Little Miami
River Watershed in Southwest Ohio, USA. U.S. Environmental
Protection Agency, Office of Research and Development, National Center
for Environmental Assessment, Cincinnati, OH.
Description:
Abstract:
This is a final classification. This data set is the classification of the
Little Miami River Watershed. This data set consists of 82 mosaicked
flight lines that were analyzed according to the protocols as set forth
in a Quality Assurance Project Plan (QAPP) to determine land use/land
cover. It was created using CASI hyperspectral data by applying the object
segmentation approach in eCognition (version 3.0) and the Spectral Angle
Mapper (SAM) approach in ENVI (version 4.0). Then data were manually
edited and mosaicked. This layer is also available in selected
sub-watersheds.
Purpose:
To classify existing hyperspectral imagery and produce several land
D-1
-------
use/land cover products suitable for generating landscape metrics and
analyses by the EPA.
Time_Period_of_Content:
Time_Period_lnformation:
Range_of_Dates/Times:
Beginning_Date: 20020724
Ending_Date: 20020809
Currentness_Reference: Dates of the CASI collection
Status:
Progress: Complete
Maintenance_and_Update_Frequency: None
Spatial_Domain:
Bounding_Coordinates:
West_Bounding_Coordinate: -84.507357
East_Bounding_Coordinate: -83.547827
North_Bounding_Coordinate: 39.991621
South_Bounding_Coordinate: 38.874112
Keywords:
Theme:
Theme_Keyword_Thesaurus:
Theme_Keyword: Land Cover Analysis
Theme_Keyword: Hyperspectral
Place:
Place_Keyword_Thesaurus:
Place_Keyword: Little Miami River Watershed
Place_Keyword: Ohio
Access_Constraints: None
Use_Constraints:
Data set is not for use in litigation. While efforts have been
made to ensure that these data are accurate and reliable within
the state of the art, EPA, cannot assume liability for any
damages, or misrepresentations, caused by any inaccuracies in the
data, or as a result of the data to be used on a particular
system. EPA makes no warranty, expressed or implied, nor does
the fact of distribution constitute such a warranty.
Native_Data_Set_Environment: Erdas Imagine signed 32bit integer (.img)
Data_Quality_lnformation:
Attribute_Accuracy:
Attribute_Accuracy_Report:
According to accuracy assessment performed by Forest One,
the overall accuracy is 83.92% and 81.99% Kappa.
Each class accuracy is as follows: (Producers Accuracy/Users Accuracy)
1: Lotic = 86.73%/98.84%
2: Lentic = 70.83%/85.00%
3: Forest = 95.00%/89.62%
4: Corn = 88.33%/95.50%
5: Soybean = 87.70%/90.68%
6: Wheat = 47.37%/100.00%
7: Dry Herbaceous = 82.00%/58.57%
8: Grass = 75.00%/80.65%
9: Urban Barren = 82.61 %/90.48%
10: Rural Barren = 64.52%/45.45%
11: Urban/Built = 95.00%/92.23%
The validation points were both assembled from EPA provided ground truth
and interpreted in the lab using color aerial photographs.
There were 902 points used for accuracy assessment total.
Field collected validation points were collected concurrently
(or nearly so) with the imagery in July and August of 2002. The field data
and 2002 unrectified color aerial images were used for ground truthing of
agricultural classes. 2003 high-resolution color aerial orthophotos were
used for ground truthing of the rest of the classes: lotic, lentic,
forest, grass, urban barren, rural barren, and urban/built.
Logical_Consistency_Report:
Tests for logical consistency indicate that all row and column
positions in the selected latitude/longitude window contain data.
Conversion and integration with vector files indicates that all
positions are consistent with earth coordinates covering the same
area. Attribute files are logically consistent.
D-2
-------
Completeness_Report:
Data exists for all classes.
All pixels (other than background) have been classified.
Lineage:
Source_lnformation:
Source_Citation:
Citationjnformation:
Originator: Hyperspectral Data International, Inc.
Publication_Date: 20030428
Title:
High Resolution Remote Sensing of the Little Miami River
Watershed in Southwest Ohio, USA
Publicationjnformation:
Publication_Place: NA
Publisher: NA
Online_Linkage: NA
Type_of_Source_Media: DVD+R and CD-ROM
Source_Time_Period_of_Content:
Time_Period_lnformation:
Range_of_Dates/Times:
Beginning_Date: 20020724
Ending_Date: 20020809
Source_Currentness_Reference: Unknown
Source_Citation_Abbreviation: NA
Source_Contribution: NA
Process_Step:
Process_Description:
This dataset was created by Forest One (Joon Heo - Principal
Investigator, and Sitansu Pattnaik), Earth Satellite Corp.
(Francois Smith, Christopher Jengo, Christopher Bolton, and
Michael Diller),and Hyperspectral Data International, Inc.
(Herbert Ripley, William Jones, Laura Roy, and Michelle Warr) under
EPA purchase orders 3c-R337-NTSA, and 1C-R328-NALX. Project
Officer: Dr. Michael E. Troyer. The study area is the Little
Miami River Watershed in Southwest Ohio.
Summary:
This section outlines the classification procedure for the
Little Miami River Watershed. Atmospheric correction and
rectification were applied to the raw hyperspectral imagery.
EPA field data, rectified aerial photography (Orthophotos),
and unrectified aerial images were used for guidance
in training point selection. The hyperspectral data were then
segmented into three Level 1 classes (Urban, Rural, Water) in
eCognition. The training sets were used as inputs into Level
2 classifications within the Urban and Rural Level 1 classes.
Level 2 classification was performed with the Spectral
Angle Mapper (SAM) method within ENVI. The resulting clusters
of data were combined into categories related to the
classification scheme. QA/QC was performed often resulting
in a re-combination of clusters to better represent the final
classes. Some categories were lumped to correspond to the final
classification scheme. Data were classified by flight line.
Then the data were manually edited and mosaicked. Then a final
QA/QC was completed.
Pre-Processing steps:
The data received for this project were in a state ready
for classification. Hyperspectral Data International, Inc
performed atmospheric correction via Atmosphere Correction
Now (ACORN) software at the time of data acquisition.
Initial data examination by Forest One and Earth Satellite
Corp. determined that cross-track illumination variations
on the boundary could be handled by manual editing after
classifications. Classification was performed on data
geo-rectified using nearest neighbor resampling.
Field-Collected Data:
The EPA made available field data collected by EPA
personnel during the time of image data acquisition in 2002.
D-3
-------
The 390 points were used as a guide in selecting training
sets for classification.
Classification:
Classification for this project was performed in a hierarchical
manner. First, a Level 1 classification was performed.
Training sets guided by EPA groundtruthing and high-resolution
color aerial photography were selected for the Level 1 classes
(Urban, Rural, Water). eCognition was used to perform the
segmentation. QC work on the segmented image was performed
to reduce the occurrence of obviously misclassified pixels. For
instance, dark rooftops are often misclassified as water. The
next step was to perform Level 2 classification for the Urban
and Rural Level 1 classes. Training sets were chosen for the
Level 2 classification. For Urban, these included Water, Grass,
Urban Barren, Urban/built, and Forest. For Rural, these included
Water, Corn, Soybean, Dry Herbaceous, Rural Barren, Urban/Built,
and Forest.
Water and Forest were included in Level 2 classification in case
any pixels remained improperly classified from Level 1. The Level
2 classification was performed using the Spectral Angle Mapper
algorithm in ENVI. The Level 1 and Level 2 classes were then
combined, and QA/QC was performed to locate any misclassified pixels.
Also, water was differentiated to lotic and lentic at this point.
The final classes for this project are:
-99:Background
1: Lotic
2: Lentic
3: Forest
4: Corn
5: Soybean
6: Wheat
7: Dry Herbaceous
8: Grass
9: Urban Barren
10: Rural Barren
11: Urban/Built
99: Unclassified
Ancillary Datasets:
Non-CASI datasets used are 2003 high-resolution color aerial
orthophotos from Aerials Express, acquired in July, 2003, and 2002
unrectified color aerial images from the Center for Mapping at the
Ohio State University, The acquisition dates of 2002 images were
between Aug. 30, 2002 and Sept. 13, 2002. EPA provided ground
truth, 35 shape files of sub-basins as well as transportation and
hydrographic vectors for the Little Miami River Watershed.
Post-Processing Steps:
The 82 flight lines were classified separately and then mosaicked
to produce the land use/land cover for the entire watershed at 4m
spatial resolution. The 4m dataset was aggregated to produce an
additional 30m spatial resolution land use/land cover product too.
Data products at 4m and 30m spatial resolution were also created for
35 selected sub-basins. Data were later subset by sub-watershed.
Process_Date: 20040215
Process_Contact:
Conta conformation:
Contact_Person_Primary:
Contact_Person: Dr. Michael E. Troyer
ContactjDrganization: U.S. Environmental Protection Agency
Contact_Address:
Address_Type: mailing and physical address
Address: 26 West Martin Luther King Drive
City: Cincinnati
State_or_Province: OH
Postal_Code: 45268
Country: USA
Contact_Voice_Telephone: 513-569-7399
Contact_Electronic_Mail_Address: troyer.michael@epa.gov
D-4
-------
Spatial_Reference_lnformation:
Horizontal_Coordinate_System_Definition:
Planar:
Map_Projection:
Map_Projection_Name: Albers Conical Equal Area
Albers_Conical_Equal_Area:
Standard_Parallel: 29.5
Standard_Parallel: 45.5
Longitude_of_Central_Meridian: 96 West
Latitude_of_Projection_Origin: 23 North
False_Easting: 0.00000
False_Northing: 0.00000
Planar_Coordinate_lnformation:
Planar_Coordinate_Encoding_Method: Row and column
Coordinate_Representation:
Abscissa_Resolution: 4 meters
Ordinate_Resolution: 4 meters
Planar_Distance_Units: Meters
Geodetic_Model:
Horizontal_Datum_Name: North American Datum 1983
Ellipsoid_Name: GRS80
Semi-major_Axis: 6378137.0
Denominator_of_Flattening_Ratio: 298.257
Entity_and_Attribute_lnformation:
Detailed_Description:
Entity_Type:
Entity_Type_Label: Little Miami River Watershed, Southwest Ohio, USA
Entity _Type_Definition:
Little Miami River Watershed as delineated by EPA
and extent of imagery
Entity_Type_Definition_Source: EPA
Attribute:
Attribute_Label: Land Cover Classification
Attribute_Definition: Land Cover Classification as determined by EPA
Attribute_Definition_Source: EPA
Attribute_Domain_Values:
Enumerated_Domain:
Enumerated_Domain_Value: -99 Background
Enumerated_Domain_Value_Definition:
This class contains no data due to data voids.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 1 Water - Lotic
Enumerated_Domain_Value_Definition:
Open water associated with running water system, such as a river
or stream. Such waterways typically have a defined channel
and an associated floodplain.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 2 Water - Lentic
Enumerated_Domain_Value_Definition:
Open water associated with still water system, such as lakes,
reservoirs, potholes, and stock ponds. Such bodies typically
do not have a defined channel or associated floodplain.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 3 Forest
Enumerated_Domain_Value_Definition:
Contains either or both deciduous and coniferous trees in any degree
of mixture, single stemmed, woody vegetation with canopy spanning
greater than 4m and tree canopy accounting for 25-100 percent of
the cover.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 4 Corn
Enumerated Domain Value Definition:
D-5
-------
Area under cultivation of food and fiber,
where corn is the primary crop.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 5 Soybean
Enumerated_Domain_Value_Definition:
Area under cultivation of food and fiber where soybeans are the
primary crop.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 6 Wheat
Enumerated_Domain_Value_Definition:
Area under cultivation of food and fiber where wheat is the primary
crop.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 7 Dry Herbaceous
Enumerated_Domain_Value_Definition:
Dominated by dry and/or less vigorous herbaceous types of
vegetation; herbaceous vegetation accounts for no less than 25%
of the cover. This class mainly includes naturally occurring and
unmanaged herbaceous vegetation, and dried out, unhealthy, or
stressed crop. Dry Herbaceous vegetation prevailed in crop fields
as well as natural fields, due to a high degree of drought in the
summer of 2002. These dry herbaceous types of vegetations had little
chlorophyll content and very similar spectral signatures. Not
enough variation was present in the spectral signatures to further
classify this class into different vegetative species.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 8 Grass
Enumerated_Domain_Value_Definition:
Dominated by cultivated grasses planted in developed settings for
recreation, erosion control, or aesthetic purposes. Examples include
well-watered parks, lawns, golf courses, airport grasses, and
industrial site grasses.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 9 Urban Barren
Enumerated_Domain_Value_Definition:
Composed of bare soil, rock, sand, silt, gravel, or other earthen
material with little (less than 25%) or no vegetation within urban
areas. Examples in this class include exposed soil in urban areas
and construction sites.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 10 Rural Barren
Enumerated_Domain_Value_Definition:
Composed of bare soil, rock, sand, silt, gravel, or other earthen
material with little (less than 25%) or no vegetation in rural areas.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated_Domain_Value: 11 Urban/Built
Enumerated_Domain_Value_Definition:
Areas covered by structures and impervious surfaces in urban,
suburban, and rural areas. Buildings, parking lots, and
paved roads typically fall into this class.
Enumerated_Domain_Value_Definition_Source: EPA
Enumerated_Domain:
Enumerated Domain Value: 99 Unclassified
D-6
-------
Enumerated_Domain_Value_Definition:
This class includes areas of image gaps among flight lines and
cloud cover where land cover classification is impossible.
Enumerated_Domain_Value_Definition_Source: EPA
Distributionjnformation:
Distributor:
Contactjnformation:
Contact_Organization_Primary:
Contact_Organization:
National Center for Environmental Assessment, Office of Research
and Development, U.S. Environmental Protection Agency
Contact_Person: Dr. Michael E. Troyer
Contact_Address:
Address_Type: mailing and physical address
Address: 26 West Martin Luther King Drive.
City: Cincinnati
State_or_Province: OH
Postal_Code: 45268
Country: USA
Contact_Voice_Telephone: 513-569-7399
Contact_Electronic_Mail_Address: troyer.michael@epa.gov
Resource_Description: Little Miami River Watershed Hyperspectral Classification
Distribution_Liabiiity: NA
Standard_Order_Process:
Digital_Form:
Digital_Transfer_lnformation:
Format_Name: Erdas Imagine signed 32bit integer (.img)
Digital_Transfer_Option:
Offline_Option:
Offline_Media: DVD or CD-ROM
Recording_Format: ISO 9660
Compatibilityjnformation:
ISO 9660 format allows the media to be read
by most computer operating systems.
Fees: NA
Metadata_Reference_lnformation:
Metadata_Date: 20040217
Metadata_Review_Date: 20060330
Metadata_Contact:
Contactjnformation:
Contact_Organization_Primary:
Contact_Organization:
National Center for Environmental Assessment, Office of Research
and Development, U.S. Environmental Protection Agency
Contact_Person: Dr. Michael E. Troyer
Contact_Address:
Address_Type: mailing and physical address
Address: 26 West Martin Luther King Drive
City: Cincinnati
State_or_Province: OH
Postal_Code: 45268
Country: USA
Contact_Voice_Telephone: 513-569-7399
Contact_Electronic_Mail_Address: troyer.michael@epa.gov
Metadata_Standard_Name:
FGDC (Federal Geographic Data Committee)
CSDGM (Content Standard for Digital Geospatial Metadata)
Metadata Standard Version: FGDC-STD-001-1998
D-7
------- |