Kentucky GAP Analysis
Vegetation Mapping Annual Report
April 20, 1998
The vegetation mapping portion of the Kentucky GAP
project began in October 1997 with the initiation of a pilot project in west central
Kentucky. The pilot project was designed to investigate fundamental questions arising from
the implementation of a large-scale mapping project that relies on remotely sensed data as
primary input. Image related issues addressed included pre-processing, classification,
post classification image analysis, collection of ancillary data, GIS integration and
modeling. Various sources of ground truth were acquired, input into digital form,
summarized and cross-walked to the Southeastern Region TNC classification. Preliminary
results indicated that Landsat TM data could be used effectively to at least Anderson
level II, but the integration of spatially explicit ancillary data was necessary for
decision tree modeling purposes to achieve a more refined vegetation classification.
Efforts in the immediate future will focus on refining the vegetation map of the pilot
area and the re-assessment of air video.
TABLE OF CONTENTS
1.1 Vegetation of Kentucky *
1.2 Study Area *
Data Collection *
2.1 Ground Truth Data *
3 Methods *
3.1 Computer Assisted
2.1.1 Mammoth Cave National Park (MCNP) *
2.1.2 Pennyrile State Forest (PSF) *
2.1.3 Forest Stewardship Management Plans *
2.1.4 Fort Campbell *
2.1.5 Air Video *
2.2 Digital Data *
2.2.1 Satellite Imagery *
2.2.2 KLC Imagery *
2.2.3 MRLC Imagery *
2.2.4 Elevation Models *
2.3 GIS Databases *
2.3.1 Kentucky Land Cover *
2.3.2 National Wetlands Inventory *
2.3.3 STATSGO Soils Map *
2.3.4 Other Ancillary Data *
2.4 Metadata *
3.1.1 Image Processing *
220.127.116.11 Geometric Rectification *
18.104.22.168 Radiometric Rectification *
22.214.171.124 Topographic Attenuation *
3.1.2 Image Stratification *
4 Discussion *
5 References *
6 Appendix A List of
available imagery *
7 Appendix B List of
hardware and software *
3.1.3 Unsupervised Classification *
3.1.4 Modeling *
3.2 Vegetation Mapping *
3.2.1 TNC Classification *
3.2.2 Class Assignment and Labeling *
3.3 Accuracy Assessment *
A Memorandum of Agreement between the Mid America Remote
sensing Center (MARC) at Murray State University and the Kentucky Department of Fish and
Wildlife Resources (KDFWR) was completed in July 1997 initiating the vegetation mapping
portion of the Kentucky Gap Analysis Project. A vegetation ecologist (Mark Drew) was hired
in August 1997 and an image analyst (Geoff Ghitter) was hired in October 1997.
In October 1997 a pilot project was initiated with the
- to acquire information regarding vegetation distribution,
soils, physiography and geology and other appropriate data sets from state and federal
- to conduct field work and to otherwise collect appropriate
ground truth data to be used for image classification and accuracy assessment,
- to test the applicability of various classification schemes
over natural areas,
- to test appropriate methodologies for obtaining suitable
classification results from Landsat Thematic Mapper (TM) digital imagery and,
- to conduct research and to communicate with GAP cooperators
for the purpose of refining the chosen methodology to suit Kentuckys unique
physiographic characteristics and the type of ancillary data that are available.
The purpose of this report is to update and apprise the
state, federal and private cooperators with an interest in Kentucky GAP of the
progress-to-date regarding these objectives.
Vegetation of Kentucky
The vegetation of Kentucky is diverse in character and
vast in aerial extent. An estimated 40% of Kentucky remains forested and in a natural or
semi-natural condition. Kentucky is located entirely within the Deciduous Forest Formation
of eastern North America, and has been described as Mixed Mesophytic Forest in eastern
Kentucky and Western Mesophytic Forest throughout most of central and western Kentucky
(Braun 1950). The flora of Kentucky contains a variety of geographical elements ranging
from Appalachian species and species with northern affinities in the east to Coastal Plain
and southern species in western and south-central Kentucky (Wharton & Barbour, 1973).
The diverse vegetation of Kentucky is, to some extent, a function of the diverse geology
and soils. Ages of exposed bedrock and sediments range from Ordovician to recent; soils
are derived from limestone, dolomite, sandstone, shale, loess, and alluvium (Karan &
An image subset (Path 21, Row 34) was selected to serve as
our pilot project study area. Chosen for its proximity to Murray and the availability of
substantial and varied ground truth information, the study area is 98 km x 160 km and
extends from Pennyrile State Forest in the west to the eastern boundary of Mammoth Cave
National Park (MCNP) and from the Tennessee border in the south to the northern border of
MCNP (Figure 1). Located in west-central Kentucky, parts of five of the states nine
physiographic regions fall within its borders .
The pilot area is entirely within the Western Mesophytic
Forest region (Braun 1955), and is floristically diverse containing a possible 37 forest
alliances. The majority of the area is comprised of the Western Coalfield (WC) and Western
Pennyroyal (WP) soil regions with small portions of Floodplains/Terraces and Thick Loess
Belt soil regions. The geologic formations of the WC soil region are mainly sandstone,
siltstone, and shale of the Pennsylvanian system, and is characterized by numerous faults
and escarpments. The geologic formations of The WP are limestones of the Lower
Mississippian system characterized by a gently sloping karst landscape. Most of the WP
soil region is presently cleared for agriculture, whereas the WC soil region is relatively
densely forested. Climate of the pilot area is characterized by warm humid summers and
moderately cold winters. Average annual precipitation is approximately 125 cm. Mean
elevation is 176 m and ranges from 102 m to 294 m.
Ground Truth Data
Mammoth Cave National Park
Vegetation data were provided to the National Park
Service and KY GAP in the form of a final report by Ball State University. Summarized
tables of forest species composition including density, dominance, relative dominance,
frequency, relative frequency, and importance value were provided for plot data and
point-centered quarter (PCQ) transect data. Geographic locations of each plot and PCQ
transect were provided on photocopied 1:24000 scale topographic maps. Plot and PCQ
transect data was then screen digitized over digital raster graphics in order to label the
classified images. Preliminary results of data in similar format was received from a
forest succession study in Mammoth Cave National Park authored by Virginia McDaniel of
Western Kentucky University.
A GIS coverage of MCNP showing 17 different forest classes
was received in mid-November was exported to Arc/Info format from the GRASS GIS in use at
the park. This coverage was produced by digitizing a map created in 1934 by Ellsworth, who
spent three years surveying the proposed national park site and creating the map. This
coverage did not overlay precisely with the rectified digital images and had to be
subjected to further processing.
An Arc/Info routine was used to rubber sheet
the coverage by identifying control points from 1:24000 USGS map sheets and the Arc/Info
coverage. A total of 70 points were identified, however, unlike raster image rectification
wherein each control point has an individual influence on the transformation and can be
quantitatively measured as to its accuracy, control points in Arc/Info are considered
correct by default and therefore the coverage was transformed around those
points. A visual assessment of the output indicated that it fitted much better than the
original coverage. However, there was no way to judge this quantitatively. Since this
coverage is being used for visual assessment and guidance only, the lack of quantitative
accuracy measures was not considered a hindrance.
Pennyrile State Forest
Pennyrile State Forest (PSF) personnel have produced a
site classification based on a grid of over 1500 sample plots. Quantitative species data
were collected within variable radius prism plots nested within one-acre. Each one-acre
plot was assigned to a forest class based on a classification developed by the PSF
personnel. In order to cross-walk these classes into the TNC alliance level
classification, a subset of plots was re-visited and additional species data were
recorded. The 1500 plots were digitized from topographic base maps, but because they had
not been accurately located using GPS technology, a window of pixels surrounding each plot
is being considered when assigning possible alliances.
Forest Stewardship Management
Approximately 100 FSM plots have been collected and
digitized. These plot data were borrowed from the Kentucky Division of Forestry and were
originally used for assessment of woodlot potential on private lands. Data were gathered
from five counties (Muhlenberg, Todd, Logan, Butler, Christian) that all fall within the
boundaries of the pilot project. If this source of data proves to be useful, attempts will
be made to collect data from all of Kentucky. Species data are in order of dominance. As
many as 20 species are recorded at each plot. These data have been cross-walked to the TNC
classification and will be used to assess the performance of the classifier and to aid in
air video site selection. These data have been fitted to TNC alliances and will be used
for labeling unsupervised classes.
Plot data from Fort Campbell (Kentucky/Tennessee) has been
received in Arc/Info format and its associated attributes have been input to the main GAP
database. Diameter at breast height (DBH) was recorded for overstory species in each of
176 one-fifth acre plots. Total basal area by species by plot was calculated. These
calculations will be used to assign possible alliances to each plot and subsequently to
label unsupervised classes.
Air video has been used as a source of ground truth data
and for post-classification accuracy assessment (Slaymaker, et al., 1996) and has been
adopted by various state GAP projects nation-wide. However, initial assessment of air
videography as a tool in western Kentucky has so far not proven to be successful. The main
problems stem from: 1) lack of sufficient resolution in the chosen video format causing a
severe lack of discriminating power that is compounded by; 2) the extreme heterogeneity of
the eastern deciduous forest which inhibits the analyst from making conscientious
identification of individual tree species as stipulated by the methodology (Slaymaker, et
al., 1996) and; 3) the air video mission was conducted in September (1996), before maximum
leaf color separation.
Reliable ground truth data are a crucial element of any
mapping project using remotely sensed data as a primary input. The large-scale character
of state GAP projects necessitate the collection of many hundreds (or thousands) of points
that serve as input prior to classification or as a means of accuracy assessment.
Therefore, a means for extracting reliable data without field checking every point and
that satisfy classification and accuracy assessment requirements must be examined.
With the recent acquisition of a higher resolution
printer, we continue to assess the usefulness of air video in the deciduous forest of
Kentucky. We hope to train analysts to distinguish enough species from a given frame to
place the frame in an alliance with confidence and repeatability. If this approach does
not provide reliable results, two other options will be considered:
- mounting a new air video mission at a time of year more
conducive to separation of species or alliances and/or,
- consider a new form of data acquisition that would include
the use of a high resolution digital frame camera as a substitute for aerial videography.
This project will use Landsat Thematic Mapper (TM) images
of that were acquired by MARC for use in the Kentucky Landcover Mapping project (KLC).
These images are being used in conjunction with the standard Multi-Resolution Land
Characteristics Consortium (MRLC), which are also Landsat data, because the acquisition
dates of the KLC images are more recent (1994/95 compared to 1991/92). MRLC data were
acquired and used as ancillary data sets to compliment the TM imagery already in house,
especially for the discrimination and classification of coniferous forest species. Two
dates for each of the nine scenes will be used in the analysis for the Kentucky Gap
Project. Appendix A lists scene path and row information and acquisition dates for each
Two dates are being used to maximize the useful amount of
information available to a classification procedure. Two dates (usually early spring (leaf
on) and late fall (leaf off)) have been shown to increase classification accuracy however,
the majority of KLC images are both leaf on. Differences in greenness between
the spring/fall acquisition dates may be sufficient to make the use of both scenes
worthwhile since there may be color changes in the foliage that would provide additional
useful information in the classification process.
Note: Satellite imagery from KLC has been row-shifted so
that two images rather than three images per path cover the state from south to north. The
MRLC data conforms to regular path/row criteria and often three images are required for
full coverage for a path.
In March 1997, 43 compact discs containing digital
satellite scenes were received from the Multi-Resolution Land Characteristics Consortium
(MRLC) to be used to complement the data already in house. Among these data were nine
scenes imaged during leaf-off conditions adding new and valuable information to that
already available. The leaf off images covering the pilot project study area were employed
to help identify and classify stands of coniferous forest that had previously been subject
to confusion (see section 3.1.3).
USGS, Level 1, 7.5 minute digital elevation models (DEMs)
accessed from the Kentucky Geological Survey website were subset and mosaiced to fit the
study area boundaries. A filtering routine (Brown and Bara, 1994) was used to remove
systematic noise that is inherent in Level 1 DEM products (Garbrecht and Starks, 1995). An
interpolation routine was implemented to smooth slivers of no-data left between the
DEMs and their derivatives (slope, aspect) have become
standard tools used in land cover classifications. Although the pilot project study area
in western Kentucky does not reflect extremes in elevation, vegetation responses to
topographic variables can be modeled to aid in the classification process. In addition
DEMs can be used to visualize the results of classification output when they are draped
over the DEM and relationships between unsupervised classes and landscape variables can be
Kentucky Land Cover (KLC) is a remote sensing
project being conducted at MARC for Kentuckys Natural Resources and Environmental
Protection Cabinet, Division of Water that is in the final stages of mapping the entire
state to Anderson level II (Anderson, 1976). In order to present a complete coverage of
the state, and to minimize time spent by GAP staff in classifying areas of no particular
interest for GAP analysis, all classes not represented by the TNC classification (e.g.
urban, agricultural) will be derived from KLC.
The National Wetlands Inventory (NWI) data
depict the location and classification of wetland regions as defined by the U.S. Fish and
Wildlife Service (Cowardin, 1979). Digital coverages on a county-by-county basis are
available for the entire state and were used in the pilot project to delineate and subset
wetland areas for independent classification. Wetlands data from the 16 counties wholly or
partially in the study area were merged into a single database and then clipped to match
the study area boundaries.
The NWI databases consist of many thousands of polygons,
most of which fall below the minimum mapping unit (MMU) standard as described in the GAP
handbook. Additionally, software limitations (i.e. the inability to deal with more than a
certain number of polygons) necessitated a reduction in the number of polygons contained
in these coverages prior to subsetting. All polygons below the MMU threshold (2 ha or 22
pixels) were either eliminated outright, or if spatially adjacent to larger polygons,
merged. Following this procedure the number of wetland polygons in the pilot study area
was reduced from over 20,000 to 3300 while not significantly altering the polygons of
Non-forested wetland polygons will be used in their
present form (after cross-walking to TNC alliances) while forested polygons will be
digitally classified and subjected to the same classification, ground truth and accuracy
assessment procedures as the rest of the states forested areas.
STATSGO Soils Map
In 1985 the Natural Resources and
Environmental Protection Cabinet (NREPC) Division of Conservation began digitizing the
interpretive general soil maps as part of the national Cooperative Soils Survey Program.
In counties where no published soil surveys existed, the experience of a local soil
scientist was used to extrapolate lines from adjoining mapped areas or to update previous
The soil map used was acquired from the NREPC web site but
originated from the State Soil Geographic (STATSGO) data base. Soil maps for STATSGO are
compiled by generalizing more detailed (SSURGO) soil survey maps. Where more detailed soil
survey maps were not available, data on geology, topography, vegetation, and climate were
assembled, together with Landsat images. Soils of like areas were studied, and the
probable classification and extent of the soils were determined. Map unit composition for
STATSGO maps were determined by transecting or sampling areas on the more detailed maps
and expanding the data statistically to characterize the whole map unit. Using the United
States Geological Surveys (USGS) 1:250,000 scale, 1- by 2-degree quadrangle series
as a map base, the soil data were digitized by line segment (vector) method to comply with
national guidelines and standards. Data for the STATSGO data base were collected in 1- by
2-degree topographic quadrangle units and merged and distributed as statewide coverages.
Features were edge matched between states. The map unit composition and the proportionate
extent of the map unit components were also matched between states (USDA, 1994).
Soil units are a function of the parent material and so
correspond roughly to the physiographic provinces of Kentucky. Descriptions of the soil
regions and their corresponding physiographic names are given in Table 1.
Table 1 Physiographic areas and
description of soils
soils on floodplains, terraces and colluvial fans
tertiary and Cretaceous coastal plain sediments
sandstone, siltstone and shale
Mississippian limestone of eastern Pennyroyal
limestone of western Pennyroyal
Mississippian and Devonian siltstone and shale
and Ordovician limestone and calcareous siltstone
Mississippian and Pennsylvanian sandstone, siltstone and shale
Other Ancillary Data
Other ancillary data for image
classification and map creation include:
- National Aerial Photography Program (NAPP) 1:40,000 color
- 1:24,000 USGS topographic maps,
- Digital Raster Graphics (scanned images of USGS topographic
- GIS coverages from the Kentucky Natural Resources and
Environmental Cabinet including roads, state boundary, quad boundaries (1:24,000,
1:100,000) and other cartographic coverages.
We have instituted an on-going program to keep metadata
current. Sample metadata output for the pilot project will be compiled when the pilot
project nears completion.
Due to the broad scope and large geographic extent of this
project, a single methodology designed to address every issue, data type or source would
be impossible to implement. Instead we intend to use a variety of methods that will vary
depending on the particular source and type of data for a particular region or image, the
amount of reliable ground truth available and the variety of dependable ancillary data for
In general, in regions where only point data exists, that
data will be used to assess and name classes derived from an unsupervised classification.
Unsupervised classes will be merged or deleted and a rule based decision tree model will
be implemented using Arc/Info Grid that will sort pixels into classes
depending on their spectral and spatial characteristics. A model for this method has been
employed by Tennessee GAP (Jeanette Jones, pers. comm.) and will be available as a
Computer Assisted Mapping
In all, 27 Landsat TM scenes will be used for vegetation
mapping, 18 (nine spring, nine fall) from KLC and nine (leaf-off) from MRLC. There is some
debate in the remote sensing community as to the proper strategy to employ when using
multiple images in large scale mapping projects (Vogelmann, et al., 1998). Should the
images be mosaiced first and then classified, or should each image be treated separately
and the classification results aggregated later? Mosaicing images prior to classification
allows for consistent results across physiographic regions or vegetation classes that
cross image boundaries (Edwards et al., 1995) and suspends the need for tedious edge
matching, however the loss of radiometric consistency from scene to scene may bring into
question the validity of the spectral classes.
Images from the same path are from the same date affording
the opportunity to mosaic image pairs with a minimum loss of confidence in the radiometric
consistency. However, it was decided that mosaicing would not occur between images of
different paths. Thus the entire state will be covered by five image databases, each of
two mosaiced scenes. The leaf-off MRLC data will complement these scenes and will be
subset and fitted according to need. Some of these data were also collected on the same
date along satellite paths, and will be mosaiced to reflect this.
Image processing consists of the following steps:
- geometric rectification,
- radiometric rectification (normalization),
- image mosaicing, and
- image stratification.
Each procedure will be conducted according to the
specifications mentioned in the following sections.
Geometric rectification was performed on all image data to
a Root Mean Square (RMS) accuracy of ± .5 pixels. Ground control points for each corner
of each image were provided by EDC. However, additional control points were added in an
even manner across each image to provide for optimal registration and resampling. A
first-order transformation and a nearest-neighbor resampling algorithm was used to correct
each scene that was projected to a UTM grid using the NAD 27 datum and the Clarke 1866
ellipsoid. Pixels were (or will be) resampled to standard TM unit size of 30m per side.
Kentucky straddles two UTM zones (16 and 17) so a decision
will have to be made as to how to handle the image(s) where the change occurs. It may be
easiest to re-project the parts of the state that fall in zone 17 (about 1/6 of the
eastern portion of the state) to zone 16.
Ground control points were collected during the Kentucky
Land Cover project for most of the images. However some image dates were not used and
therefore have no control points. New control points have been chosen for those scenes.
Control point information including RMS error and location (UTM coordinate and quad name)
are included in the KY-GAP project binder under the heading of GCP Data.
MRLC data were received rectified and projected to UTM
coordinates using GRS 1980 ellipsoid, NAD 83 datum. These images were re-projected to the
standards mentioned above.
Radiometric rectification is the process whereby
distortions caused by the atmosphere during scene acquisition are accounted for and
corrected. However, the amount and severity of distortion can vary across a single scene
and any rectification procedure, other than that with precise atmospheric and
meteorological data, is only general and the effects on pixel brightness across a scene
may be unknown. Various researchers have approached this problem from differing points of
view, some preferring to correct all images, other preferring only to correct images when
multi-temporal data are being used in a change detection analysis. Radiometric
rectification is also very important if scenes from multiple dates are being mosaiced to
form one scene upon which analysis will take place.
Adopting somewhat of a minimalist approach, KY GAP will
use raw band values whenever possible. Since we are not mosaicing scenes from different
dates, any subsequent analysis procedures should not be substantially affected. With the
exception of topographic attenuation (section 126.96.36.199) no other radiometric correction
Topographic variables such as slope and aspect may
introduce radiometric distortion on the recorded signal in a satellite image (Jensen,
1996). Due to the orientation of the sensor in relation to the suns elevation and
azimuth over the target imaging area, certain areas of interest may appear in shadow while
others (south east facing slopes, for example) may be overexposed. During visual
inspection of early classification results, it was obvious that the image used for the
pilot project needed some kind of correction to obviate the results of topography on scene
illumination. However, it was not clear if vegetation patterns observed in the classified
image were directly a result of the topographic effect or if the resultant classes were
true, a result of their position on the landscape.
A hybrid solution designed to address this issue was
implemented by applying a topographic correction to the six fall bands from the 12 band
spring-fall data set used to create the unsupervised classification. Since variations of
topographic corrections are known to over correct in some situations and under correct in
other situations (Meyer et al., 1993), this solution allowed for removal of topographic
effect in the fall imagery while maintaining vegetation classes that may truly exist as a
result of their position on the landscape. In this way the effects of the over corrections
known to occur in weakly illuminated pixels would be minimized in the ultimate
The cosine correction described by Jensen, 1996, was
implemented on the fall data set using the following equation:
LH = radiance observed for a
LT = radiance observed over sloped
q 0 = suns zenith angle
i = suns incidence angle in relation to the
normal on a pixel
Initial examination of classification results before and
after topographic normalization indicated that a better spectral
classification has resulted. That is, the effects of topography on pixel brightness have
been mitigated to the extent that we are more confident that the spectral classes
represent ground cover classes, not classes of bright or shadowed areas as a response to
Image stratification has been used in many studies
(Bolstad and Lillesand, 1991; Lillesand, 1996; Zhu and Evans, 1994) as a strategy to
reduce confusion between classes that occur on different physiographic or soil units and
to include (or exclude) certain species of vegetation with known distributions during the
modeling phase of the project. Stratification can take many forms and is partly dependent
on the availability and quality of ancillary data as well as the specific physiographic
characteristics of the landscape that are unique to each mapping project.
Stratification involves identification of a particular
area with unique characteristics, separating it from the main image and classifying it
independently. After classification, the separate, classified entities are edge matched
and joined together to produce a final map product. In many cases physiographic provinces
have been used as a coarse level stratification. However, a suitably accurate
physiographic map of Kentucky was not available. In its stead a 1:250,000 map of the soils
of Kentucky (STATSGO) was used to mimic the effects of physiographic stratification and
potentially add other information that could be used in subsequent modeling routines, pH,
for example. The stratification scheme adopted for the KY-Gap pilot project is as follows:
- subset all urban, agricultural and other non-forested lands
(i.e. everything except classes 4 and 6 according to Anderson, 1976.),
- subset wetland areas (data from the National Wetlands
Inventory digital database, Anderson class 6),
- identify and separate coniferous land cover from deciduous
land cover to be classified independently, and
- subset the remainder of the image according to its unique
Following this formula a set of seven independent images
comprising the pilot project study area were created:
- all wetlands in the study area,
- all coniferous in the study area, and
- deciduous only for each of five soil zones.
Testing is currently being conducted to see if there is
indeed a difference in the unsupervised classes created from each independent image. We
have already observed high amounts of class confusion between wetlands and coniferous
areas when classified from the same image. Testing is just being initiated to try and
discern the difference between the five deciduous-only classifications derived from
different soil zones.
Unsupervised image classification is the main method for
arriving at a set of ground cover classes derived from the satellite imagery that will be
used as the primary input for a model-based final map product. Unsupervised classification
is a numerical approach for grouping satellite data into clusters based on the spectral
properties of each pixel in an n-band data set. Once the data are classified, the
analyst then attempts a posteriori to assign these spectral classes to the
information classes of interest (Jensen, 1996). Often, however, the spectral classes
cannot be assigned to a particular information class because they may represent a mixture
of classes on the earths surface or because certain properties inherent in the data
may cause the classifier to group pixels based on similar spectral properties when in fact
they are completely different from the information classes of interest.
An iterative strategy to maximize the discrimination power
of unsupervised classification is called cluster busting. Cluster busting is
used to take general clusters or groups of clusters, remove them from a classification
(i.e. treat them as a separate entity) and then subject those clusters to a new
unsupervised classification of only the areas of interest. For example, the first
unsupervised classification was used to discriminate forested areas from non-forested
areas, the second iteration to discriminate coniferous forest from deciduous forest and
the third iteration to discriminate between different coniferous or deciduous spectral
clusters. Only after the third iteration were clusters initially labeled or considered for
splitting and merging.
National Aerial Photography Program (NAPP), 1:40,000,
color infrared photography available in-house for selected flight lines, was used where
possible, to label classes, or to record information as to the similarity in vegetative
characteristics for other classes. For example, of the 20 coniferous classes generated
from the unsupervised classification (third iteration), two classes were assigned to
particular coniferous species, two to the same species with a slightly different density
class, three classes to water/land edge pixels (which exhibit similar spectral
characteristics to coniferous vegetation), 8 classes to a variety of coniferous/deciduous
canopy mixes (60/40 coniferous deciduous, e.g.) while five classes remained unnamed;
either because their occurrence over the landscape was not continuous (i.e. only speckled
examples of the class occurred) or no significant cover type association could be
Recognizing the limitations of Landsat TM
digital imagery to spectrally discriminate forest cover types to the TNC alliance level,
efforts to increase accuracy and identify cover types not separable in digital
classifications are being tested. At present an intensive literature review is being
undertaken with the goal of further identifying physical variables that can be modeled to
represent significant vegetative features of the Kentucky landscape. Once modeled, these
parameters can be incorporated into a routine to further discriminate the spectral classes
from the unsupervised classification.
An example of landform modeling using DEM derivatives is a
concavity/convexity model designed to identify parts of the landscape where certain
vegetation species are known to exist. In this model (McNab, 1989; McCombs et al., 1997)
areas of concavity and convexity in the landscape were modeled using the DEM and
subsequently used to identify landforms that provide ecological niches for species that
may not be separable using spectral class alone.
The Nature Conservancy (TNC) and state
Natural Heritage Programs are currently developing a national classification for use in
the conservation of biological diversity. The classification is hierarchical with
physiognomic criteria defining the higher (formation) levels and floristic criteria at the
lower (association and alliance) levels. The Southeast Region of TNC has completed
alliance descriptions (TNC 1997) which are currently in use by the Kentucky GAP project.
The overall goal of the vegetation mapping component of Kentucky GAP project is to produce
a statewide map to the level of the alliance, or aggregations of alliances.
Class Assignment and Labeling
Forested areas were identified
using color infrared (CIR) photography to label forested areas on a 75 class unsupervised
classification. Pixels labeled as forested were then extracted and an additional
unsupervised classification was performed on the forested areas only. CIR photography was
again used to identify unsupervised classes dominated by coniferous species. Additional 25
class unsupervised classifications were then performed on coniferous and deciduous areas.
Digital ground truth from MCNP (PCQ transect locations, plot locations, McDaniel plot
locations) and from Pennyrile SF (1500 digitized points) were overlaid onto the classified
images to aid in the assignment of alliances (or groups of alliances) to unsupervised
classes. Since data points from Pennyrile SF were not GPSed, each point was analyzed
in the context of a 3x3 pixel window.
Results of these steps indicated that the deciduous class
could be reliably split into two basic classes: one dominated by Quercus species and one
dominated by Fagus grandifolia and Acer saccharum. Unfortunately, it appears that the
spectral and/or spatial resolution of Landsat TM will not allow additional separation of
deciduous forest classes in the pilot area. A model that incorporates landform index,
slope, and aspect is in the developmental stages; we hope the model will allow accurate
delineation of 1) a mesic oak community dominated by Quercus alba, Q. rubra, and Carya
spp. located on concave landforms and lower slope positions, 2) a dry oak hickory
community dominated by Quercus alba, Q. coccinea, and Q. velutina found on side slopes and
sheltered ridgetops, and 3) a xeric oak community dominated by Quercus prinus, Q.
coccinea, and Q. velutina.
Accuracy assessment has not yet taken place
pending the completion of a final map product. Also, issues relating to the collection of
data points for use in any such assessment (procedures and methods, see section 2.1.5)
need to be addressed before a scientifically defensible accuracy assessment can occur.
The Kentucky GAP vegetation mapping project began in
October 1997 and a pilot project designed to test methods and procedures was instituted
shortly thereafter. A good part of the time during the early stages of the project were
devoted to logistics, that is, identifying and gathering the resources required to
implement the operational phase of the project (Appendix B).
Each state GAP project is unique and although we can build
on and benefit from the experiences of other states, each individual state project must be
designed to synthesize the particular array of resources at its disposal. Ideas and
methodologies from other states cannot just be replicated, instead they must be made to
fit the particular circumstances both of the physical characteristics of the Kentucky
landscape and the resources available to the Kentucky GAP project. For example, many
states have used a pre-existing physiographic map to subset the area into physiographic
regions where independent classification and analysis takes place. A suitable
physiographic map was not available in Kentucky, so instead we were able to use a soils
map which closely approximates the physiographic provinces.
Upon completion of the path/row in which the pilot area is
located, attention will be focused upon far western Kentucky. Central and eastern Kentucky
will then be classified. Methodologies described above will be modified if necessary from
region to region as the data or the land cover dictate.
Anderson, J.R.; Hardy, E.E.; Roach, J.T.;
and Witmer, R.E., 1976. A Land Use and land Cover Classification System for Use with
Remote Sensor Data. USGS Professional Paper 964, Washington D.C., 28 pages.
Badger, K.S., 1997. Mammoth Cave National Park vegetation
study. Unpublished report to the National Park Service.
Bolstad, P.V. and Lillesand, T.M., 1991. Improved
classification of forest vegetation in Northern Wisconsin through a rule-based combination
of soils, terrain and Landsat Thematic Mapper data. Photogrammetric Engineering and Remote
Sensing, 57 (1), 66-74.
Braun, E.L. 1950. Deciduous forests of eastern North
America. Hafner Publishing Co. New York, New York.
Brown, D.G. and Bara, T.J., 1994. Recognition and
reduction of systematic error in elevation and derivative surface from 7½ minute DEMs.
Photogrammetric Engineering & Remote Sensing; 60 (2), 189-194.
Cowardin, L.M., Carter, V., Golet, F.C. and LaRoe, E.T.,
1979. Classification of wetlands and deepwater habitats of the U.S. U.S. Fish and Wildlife
Service, Washington D.C., Publication number FWS/OBS-79/31, 103 p.
Edwards, T.C., Homer, C.G., Bassett, S.D., Falconer, A.,
Ramsey, R.D., and Wight, D.W., 1995. Utah Gap Analysis: An environmental information
system, Final Project Report 95-1, Utah Cooperative Fish and Wildlife Research Unit, Utah
State University, Logan, Utah, 60 p.
Ellsworth, I.J., 1934. Forest cover type map survey of
proposed Mammoth Cave National Park, United States Department of the Interior,
Garbrecht, J. and Starks, P. 1995. Note on the use of USGS
Level 1 7.5-minute DEM coverages for landscape drainage analyses. Photogrammetric
Engineering & Remote Sensing; 61 (5), 519-522.
Karan, P.P., and C. Mather (eds.). 1977. Atlas of
Kentucky. University Press of Kentucky. Lexington, Kentucky.
Jensen, J.R., 1996. Introductory digital image processing:
a remote sensing perspective, 2nd Edition. Prentice-Hall Inc., Upsaddle River,
NJ. pp. 316.
Lillesand, T.M., 1996. A protocol for satellite-based land
cover classification in the upper Midwest, in Gap Analysis. In: Gap Analysis, A Landscape
Approach to Biodiversity Planning; Scott, J.M., Tear, T.H. and Davis, F.W. eds. American
Society for Photogrammetry and Remote Sensing, Bethseda MD., pp. 103-118.
McCombs, J. Klopfer, S. Morton, D. and Waldon, J., 1997.
An alternative approach to land cover mapping in complex terrain. In: GAP Analysis,
Bulletin No. 6. USGS Biological Resources Division, 62 pages.
McDaniel, V., 1997. Succession in old fields at Mammoth
Cave National Park. Unpublished report to the National Park Service
McNab, W.H., 1989. Terrain shape index: quantifying the
effect of minor landforms on tree height. Forest Science, 35 (1), 91-104.
Meyer, P., Itten, K.I., Kellenberger, T., Sandmeier, S.,
and Sanmeier, R., 1993. Radiometric correction of topographically induced effects on
Landsat TM data in an alpine environment. ISPRS Journal of Photogrammetry and Remote
Sensing 48 (4), 17-28.
Slaymaker, D.M., Jones, K.M., Griffin, C.R. and Finn,
J.T., 1996. Mapping deciduous forest in southern New England using aerial videography and
hyperclustered multi-temporal Landsat TM imagery. In: Gap Analysis A Landscape Approach to
Biodiversity Planning; Scott, J.M., Tear, T.H. and Davis, F.W. eds. American Society for
Photogrammetry and Remote Sensing, Bethseda MD., pp. 87-102.
The Nature Conservancy, 1997. An Alliance level
classification of the vegetation of the southeastern United States. The Nature
Conservancy, Southeast Conservation Science Department, Community Ecology Group, Chapel
Hill , NC.
United States Department of Agriculture, 1994. State Soil
Geographic Database (STATSGO). Miscellaneous publication number 1492.
Vogelmann, J.E., Shol, T., and Howard, S.M., 1998.
Regional characterization of land cover using multiple sources of data. Photogrammetric
Engineering & Remote Sensing; 64 (1), 45-57.
Wharton, M.E., and R.W. Barbour. 1973. Trees & Shrubs
of Kentucky. The University Press of Kentucky. Lexington, Kentucky.
Zhu, Z. and Evans, D.L., 1994. U.S. Forest Types and
Predicted Percent Forest Cover from AVHRR Data. Photogrammetric Engineering and Remote
Sensing; 60 (5), 525-531.
Appendix A List of
MARC TM Satellite Coverage on CD
*MRLC data is geo-corrected to GRS 1980 NAD
*KLC data can be geo-corrected with points that are
already in house
Bold denotes imagery to be used by KY-GAP
Appendix B List
of hardware and software
200 MHz Pentium desktop PC running Microsofts NT
operating system; configured with 128 Mb ram and 10 gigabytes of hard disk space
166 MHz laptop computer running Microsofts Windows
95 operating system configured with 64 Mb ram and 4 gigabytes of hard disk space
2 Sony Hi-8 video players
2 Panasonic color video monitors
Sony color video printer
HP 670C color inkjet printer
Tecmar tape backup (4 GB)
Software being used in the Kentucky GAP
- Erdas Imagine 8.3.1 image processing
- SAS Statistical Package
- ESRI ArcView GIS (plotting, map production, image
- ESRI Arc/Info GIS (GIS functions, AML routines, GRID)
- Windows Office Professional including Word (word
processing), Access (database), Excel (spreadsheet)
- Visual Basic 5.0 (specialized Windows programming)