Content uploaded by Jane R Foster
Author content
All content in this area was uploaded by Jane R Foster
Content may be subject to copyright.
USGS-NPS Vegetation Mapping Program
VEGETATION OF SHENANDOAH NATIONAL PARK IN RELATION TO
ENVIRONMENTAL GRADIENTS
Final Report (v1.1)
2006
Prepared for:
US Department of the Interior
National Park Service
Prepared By:
John Young
US Geological Survey
Leetown Science Center
Kearneysville, WV 25430
Gary Fleming
Virginia Department of Conservation and Recreation
Division of Natural Heritage
Richmond, VA 23219
Phil Townsend and Jane Foster
Department of Forest Ecology and Management
University of Wisconsin-Madison
Madison, WI 53706
* This document is designed to be viewed in color *
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Table of Contents:
LIST OF TABLES……………………………………………………………………… iii
LIST OF FIGURES……………………………………………………………………. iv
LIST OF CONTACTS AND CONTRIBUTORS……………………………………. v
ACKNOWLEDGEMENTS……………………………………………………………. vi
LIST OF ABBREVIATIONS AND ACRONYMS………………………………….. vii
EXECUTIVE SUMMARY……………………………………………………………..viii
1. INTRODUCTION……………………………………………………………………1
1.1. Background..............................................................................................2
1.2. Scope of Work..........................................................................................5
1.3. Study Area Description……………………………………………………...7
2. METHODS………………………………………………………………………....11
2.1. Environmental Gradient Modeling……………………………………….11
2.2. Sample Site Selection............................................................................19
2.3. Field Survey Methods………………………………………………………20
2.4. Plot Data Management and Classification Analysis...........................25
2.5. Image Processing and Classification……………………………………31
2.6. Accuracy Assessment……………………………………………………...40
3. RESULTS…………………………………………………………………………..43
3.1. Landforms and Ecological Land Units…………………………………..43
3.2. Vegetation and Accuracy Assessment Plots…………………………..46
3.3. Vegetation Classification Scheme………………………………………..50
3.4. Vegetation Map………………………………………………………………73
3.5. Accuracy Assessment……………………………………………………...83
4. DISCUSSION/CONCLUSION…………………………………………………….86
5. LITERATURE CITED………………………………………………………………89
APPENDICES
1. Vegetation Classification, Clustering, and Ordination charts
2. Subset from the National Vegetation Classification for Shenandoah
National Park
3. Key to the Natural Vegetation of Shenandoah National Park
4. Accuracy Assessment Field Form
5. Field plot metadata
6. Accuracy assessment tables
ii
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
LIST OF TABLES
TABLE 2.1 Selected publications using environmental gradient models for vegetation
community analysis……………………………………………………………………………….. 13
TABLE 2.2 Rock formations and rock type groupings of Shenandoah National Park.…… 16
TABLE 2.3 Class codes for Ecological Land Unit map (ELU)……………………………….. 17
TABLE 2.4 Environmental gradients derived for Shenandoah National Park from a
15-meter resolution digital elevation model …………………………….…..…………….…18-19
TABLE 2.5 Cover class scores used in field sampling and data analysis ………….……… 21
TABLE 2.6 Topographic / hydrologic environmental indices recorded at each
plot-sampling site …………………………………………………………………………………. 24
TABLE 2.7 Ordinal variables used in analysis for scalar topographic and soil moisture
variables estimated in the field ………………………………………………………………….. 26
TABLE 2.8 Aggregate geological classes used as dummy variables in data analysis……. 26
TABLE 2.9 Image and topographic variables selected for use in canonical linear
discriminant analysis models ……………………………………………………………………. 37
TABLE 3.1 Mapped vegetation communities of Shenandoah National park, by area ……. 81
TABLE 3.2 Cumulative accuracy by cumulative park area …………………………………. 85
iii
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
LIST OF FIGURES
FIGURE 1.1 Elevation classes of Shenandoah National Park ……………..………………. 8
FIGURE 1.2 Generalized surficial geology of Shenandoah National Park......................... 9
FIGURE 2.1 Example spectral response from AVIRIS hyperspectral sensor ..…………….34
FIGURE 2.2 Illustration of loading of outcrop classes on two canonical axes …….……….36
FIGURE 2.3 Weighted importance of variables used in July 2001 AVIRIS image model....38
FIGURE 2.4 Weighted importance of variables used in multi-temporal Landsat image
model ……...........................................................................................................................39
FIGURE 3.1 Landforms of Shenandoah National Park by area ….………………………….44
FIGURE 3.2 Example landform mapping results ………………………………………………45
FIGURES 3.3a, 3.3b, 3.3c Field sampling and AA plot locations…………...………..47,48,49
FIGURE 3.4 Number of ecological land units (ELU) per vegetation community type ……..73
FIGURES 3.5a, 3.5b, 3.5c Final vegetation map result, by district…………………...75,76,77
FIGURE 3.5d Map legend for final vegetation map …………...………………………………78
FIGURE 3.6 Image sources used as inputs to final vegetation mapping result …..………..79
FIGURE 3.7 Final class probabilities by pixel for vegetation mapping result .………………80
FIGURE 3.8 Mapped vegetation communities of Shenandoah National Park, by area …...81
iv
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
LIST OF CONTACTS AND CONTRIBUTORS
John Young, Biogeographer
USGS Leetown Science Center
11649 Leetown Road
Kearneysville, WV 25430
304-724-4469
jyoung@usgs.gov
Gary Fleming, Vegetation Ecologist
Virginia Department of Conservation and Recreation
Division of Natural Heritage
217 Governor St., 3rd Floor
Richmond, VA 23219
(804) 786-9122
gary.fleming@dcr.virginia.gov
Karen Patterson, Ecologist
Virginia Department of Conservation and Recreation
Division of Natural Heritage
217 Governor St., 3rd Floor
Richmond, VA 23219
(804)786-5990
karen.patterson@dcr.virginia.gov
Phil Townsend, Associate Professor
Department of Forest Ecology and Management
University of Wisconsin-Madison
1630 Linden Drive, Russell Labs
Madison, WI 53706
608.262.1669
ptownsend@wisc.edu
Jane Foster, Doctoral Candidate
Department of Forest Ecology and Management
University of Wisconsin-Madison
1630 Linden Drive, Russell Labs
Madison, WI 53706
608-265-6321
jrfoster@wisc.edu
Wendy Cass, Botanist
Shenandoah National Park
3655 US Hwy 211 E
Luray, VA 22835
540-999-3432
Wendy_Cass@nps.gov
Lesley Sneddon, Sr. Regional Ecologist
NatureServe
11 Avenue de Lafayette, 5th floor
Boston, MA 02111
617-542-1908 x245
lesley_sneddon@natureserve.org
v
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
ACKNOWLEDGEMENTS
This project was the collective effort of many people and could not have been
completed without the group of dedicated and committed botanists, geographers,
and environmental scientists that came together to assess the vegetation of
Shenandoah National Park (SNP). However, this project could not have even
started without the guidance and logistical support of Wendy Cass, Botanist at
Shenandoah National Park. Gordon Olson, Natural Resources Branch Chief,
SNP, provided enthusiastic support for the project. We also thank Tom Blount of
the National Park Service for getting the ball rolling before moving to his next
NPS post. John Karish, NPS Northeast Region Chief Scientist was instrumental
in securing funding for the research. Mike Story of the NPS Vegetation Mapping
Program gave additional guidance and feedback. Dan Hurlbert, GIS Coordinator
for Shenandoah National Park, provided much of the GIS data used in this
project. Alan Williams of Shenandoah National Park provided additional
database support.
We also thank Dean Walton, formerly of the Virginia Department of Conservation
and Recreation, Division of Natural Heritage (DCR-DNH) for leading the field
portion of this project from 2000-2003. It was a privilege to work with such a
knowledgeable and dedicated naturalist. We also thank Karen Patterson of
DCR-DNH for data analysis and classification support and for leading the
accuracy assessment survey planning. Chris Ludwig of DCR-DNH helped to
initiate and guide the project through its completion. Allen Belden and Nancy Van
Alstine conducted the field accuracy assessment in 2004.
From the USGS, we thank Ann Rafter for GIS, image processing and field
support. We also thank Nissa Thomsen for field and mapping support and Dave
Morton, now with the Virginia Department of Game and Inland Fisheries, for
helping to initiate the project and procure imagery.
Lesley Sneddon of NatureServe led the crosswalk of vegetation communities to
their USNVC equivalents and helped design the key to identifying community
types at SNP. Jim Drake of NatureServe provided guidance on project planning.
We thank Wendy Cass, John Karish, Mike Story, Jim Comiskey, Kristina
Callahan, and Dan Hurlbert for helpful reviews of this report. We have
endeavored to incorporate their suggestions into this revised report.
Finally, we wish to thank the staff of Shenandoah National Park for graciously
accommodating our needs for lodging, keys, and entrance to the park. It was an
honor to work in such a beautiful and special place.
vi
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
LIST OF ABBREVIATIONS AND ACRONYMS
AA = Accuracy Assessment
ASCII = American Standard Code for Information Interchange
AVIRIS = Advanced Visible and Infra-Red Imaging Sensor
CART = Classification and Regression Tree
CLDA = Canonical Linear Discriminant Analysis
CTI = Compound Topographic Index
DEM = Digital Elevation Model
DOQQ = Digital Ortho-Photograph Quarter Quadrangle
ELU = Ecological Land Unit
GCP = Ground Control Point
GIS = Geographic Information System
GPS = Global Positioning System
LSC = Leetown Science Center
Landsat TM = Landsat Thematic Mapper
NAD = North American Datum
NASA = National Aeronautic and Space Administration
NMDS = Non-Metric Multidimensional Scaling
NPS = National Park Service
RMI = Relative Moisture Index
RMSE = Root Mean Square Error
SAF = Society of American Foresters
SHEN = Shenandoah National Park
TCI = Topographic Convergence Index
TRMI = Topographic Relative Moisture Index
USGS = United States Geological Survey
USNVC = United States National Vegetation Classification System
UTM = Universal Transverse Mercator
VAGAP = Virginia Gap Analysis Project
VANHP = Virginia Natural Heritage Program
WGS = World Geodetic System
vii
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
EXECUTIVE SUMMARY
This report documents the results of a four year research project to assess and
map vegetation communities of Shenandoah National Park. The project was a
collaborative effort between Shenandoah National Park, the US Geological
Survey-Leetown Science Center, the Virginia Department of Conservation and
Recreation-Division of Natural Heritage, the University of Maryland Center for
Environmental Science-Appalachian Laboratory, and NatureServe. While set up
as a research project rather than strictly a mapping effort, the result of the project
is a new map of vegetation distribution in Shenandoah National Park based on
U.S. National Vegetation Classification System standards. Additional products
include a classification scheme of vegetation communities in the Park based on
field sampling of 311 vegetation plots, maps of landforms, ecological land units,
and environmental gradients, and an assessment of the relationship between
vegetation community distribution and environmental gradients. Additionally we
tested the capability of mapping vegetation communities using hyperspectral
remote sensing, environmental gradient maps, and statistical modeling.
We classified 34 vegetation communities at the association level of the National
Vegetation Classification System within Shenandoah National Park. Three
community types were newly classified and described. We mapped vegetation
communities to the association level using AVIRIS spring 2000 and summer
2001 imagery, and we filled in missing areas with Landsat TM imagery. We
validated the results using internal cross validation and through an accuracy
assessment conducted at 224 field plots in the summer of 2004, and at 68 field
plots in 2005. Results of accuracy assessment range from 88% overall accuracy
from internal validation to 67% overall accuracy from field validation. As a
whole, statistical analysis indicated that bedrock parent material, soil fertility,
elevation, and topographic position are the most important, interrelated
environmental factors influencing major vegetation patterns in Shenandoah
National Park. In addition, vegetation communities occurred across a number of
environmental types as categorized into ecological land unit types.
The results of this project demonstrate innovative application of the latest
techniques in vegetation mapping using hyperspectral imaging and landscape
modeling. Mapping results are generally consistent with previous mapping
efforts (by area) and show that the USNVC can be reliably mapped with
techniques other than manual interpretation of aerial photography. In fact, by
exploiting the fine spectral specificity of hyperspectral imaging and the spatial
heterogeneity of environmental gradient models, much more information can
potentially be extracted on growing environments than can be visually
interpreted. The fact that association-level classes can be mapped using these
techniques is an important finding.
viii
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
1. INTRODUCTION
Accurate and up-to-date vegetation maps are fundamental to management of
national parks. Activities as diverse as park planning, fire management, wildlife
research, and visitor interpretation all require current maps of vegetation
distribution. In recognition of this need, the US Geological Survey (USGS) and
the National Park Service (NPS) jointly initiated a program for mapping
vegetation in National Parks to the United States National Vegetation
Classification system (USNVC) standard. Procedures and protocols were
developed for field sampling, photo interpretation, and accuracy assessment and
a funding program was established to initiate mapping at NPS units. As of 2006,
109 of 270 NPS units nationwide have vegetation mapping projects underway
(Brown 2006). However, since the program’s inception in 1994, only 39 parks
have completed vegetation mapping projects (Brown 2006). Clearly these
projects require a substantial investment of time and resources for field sampling,
vegetation classification, and image interpretation. Thus, research that can that
improve the efficiency, reliability, and information content extracted from
vegetation mapping protocols is needed.
Planning began in 1999 for a research project funded by the USGS Natural
Resources Preservation Program (NRPP) to assess vegetation community
distribution in relation to environmental gradients in Shenandoah National Park
(SHEN). While not officially a part of the USGS-NPS vegetation mapping
program, this project was conducted to provide SHEN with updated maps of
vegetation distribution using the USNVC standard, and at the same time to
investigate new methods of mapping vegetation communities and their growing
environments within the park.
This project was initiated in 2000 as a partnership between Shenandoah National
Park, the USGS-Leetown Science Center (USGS-LSC), Virginia Department of
Conservation and Recreation – Division of Natural Heritage (VANHP), and
NatureServe. In 2002, researchers at the University of Maryland, Center for
Environmental Science-Appalachian Laboratory (now of the University of
Wisconsin, Madison) were brought into the project to provide hyperspectral
remote sensing expertise. Field sampling was conducted in 2001, 2002, and
2003. Image analysis and environmental gradient modeling were conducted in
2001-2004, and draft vegetation association maps were completed in 2004. An
initial accuracy assessment was conducted in 2004, and an additional accuracy
assessment was completed in 2005. Revised map products were produced in
2005.
Environmental gradient modeling, sampling design, and remote sensing activities
were led by the USGS-LSC (Young); field sampling, vegetation community data
analysis, community classification, and accuracy assessment field work were led
by the VANHP (Fleming, Walton, and Patterson); crosswalk of the vegetation
classification into USNVC classes was led by NatureServe (Sneddon);
1
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
hyperspectral image analysis and vegetation community mapping were led by
Univ. of Wisconsin-Madison (Townsend and Foster).
1. 1. Background
One of the goals of the U.S. Geological Survey-Leetown Science Center (USGS-
LSC) is to provide clients within the U.S. Department of the Interior with biological
research results that will assist in managing the Nation’s public lands. A critical
information need within many national parks is accurate and up-to-date
information on vegetation composition, distribution, and change. Shenandoah
National Park (SHEN) in particular has pressing management issues that rely on
an accurate vegetation map including visitor safety, fire management, forest
insect pest management, and threatened and endangered species preservation.
The park’s landscape is the result of prior land use history and 70 years of
protection as a national park. Historically, park management has promoted
forest protection with an emphasis on fire suppression and minimal vegetation
manipulation. However, SHEN forests have undergone dramatic changes in
forest composition in the last decade as a result of gypsy moth defoliation,
hemlock woolly adelgid infestation, southern pine beetle infestation, ice storms,
large fires, and floods.
Shortly after the park’s establishment in 1935, vegetation communities were
mapped by Berg and Moore (1941). The park was mapped on a topographic
base in 19 sections at a scale of 1 inch to 1 mile (1:63,360). Twelve forest cover
types were mapped using a Society of American Foresters (SAF) classification
scheme. In addition to canopy cover type, age classes and acreage burned were
also recorded for cover types. No accuracy assessment was conducted for the
mapping effort and only forest cover types were mapped (no ground or shrub
cover estimates were provided). However, accuracy of this mapping effort is
considered good (Teetor 1988) and this map should serve as an excellent
reference for examining successional changes and disturbance patterns that
have occurred in the park over the past 70 years.
Subsequent to the Berg and Moore (1941) map, the park’s vegetation map of
record was a map developed from 1985-1988 (Teetor 1988) using low altitude
aerial photography imaged as 35 mm color infrared slides by the Virginina
Department of Game and Inland Fisheries (circa 1983-1984). The classification
was focused on forest canopy species and was based on Society of American
Foresters (SAF) cover classes. This effort provided an adequate map for
characterizing broad forest cover types and has been used extensively for
resource management, monitoring, and research projects. An extensive
accuracy assessment was conducted with over 2000 ground plots, and overall
accuracy of the mapping effort was reported to be 70% (Teetor, 1987). However,
this map was based solely on dominant overstory vegetation, and the park is
divided into only 7 forest cover types. Forest cover class boundaries were
interpreted by tracing over aerial slide images projected on a wall. This mapping
2
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
method undoubtedly introduced positional errors into the final map as systematic
errors inherent in aerial photography were not controlled. In addition, conversion
of this map between new implementations of GIS software packages over the
years resulted in introduced errors including shifts in vegetation types and open,
unclassified polygons along stand boundaries. Park resource inventories and
monitoring efforts now depend heavily on the use of digital maps for planning and
assessing ecological condition. Discrepancies in the accuracy of this vegetation
map and the massive changes that have occurred in the forests of the park
during the last 25 years justified the need for a new assessment of vegetation
composition and distribution.
In addition to the Teetor (1988) map, several other efforts have attempted to map
vegetation in SHEN. Cibula (1981) used Landsat Multi-Spectral Scanner (MSS)
imagery to map forest types. However, Teetor (1988) reported the map to be
unreliable, perhaps due to the coarse resolution of the early Landsat sensors (80
meter pixel size). The Southern Appalachian Man and the Biosphere (SAMAB)
program developed a regional land cover map through a private contractor
(Pacific Meridian) using Landsat TM imagery (circa 1990-1994). This map was
based on SAF cover types, and omitted understory vegetation and ecological
parameters. Accuracy of this map is unknown. Due to these limitations,
researchers and managers never adopted the SAMAB map. The Virginia Gap
Analysis Project (VAGAP) derived vegetation types for the entire Commonwealth
of Virginia from Landsat TM imagery (c. 1993). This effort attempted to go
beyond dominant overstory vegetation to community types based on the newly
implemented United States National Vegetation Classification System (USNVC).
Despite the improvement in the classification system, VAGAP could still not
provide the range of vegetation types, accuracy, and scale of treatment needed
by researchers and managers.
The USNVC is a hierarchical classification system, defining communities by
physiognomic structure at broad levels and then floristically at finer levels
(Grossman et al. 1998, Anderson et al. 1998, NatureServe 2002). Unlike the
Society of American Foresters cover classes that focus only on dominant tree
cover, the USNVC defines plant communities at the lowest level of the hierarchy
(the association level) on the basis of characteristic herbaceous and shrub
species, in addition to forest canopy species. The USNVC system also has the
potential to define characteristic mixed forest associations that may better reflect
natural conditions rather than attempting to lump forest types under a single
dominant tree species type as does the SAF system. The USNVC was recently
adopted as a Federal standard guiding vegetation mapping at U.S. government
agencies, state agencies, and non-governmental organizations.
Vegetation mapping in the deciduous forests of the eastern U.S. can be a difficult
task. Classification of spectral patterns from aerial or satellite imagery into
species or community-based categories can be challenging due to the nature of
the mixed forest communities. Since aerial and satellite imagery provide an
3
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
image from the top of the canopy downward, it is often difficult if not impossible to
directly image ground cover and shrub species. This difficulty often necessitates
mapping only to broadly defined groups when using remotely sensed imagery
alone, and past efforts (as evidenced above) have been somewhat unsatisfactory
from an ecological perspective. However, vegetation does respond predictably to
ecological gradients in the steep terrain of the park. Integrated ecological
modeling and predictive mapping approaches have shown promise for mapping
plant communities by exploiting the predictable relationship between vegetation
distribution and environmental gradients (Bridge and Johnson 2000, Franklin
1995, Swanson et al. 1988). For example, eastern hemlock (Tsuga canadensis)
is known to track closely with gradients in soil moisture and available light,
occurring in regular pattern on moist, cool, north-facing slopes and in moist,
shaded ravines. Conversely, pitch pine (Pinus rigida) commonly occurs on drier,
well-drained soils with a more southerly exposure. While this is not an entirely
deterministic relationship (past land disturbances, successional and gap
dynamics, soil characteristics, and micro-climates also strongly influence
vegetation occurrence), knowledge of the tendency of vegetation to occur in
definable ecological associations may allow a predictive approach to mapping.
Researchers in the United States and elsewhere have had success using a
predictive approach to mapping forest composition by modeling ecological
associations between vegetation and environmental gradients (extensively
reviewed in Franklin 1995). Several authors have created models of
environmental gradients by deriving measures of soil moisture availability,
available light, and topographic shape from digital geospatial data (Gallant and
Wilson 1996, Iverson et al. 1997, Dubayah and Rich 1995). By assessing
components of landform using digital elevation models, it is possible to model the
spatial distribution of climatic and topographic variables that have strong
relationships to vegetation occurrence. These data are especially useful for
spatial extrapolation of vegetation distribution between observations collected at
field plots (Hong et al. 1998). Combining field collected vegetation plot data,
satellite image derived measures of plant characteristics (e.g. tasseled cap and
other vegetation indices), and statistical classification, regression, and ordination
techniques has resulted in useful predictive models of vegetation distribution
(Franklin et al. 2000, Davis and Goetz 1990). However, predictive models that
are based solely on direct gradients are useful only for prediction of potential
natural vegetation patterns. Since current vegetation pattern is highly correlated
to past disturbance (Glenn et al., 1999), information on disturbance regimes
(derived from satellite or aerial imagery or historic land use maps) should be
incorporated in predictive models to get an accurate idea of current vegetation
distribution.
Data-rich hyperspectral imagery represents one of the latest advances in sensor
technology applicable to landscape mapping of vegetation (Treitz and Howarth
1999). Measurement of electromagnetic radiation from hundreds of spectral
bands introduces hundreds of new predictor variables that may aid in
4
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
discrimination of vegetation communities. Several studies have documented
attempts to harness the potential discriminating power of hyperspectral data for
vegetation mapping in recent years (Martin, Newman et al. 1998; Treitz and
Howarth 1999; Cochrane 2000; Gong, Pu et al. 2001; Foster and Townsend
2004; Thenkabail, Enclona et al. 2004). These studies are at the forefront of a
broad and diverse effort to exploit hyperspectral data for the analysis of
vegetative life forms and communities. As these goals are pursued, many hopes
for the application of hyperspectral image data to vegetation mapping remain
unrealized, and limitations continue to exist in the current availability and
coverage of hyperspectral data. One area that deserves further investigation is
the fusion of statistical analysis, environmental gradient modeling, and
assessment of spectral reflectance derived from hyperspectral and multispectral
sensors for mapping vegetation communities to USNVC standards.
Few researchers have investigated whether the newly adopted USNVC can be
used successfully with this approach, and no effort has been attempted to fully
incorporate this type of information to map the heavily disturbed vegetation of
Shenandoah National Park. If reliable relationships between ecological
gradients, spectral reflectance, and vegetation patterns can be established at
SHEN, then knowledge of current vegetation distribution, potential successional
dynamics, and impacts of future disturbances will be greatly enhanced.
1. 2. Scope of Work
The overall objective of this project was to assess the distribution of vegetation
communities in SHEN in relation to ecological units defined by terrain and
landscape structure. Supporting and concurrent objectives included 1)
classification of vegetation communities into a USNVC hierarchy using data
collected at field plots, 2) research and development of ecological gradient
models based on terrain analysis, 3) investigations of newly available remote
sensing technology for mapping vegetation to the USNVC, 4) construction of a
statistical model that predicts the distribution of USNVC vegetation classes from
field plots, terrain-based ecological gradient models, and vegetation spectral
responses mapped from satellite imagery, 5) delineation of riparian and wetland
areas of the park, and 6) statistically valid accuracy assessment of vegetation
classifications and ecological models.
The first objective of this project was to evaluate and assess information on
distribution and composition of vegetation from current plot databases. Several
groups had collected information on vegetation occurrence and distribution within
SHEN at the initiation of this project. The VANHP located 103 vegetation plots
within the park prior to 2000 and had collected a variety of detailed floristic,
structural, and environmental data at each plot. Of these plots, 93 were sampled
in the period 1999-2001, just prior to initiation of this project. Shenandoah
National Park’s research and monitoring effort placed 400 plots throughout the
park, stratified by vegetation types predicted from the 1987 vegetation map.
5
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Additionally, USGS-LSC placed over 100 plots in eastern hemlock stands to
assess forest composition and tree health in relation to defoliation by hemlock
woolly adelgid (Adelges tsugae). These data were incorporated in the
classification or evaluation process for this project where sufficient data existed
to evaluate vegetation community composition and environmental attributes.
Secondly, geographic information system (GIS) based terrain modeling was
conducted to determine ecological gradients within the park using landscape and
topographic data. We analyzed landscape measures relevant to the occurrence
and distribution of vegetation communities (e.g. direct and indirect gradients)
from digital elevation models and other GIS data layers such as solar
illumination, predicted soil moisture, terrain shape, slope position, and soil parent
material (e.g. geology). We used classification and ordination techniques to
determine the main ecological gradients within the park that influence vegetation
occurrence and distribution. Ecological gradients derived from a digital elevation
model (DEM) were used as a framework to assess existing vegetation
distribution from field plots, to guide remote sensing interpretations, and to
construct predictive models of plant community distributions.
The third objective of this project was to assess the applicability of newly
available satellite and aerial imagery for mapping vegetation using the USNVC.
Past efforts using moderate spatial and spectral resolution instruments such as
Landsat Thematic Mapper (TM) have been successful mapping to broad levels in
a classification hierarchy but have had difficulty achieving fine specificity. Newly
available imaging technologies offer much higher spectral resolutions (e.g.
hyperspectral), or use different approaches for vegetation classification. This
project explored the utility of emerging remote sensing methods and instruments
for mapping vegetation communities to USNVC through hyperspectral image
analysis, statistical classification techniques, and environmental gradient
modeling.
The fourth objective was to use statistical modeling techniques to extrapolate
vegetation communities observed at field plots to similar environments within the
park on the basis of ecological gradient models, spectral responses of current
vegetation from satellite imagery, and spatial information on past disturbances.
We tested and employed predictive models using Classification and Regression
Tree (CART) techniques, canonical correlation, and canonical linear discriminant
analysis.
The fifth objective was to put special emphasis on accurately delineating riparian
zones through a combination of moisture regime modeling using a DEM and
remote sensing assessments. Once predicted riparian and wetland zones were
delineated, special emphasis was given to these areas for field assessment and
survey.
6
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
The last objective was to conduct a statistically valid accuracy assessment of
predicted vegetation communities using standard procedures adopted by the
national USGS/NPS National Vegetation Mapping program. Two separate field
surveys were conducted to assess accuracy of vegetation composition predicted
from classifications and gradient models.
1.3. Study Area Description
As of 2004, Shenandoah National Park (SHEN) encompasses 195,821 acres
(79,246 ha) of mostly forested uplands in the Blue Ridge Mountains of
northwestern Virginia. Of this total, 82,661 acres (33,452 ha) or 42.2% of the
park is in Wilderness designation. The park occurs primarily in the “Northern
Igneous Ridges” and “Northern Sedimentary and Metasedimentary Ridges”
ecoregions of Omernick (Omernick, 1995), although small portions of the park
occur in the “Northern Limestone/Dolomite Valleys” and “Piedmont Uplands”
ecoregions. Elevations range from a low of 530 ft (161.5 m) near Front Royal to
4,060 ft (1237.5 m) at Hawksbill Mountain. Generally, 14% of the park is in
elevations below 1500 ft (457.2 m), 74% in elevations between 1500 ft (457.2 m)
and 3000 ft (914.4 m), and 10% is above 3000 ft (> 914.4 m) (Figure 1.1).
The geology of the park consists of three major lithologies: Cambrian-aged
siliciclastic rocks of the Chilhowee group (Antietam, Harpers, and Weverton
formations); late pre-Cambrian-aged metabasalts (a.k.a. “greenstone”) of the
Catoctin formation; and Middle Proterozoic granitic and gneissic rocks, including
the Old Rag granite and rocks formerly called the Pedlar formation (alkali-
feldspar leucogranite, charnockite, charnockite gneiss, and layered pyroxene
granulite) (Figure 1.2). Small areas of the western flank of the park are
underlain by Cambrian and Ordivician-aged carbonate rock and calcerous shales
of the Tomstown and Waynesboro Formations, and late pre-Cambrian aged
siliciclastic rocks of the Swift Run formation occur between the metabasalts of
the Catoctin formation and the granitic formations.
Geology and topography are primary drivers of vegetation distribution within the
park along with climate, fires, and prior land use history (Conner 1988). Geology
and topography act synergistically with climate to influence soil formation,
nutrient availability, moisture, solar insolation, and temperature. Siliciclastic
parent material of the Chilhowee group produces nutrient poor, well drained, and
acidic soils. Metabasalt parent material of the Catoctin formation produces
relatively nutrient rich, mesic soils. Granitic and gneissic parent material of the
Old Rag granite and former Pedlar formation complex produces intermediate
soils of higher acidity than the Catoctin metabasalt. In addition, alluvium deposits
occur along stream channels and provide soils of different nutrient and moisture
capacity than would be expected from bedrock geology alone. Previous land use
history and fire are additional factors that influence vegetation distribution, but
are themselves moderated to some extent by geology and topography.
7
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 1.1 Elevation classes of Shenandoah National Park.
8
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 1.2 Generalized surficial geology of Shenandoah National Park,
(after Morgan et al. 2004).
9
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Berg and Moore (1941) recognized the importance of geology in structuring
vegetation communities in the park in their extensive mapping effort, but it is
unclear how or if this information was used for mapping. While acknowledging
that the park was primarily in second growth, they state:
“Underlying rock formations with residual soil overburden have had a
definite influence on the existing forest cover. This influence is evidenced
by the Red Oak type associations, which are found generally throughout
the moderately moist to moist soils overlying Catoctin greenstone and,
less frequently, Hypersthene granodiorite formations; Chestnut Oak types,
which prevail on moderately dry to dry aspects underlain by granodiorite
and quartzite; and the Bear oak, Scarlet oak, and Pitch Pine types, which
are found on the dry to very dry quartzite, shale, and limestone
soils…Previously grazed and cultivated lands are confined in general to
the more fertile greenstone and granodiorite soils” (Berg and Moore 1941,
pg. 2)
Braun (1950) describes the vegetation of the Northern Blue Ridge (including
Shenandoah National Park) in relation to the physical landscape. She notes the
vegetation of the area as lacking “the luxuriance and variety which are distinctive
features of the Southern Appalachian section” due to less favorable climate,
lower altitudes, less varied topography, and a greater degree of human
disturbance (Braun 1950, pg. 221). Oak-chestnut is described as the prevalent
type (even though American chestnut (Castanea dentata) was long since
extirpated as a dominant canopy species by 1950), and human disturbance was
evident in the red cedar (Juniperus virginiana), black locust (Robinia pseudo-
acacia), and sassafras (Sassafras albidum) occupying clearings and old fields of
the lower slopes (Braun 1950). Braun describes the structure of the upland
forests in the park in relation to landscape structure:
“Forest variations along the upper slopes and crests are related to slope
exposure, steepness, and concavity or convexity of slope. Mesophytic red
oak-sugar maple-basswood communities or groups of hemlock in
northerly concavities alternate with red oak-chestnut communities whose
undergrowth contains Azalea, mountain laurel, and blueberries, or with
oak-chestnut communities with a continuous heath layer. If the crest is at
a low elevation (2500 feet or less) tuliptrees are present in the north slope
coves. On windswept slopes and knobs, chestnut oak is abundant, and
pines dominate locally…Only at the highest elevations (slopes of Hawks
Bill [sic] Mountain) are spruce and fir present.” (Braun 1950, pg. 223-224)
Teetor (1988) discussed the limiting effects of elevation, geology, and soil
moisture on structuring vegetation communities in the park and recognized the
strong influence of disturbance history, but did not specifically incorporate this
information into mapping. Instead, she compared the observed vegetation cover
10
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
types to the “expected topographic distribution” after the fact by assessing
vegetation in relation to classes of slope, aspect, and elevation. Topographic
influences are discussed by Teetor (1988) to place vegetation cover types in
context of the environmental limitations and disturbance history.
Both Berg and Moore (1941) and Teetor (1988) report chestnut oak (Quercus
prinus) as the dominant cover type occurring on dry, thin soils of the main ridge
and spur ridges overlying granitic and siliciclastic geology. However, there is
less agreement on both the type and extent of other forest types. Berg and
Moore (1941) list red oak (Quercus rubra) and scarlet oak (Quercus coccinea) as
the next most dominant cover types while Teetor (1988) lists yellow poplar
(Liriodendron tulipifera), red oak (Quercus rubra) / ash (Fraxinus americana) /
basswood (Tilia americana) and red oak (Quercus rubra) as the next most
dominant types. Methodological and classification differences may explain some
of the variation between the mapping efforts, but most of the difference must
certainly be attributed to 45 years of succession (eg. less open area and more
black locust forest cover in the Teetor (1988) map). The difficulty in mapping
mixed deciduous forests of the park may also be a factor as evidenced by the
accuracy statement of the Teetor (1988) effort: 70% overall accuracy, and
between 63% and 74% by-class accuracy.
Knowing that vegetation in SHEN responds predictably to environmental
gradients in the absence of disturbance may allow for a predictive approach to
mapping by exploiting newly available digital elevation data and satellite imagery
2. METHODS
2.1 Environmental Gradient Modeling
We mapped significant ecological gradients for SHEN as a precursor to
vegetation sampling and for use as inputs to predictive models. The overall goal
in this effort was to quantify environmental gradients that are important for
structuring vegetation communities. Methods used to quantify environmental
gradients closely followed those used by other researchers, and made use of
geographic information systems (GIS) and digital elevation models (DEM).
A number of researchers have examined vegetation data in relation to
environmental gradients derived from digital elevation data (see Table 2.1 for a
list of selected recent publications that have examined vegetation in relation to
environmental gradients). Important gradients that recur in these studies are:
slope direction (e.g. aspect), slope position, slope shape, moisture, light, and
(less commonly) rock type, and elevation. We derived environmental gradients
following the examples set in these studies to capture gradients important for
vegetation growth in the mountains of western Virginia. These are gradients of
soil moisture, light, slope orientation (aspect), slope shape, elevation, exposure,
and rock type.
11
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
We generated information on environmental gradients in two forms, as discrete
class maps for use in constructing a map of ecological land units, and as
continuous variables for use in predictive modeling of vegetation communities.
Both representations were derived using GIS operations.
2.1.1 Ecological Land Units
We followed methods in Anderson and Merrill (1998) for combining gradient
layers into an “ecological land units” map (also referred to as a “biophysical units”
map). Our goal was to use this information to create sampling strata that capture
the range of environments observed. The Anderson and Merrill (1998) method
(implemented as a set of GIS scripts by F. Biasi (2001)) builds an ecological units
map by classifying and combining individual environmental gradient maps in a
GIS. Maps of aspect, moisture, slope, and slope shape are reclassified and
assembled to produce maps of landform units. These landform units are then
combined with reclassified elevation and geologic maps to produce a final
ecological land units or “ELU” map. We used these methods as a guide to
building an ecological land units map for Shenandoah National Park, adapting
the procedures for local conditions. Individual steps in the process and maps
resulting from intermediate and final stages are described.
12
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Table 2.1. Selected publications using environmental gradient models for vegetation community
analysis.
Study Variables considered:
Newell and Peet, 1998
From: “Vegetation of Linville Gorge Wilderness,
North Carolina”, C. L. Newell and R. K. Peet,
Castanea 63(3): 275-322, September 1998.
A species composition and vegetation-
environment relationships study.
- Beer’s transformed aspect
- Distance to nearest stream
- Distance to nearest ridge
- Terrain shape index (after McNab)
- Topographic complexity
- Potential solar radiation (from
Solarflux)
- Topographic Moisture Index (after
Parker)
Anderson and Merrill, 1998
From: “ Connecticut River Watershed: Natural
Communities and Neotropical Migrant Birds”,
M.G. Anderson and M. D. Merrill, Final Report,
The Nature Conservancy, Eastern Regional
Office, Boston, MA, October 15, 1998.
An ecological communities assessment project.
Ecological Land Units were derived to assist in
a regional planning and assessment project.
- Slope (degrees)
- Moisture Index (after Moore, I.D.)
- Landscape position
- Lithology
- Elevation
Franklin, et. al. 2000
From: “Terrain variables used for predictive
mapping of vegetation communities in
Southern California”, J. Franklin, P.
McCullough, and C. Gray, in Terrain Analysis:
Principles and Applications, J. P. Wilson and J.
C. Gallant, eds., John Wiley and Sons, New
York, 2000, pp. 331-353.
A predictive vegetation modeling study.
- Slope
- Aspect
- Potential Solar Radiation (from
Solarflux)
- Upslope catchment area
- Topographic wetness
- Surface curvature
- Distance to stream
- Distance to Ridge
The primary data source for environmental gradient analysis was a 10-meter
USGS digital elevation model (DEM) for SHEN. This is a compilation of USGS
10-meter resolution 1:24,000 elevation models available for all topographic
quads covering SHEN except for the northern-most quad (Front Royal). These
data generally improve upon the 30-meter resolution digital elevation data in both
surface representation and accuracy. However, these data have artifacts from
the source data (contours) that may affect resulting models such as artificial
terracing and pits. We resampled the 10-meter DEMs covering the park as well
as the 30-meter Front Royal quadrangle to a new 15-meter merged DEM in order
to smooth out some of these inconsistencies, and to incorporate data for the
13
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Front Royal quadrangle. This provided a single consistent elevation model
covering the park. All elevation-based topographic gradient maps were
subsequently derived using this layer.
Elevations from the DEM (originally in meters) were reclassified to correspond to
3 broad elevation ranges: 0-1500 ft (0-457.2 m), 1500–3000 ft (457.2-914.4 m),
and greater than 3000 ft (> 914.4 m). These classes were determined to have
the greatest correlation with vegetation distribution in the mountains of western
Virginia based on previous vegetation assessments (G. Fleming, VANHP, pers.
comm.).
Reclassifying the DEM into elevations was accomplished using a simple recode
operation. Numeric codes were assigned to correspond to the above classes as
follows:
0 –1500 ft (0-457.2 m) = 1000
1500 – 3000 ft (457.2-914.4 m) = 2000
> 3000 ft (> 914.4 m) = 3000
We derived an index of topographic moisture from the DEM using methods
proposed by Ian Moore (1990) and adopted by Anderson and Merrill (1998) as
well as others. The basic idea is to examine the amount of water entering a point
on a map (e.g. a pixel) and compare it to the amount of water that would leave
the cell based on topography. The “relative moisture index” is computed as the
log of the ratio between the flow accumulation at each cell and the slope of the
cell. The “flow accumulation” function in ArcInfo (ESRI Inc., Redlands, CA) is
used to compute a relative amount of water entering each cell from its upstream
neighbors (values are number of upstream cells flowing into each cell). Slope is
computed for each cell in the DEM as percent slope. The formula for
computation of the moisture index (as given by Anderson and Merrill (1998)) is
then:
Relative moisture index = ln((flowaccumulation + 1) / (slope + 1))
In order to generalize the map slightly to remove spurious features, we filtered
the resulting moisture index map using a mean filter that replaces cell values at
each pixel with the mean value occurring in a 3x3 pixel scanning window. Other
moisture index maps, such as the Topographic Relative Moisture Index proposed
by Parker (1982) could be substituted here if desired.
We derived a landform index using a routine provided by Zimmerman (2000) that
computes a terrain shape or landform index in a slightly different manner than
that proposed by Anderson and Merrill (1998). Both techniques compute terrain
shape in a manner similar to that proposed by McNab (1989) whereby elevations
at each pixel are compared to the mean of elevations of neighboring pixels. If
the elevation of the focal cell is greater than that of the mean of the neighboring
14
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
cells, then this is coded as a local high or a convex shaped terrain, while
elevations lower than the mean of their neighbors are coded as local lows,
reflecting concave shapes. Differing scales of topographic shape can be
quantified in this manner by varying the size of the scanning window. Unlike the
McNab (1989) method, both the Zimmerman (2000) and the Anderson and
Merrill (1998) methods compute a weighted mean of terrain shape from
assessments at different spatial scales. We used the Zimmerman (2000) routine
since it maintains the terrain shape value for the most influential scale rather than
averaging over all scales. The basic calculation implemented in GIS is as
follows:
Terrain Shape Index = dem - focalmean(dem, circle, radius (X))
Terrain shape at each pixel on the DEM is calculated as the elevation value at
each cell on the DEM minus the mean elevation of pixels in a surrounding
circular window of size X, with the scanning window varying from 15 to 150
meters (radius).
We derived aspect (e.g. slope direction) in GIS using a standard routine that
classifies slope direction into degrees using compass directions (0-360 degrees).
We transformed aspect using Beer’s cosine transformation such that slopes
facing 50º (described as optimal for southern Appalachian vegetation response
by Newell and Peet 1998) are given a value of 2. Other slopes that are within
90° of this optimal NE direction are given a value of 1, and SW facing slopes are
given a value of 0.
The formula for computation with ArcInfo is given as:
Beers Aspect = cosine(50° – aspect) + 1
Aspects were further simplified for computation of the ELU units as either N-NE
or S-SW. These cutoffs correspond to the perpendiculars to Beer’s aspect (eg.
320° and 140°).
Slope was calculated in degrees using standard functions in GIS. Slope is
calculated as the maximum angular rate of change (in elevation) of a plane fit to
a 3x3 window surrounding each pixel on the DEM (ESRI, 1994).
We also incorporated USGS 1:24,000 digital line graph maps to represent
streams, ponds, and wetlands. We converted the hydrologic data into a grid
representation that matches the 15 m grid cells of the DEM layer.
We derived landform classes by re-classifying and combining the above maps,
following the techniques of Anderson and Merrill (1998) as defined by Biasi
(2001). Each map is reclassified into discrete classes and combined in a specific
order to derive landform classes. First, slope and landform maps are reclassified
15
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
and combined. Very steep sloped areas (greater than 35°) are classified as cliffs.
Areas of intermediate slope (24° to 35°) are classified as steep slopes. Areas of
moderate slope (6° to 24°) are classified as side slopes. Low sloping areas (< 6°)
are classified as flats. Terrain shape is used to determine if slopes are concave
or convex. Concave slopes are classified as coves or slope bottoms, while
convex slopes are classified as upper slopes or side slopes. Flat slopes are
classified as either ridge top or bottom, and coded as either moist or dry by
overlay with locations of wetlands. Next, aspect maps are reclassified and
incorporated into the landform map to determine slopes facing N-NE or S-SW.
Finally, streams and lakes are incorporated for the final landform map. The
landform map was filtered using a majority filter in a 5x5 pixel window to remove
small, spurious landforms.
We initially used a map of geologic formations mapped by Gathwright (1976).
However, subsequent to initiation of this project Morgan, et al. (2004) mapped
bedrock and surficial geology in SHEN by compiling existing bedrock geology
maps (including Gathright (1976) and others) and by compiling maps of surficial
geology from existing sources and conducting new mapping at 1:24,000 scale.
The updated bedrock geology map uses the most current nomenclature for
geologic formations in the area (e.g. it does not split the granitic types into Old
Rag and Pedlar formations), and combines some units that inter-grade (Catoctin
and Swift Run formations). The updated surficial geology map includes areas of
alluvial deposits, debris flows, and depositional (or “strath”) terraces. Since this
updated map includes additional geologic information not available in the
Gathright (1976) map used in the initial analysis, and since the terminology used
in coding this map conforms more closely to the latest geologic understanding of
SHEN, we used these maps in revised modeling of vegetation composition.
We merged the surficial and bedrock geology maps into a single map and
recoded all units to five categories to match the characterization of soil parent
material used in the vegetation plot data ordination and classification (Table 2.2,
Figure 1.2). Polygons were converted to a grid representation to match the
topographic variables. Surficial polygons coded as debris flow or strath terraces
were coded in the merged map with the predominant bedrock geology upslope
from the surficial unit. Polygons coded as alluvium were not tagged with parent
material due to the difficulty in assigning probable parent material. All other
areas not covered by a surficial geology polygon as mapped by Morgan, et al.
(2004) were coded with bedrock geology type.
Table 2.2. Rock formations and rock type groupings of Shenandoah National Park
Formation Rock Type Rock Type Code
Fluvial deposits Alluvium 100
Chilhowee Group Acidic sedimentary (siliciclastic) 200
Quartzo-feldspathic Granitic (meta-igneous) 300
Catoctin and Swift Run Mafic (meta-basalt, intergraded
conglomerate, slate)
400
Waynesboro, Tomstown Calcerous/shale 500
16
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
In the final step, maps of elevation, landforms, and geology are combined to
produce the final ecological land units map. Since the elevation map is coded
into the thousands, the geology map is coded into the hundreds, and the
landform map is coded into the tens, a simple addition of the three maps results
in the final landform class combinations (Table 2.3.)
TABLE 2.3. Class codes for Ecological Land Units map (ELU). Final ELU code is derived by
adding codes for elevation, geology, and landform. For example, low elevation-basaltic-cliffs
would be coded as 1000+100+10, or 1110.
Elevation Geology Landform
low (< 1500 ft) 1000 alluvium 100 cliff 10
mid (1500 <= 3000 ft) 2000 acidic sed. 200 steep slope 11
high ( > 3000 ft) 3000 granitic 300 slope crest 12
basaltic 400 upper slope 13
calcerous 500 flat summit/ridge 14
sideslope N/NE 20
cove/ravine N/NE 21
sideslope S/SW 22
cove/ravine S/SW 23
dry flat 30
slope bottom 33
stream 40
lake 42
spring/seep 43
2.1.2 Environmental Gradient Models for Predictive Modeling
In addition to the discrete maps of ecological land units, we also created
continuous maps of environmental gradients to use in predictive modeling. Since
vegetation composition grades continuously across the landscape and responds
to subtle changes in light availability, nutrients, and soil moisture, maps that
capture the continuous gradation of environmental influences are more
appropriate for modeling than maps representing gradients as discrete classes.
Guisan and Zimmerman (2000) divide influences on plant growth and distribution
into “resource”, “direct”, and “indirect” gradients. Resource gradients are taken
up directly by plants (e.g., water, photosynthetic active radiation, and soil
nutrients). Direct gradients influence the availability of resource gradients, such
as soil type, solar insolation, water availability, and ambient temperature. Indirect
gradients represent relatively large-scale influences, such as geology,
topography, climate, and latitude, that govern the formation of direct gradients.
Typically, indirect gradients (and, less typically, direct gradients) can be created
in GIS format for modeling environmental influences over broad areas. Resource
gradients are rarely available in spatial format useful for modeling influences on
plant communities.
We derived models of direct and indirect environmental gradients using available
methods in three general areas: general topographic measures including
17
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
measures of slope shape, measures of soil moisture, and measures of solar
illumination (Table 2.4.) All gradients were derived from the same digital
elevation model described in section 2.1.1. Some gradient models were
calculated from intermediate steps of the ELU modeling (e.g. Beers aspect
transform, slope), while others were calculated using additional specialized
scripts.
Table 2.4. Environmental gradients derived for Shenandoah National Park from a 15-meter
resolution digital elevation model. A.) Topographic measures B.) Measures of solar illumination
C.) Measures of soil moisture
2.4.A. Topographic Measures
Derived Gradient Derivation Reference
Elevation (in meters) = height
above m.s.l.
DEM USGS. 1993. US Geodata, Digital
Elevation Models, Data Users Guide.
Technical Instructions: Data Users Guide
5. U.S. Geological Survey, National
Mapping Program. Reston, Virginia.
Slope (degrees) = maximum
rate of change in z value
Elevation (DEM) ESRI, Inc. 1994. Cell-based modeling with
GRID. Environmental Systems Research
Institute, Redlands, CA. 481 pp.
(Measures of Slope Shape)
Plan curvature = across slope
(e.g. horizontal) curvature
Elevation (DEM) ESRI, Inc. 1994. Cell-based modeling with
GRID. Environmental Systems Research
Institute, Redlands, CA. 481 pp.
Profile curvature = down slope
(e.g. vertical) curvature
Elevation (DEM) ESRI, Inc. 1994. Cell-based modeling with
GRID. Environmental Systems Research
Institute, Redlands, CA. 481 pp.
Terrain shape index (TSI) =
local convexity or concavity
Elevation (DEM) McNab, H.W. 1989. Terrain shape index:
Quantifying effect of minor landforms on
tree height. Forest Science 35(1): 91-104.
Relative Slope Position (RSP)
= position on slope relative to
stream and ridgeline
Elevation (DEM) S. P. Wilds. 1996. Gradient analysis of
the distribution of flowering dogwood
(Cornus florida L.) and dogwood
anthracnose (Discula destructiva Redlin.)
in western Great Smoky Mountains
National Park. M.S. Thesis, Univ. of North
Carolina,Chapel Hill. 151pp.
2.4.B. Measures of Solar Illumination
Derived Gradient Derivation Reference
Beer’s Aspect = slope
direction (aspect) converted to
a continuous scaled variable,
set to maximum for NE slopes
(45° = coolest slope).
Aspect in degrees Beers, T.W., Dress, P.E., and Wensel, L.C.
1966. Aspect transformation in site
productivity research. J. For. 64:691.
Average Solar Illumination =
relative amount of sunlight
striking the surface throughout
the year.
Elevation (DEM)
(sun position at
solstices and
equinoxes)
ESRI, Inc. 1994. Cell-based modeling with
GRID. Environmental Systems Research
Institute, Redlands, CA. 481 pp.
18
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
2.4.C. Measures of Soil Moisture
Derived Gradient Derivation Reference
Topographic Relative Moisture
Index (TRMI) = a summed
scalar index of relative
moisture availability based on
aspect, slope, slope shape,
and relative slope position
Aspect
Slope
Plan curvature
Profile curvature
RSP
Parker, A.J. 1982. The topographic
relative moisture index: an approach to
soil-moisture assessment in mountain
terrain. Phys. Geogr. 3(2):160-168.
Relative moisture index (RMI)
= relative amount of water
flowing into a pixel (flow
accumulation) in relation to
amount flowing out based on
slope (a.k.a. “wetness index”)
Elevation (DEM)
Moore, I.D, Gessler, P.E., Nielsen, G.A.,
and Peterson, G.A. 1993, Soil attribute
prediction using terrain analysis. Soil
Science Society of America Journal
57:443-52.
Topographic convergence
index (TCI) = similar to
wetness index but calculates
the upslope contributing area
in relation to slope expressed
as percent rise
Elevation (DEM)
Wolock, D.M., and G.J McCabe, Jr. 1995.
Comparison of single and multiple flow
direction algorithms for computing
Topographic parameters in TOPMODEL.
Water Resources Research 31:1315-1324.
Compound topographic index
(CTI) = a steady state wetness
index very similar to TCI
except the tangent of slope is
used rather than rise/run.
Elevation (DEM) Moore, I.D, Gessler, P.E., Nielsen, G.A.,
and Peterson, G.A. 1993, Soil attribute
prediction using terrain analysis. Soil
Science Society of America Journal
57:443-52.
2.2 Sample Site Selection
2.2.1 ELU-based Sample Site Selection
We located new sampling sites to build the vegetation classification scheme and
to serve as training data for mapping. Sample sites were stratified within
Ecological Land Unit types proportional to the area that each type represented in
the park. We determined that we could sample vegetation at a maximum of 500
field sites during the life of the project through discussion between personnel
from USGS-LSC, VANHP, and NPS. We initially focused sampling only on the
basaltic, siliciclastic, and granitic rock types since these units form the majority of
the park by area and some sites on the carbonate bedrock of the western flank of
the park had already been sampled by VANHP prior to initiation of this project.
The proportion of area in each ELU was multiplied by 500 to get the number of
initial sample points to place in each ELU. Within each ELU, points were
randomly located using a specialized script written for Arc/Info GIS. Locations
were filtered such that they were at least 100 meters, but no more than 500
meters, from a road or trail to avoid areas likely to be influenced by human land
use and to improve access to sample sites, respectively. In addition, sample
points were restricted to areas at least 20 meters from an ELU boundary to avoid
sampling in edge habitats.
Selected sample site coordinates were exported along with a unique point
number and ELU type as an ASCII text file. Sample site coordinates were
19
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
loaded into (non-differential) Global Positioning System (GPS) units for field
navigation. Sample points were located in the field to within 5-10 meters (on
average) of their predetermined location. In some cases, pre-selected sample
sites were found in the field to be in heavily disturbed vegetation (e.g. open
canopy with dead or downed trees and thick shrub succession). In these cases,
a judgment was made in the field by VANHP botanists as to whether the
disturbance was severe enough to alter the vegetation composition. Plot
locations with subjectively determined severe disturbance were relocated in the
field to the closest intact vegetation within the same ELU.
2.2.2 Riparian and Wetland Sample Site Selection
Several ELU strata individually represented less than 0.1% of the park in area.
Due to the small size of these units and limited resources, generally no samples
were placed in these units. However, in order to meet the objective of assessing
and mapping wetland areas, we paid special attention to riparian and wetland
site selection in 2002. Since wetland and riparian areas represent only a small
proportion of area in the park as opposed to upland areas, we added additional
effort out of proportion to the numbers of samples expected from the ELU-based
sample allocation. We selected an initial sample for 2002 that included at least 3
randomly selected sample plots in each wetland and riparian ELU type found in
the park.
2.3 Field Survey Methods
2.3.1 Field Data Collection:
Plots were sampled using the relevé method (sensu Peet et al. 1998), following
standard VANHP procedures. As a rule, 400 m2 quadrats with 20 x 20 m
configurations were employed in forest and woodland vegetation, while 100 m2
quadrats with 10 x 10 m configurations were used in shrubland and herbaceous
vegetation. At some plots, however, rectangular configurations (e.g., 16 X 20 m,
10 x 40 m, or 5 x 20 m) were used to conform with narrow vegetation zones of
cliffs, ridge crests, ravines, and stream bottoms. In three cases, rocks, downfalls,
and other impediments made it impractical to sample anything larger than a 200
m2 plot. Vegetation sampling under this project was conducted during the
growing seasons of 2001, 2002, and 2003 and data were collected from 208
plots. Data from 103 additional plots sampled in the park by VANHP during the
period 1990-2000 were also utilized in the analysis. Of these plots, only 10 sites
were sampled prior to 1999, and therefore most sites were contemporaneous
with conditions encountered during 2001-2003 field sampling.
2.3.2 Vegetation Measurements:
To the extent possible, plots were placed in homogeneous stands of vegetation.
Every vascular plant taxon present was recorded and its cover, defined as the
20
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
percentage of the ground covered by the vertical projection of above-ground
biomass, was visually estimated over the full plot area. Cover was assigned
using a nine-point scale of cover classes (Table 2.5).
The overall cover of mosses, lichens, and liverworts was estimated, but the
individual covers of non-vascular taxa were not estimated. Vascular plants
thought to be characteristic of the sampled community, but located outside the
plot, were recorded parenthetically if visible from the boundary, and assigned a
cover class score of “p.” The total vegetative cover was also estimated using the
same nine-point cover-class scale used to estimate species covers.
Table 2.5. Cover class scores used in field sampling and data analysis (400 m2 plot).
Cover
Class:
Percent Cover Range: Area of Coverage: Cover Class
Midpoint (%):
(p) present outside plot - 0.05
1 < 0.1% < 20 cm20.05
2 0.1% to 1% 20 cm2 to 4 m20.55
3 1 to 2% 4 m2 to 8 m21.50
4 2 to 5% 8 m2 to 20 m23.50
5 5 to 10% 20 m2 to 40 m27.50
6 10 to 25% 40 m2 to 100 m217.50
7 25 to 50% 100 m2 to 200 m237.50
8 50 to 75% 200 m2 to 300 m262.50
9 75 to 100% 300 m2 to 400 m287.50
In addition to recording presence and cover for all species, stand structure was
quantified by measuring the size distribution and vertical stratification of woody
plants. Each woody stem (trees, shrubs, lianas) ≥ 2.5 cm dbh and < 40 cm dbh
was tallied within 5 cm diameter classes, using the measurement of the stem at
breast height (1.4 m). Diameter at breast height (dbh) classes used were 2.5-5,
>5-10, >10-15, >15-20, >20-25, >25-30, >30-35, and >35-40 cm. Stems ≥ 40 cm
dbh were individually measured to the nearest 1 cm. The maximum canopy
height was measured using a clinometer and the cover of each woody species
was estimated (if present) at each of six height strata:
herb layer, < 0.5 m
shrub layer, > 0.5 to 6 m
tree layer, > 6 to 10 m
tree layer, > 10 to 20 m
tree layer, > 20 to 35 m
tree layer, > 35 m
21
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
2.3.3 Environmental Measurements:
A standard set of environmental data was measured or estimated at each plot
(Table 2.6). Slope inclination and aspect were measured to the nearest degree
from plot center. In plots with variable microtopography, slope was measured at
several points and averaged. Elevation was determined to the nearest 10 ft (~ 3
m) using a topographic map or altimeter. The percent cover of different surface
substrates was estimated visually, with precision varying such that values
summed to 100%. Topographic position, slope shape (both horizontally and
vertically), soil drainage class, soil moisture regime, and inundation were
assessed using scalar values. Bedrock geology was determined to the greatest
precision possible by using existing geological maps, while the characteristics of
surface rocks present in a plot were recorded in the field.
Soil samples were collected from the top 10 cm of mineral soil (below the surficial
litter and humus) at 288 plots. Mineral soil was absent, or not possible to collect,
at 23 plots located on rock outcrops or boulderfields. As a rule, soil was
collected from several locations within a plot and mixed into a composite sample.
Depth of surface duff, soil color, and texture were evaluated in the field and
recorded on the plot forms. Soil samples were oven-dried, sieved (2 mm), and
analyzed for pH, phosphorus (P), soluble sulfur (S), exchangeable cations
(calcium [Ca], magnesium [Mg], potassium [K], and sodium [Na] in ppm),
extractable micronutrients (boron [B], iron [Fe], manganese [Mn], copper [Cu],
zinc [Zn], and aluminum [Al], in ppm), total exchange capacity (CEC; m.e.q./100
g), total base saturation (%TBS), and percent organic matter (%OM). Chemical
analyses were conducted by Brookside Laboratories, Inc., New Knoxville, Ohio.
Extractions were carried out using the Mehlich III method (Mehlich 1984) and
percent organic matter was determined by loss on ignition.
Evidence of any past or ongoing disturbances, including but not limited to
logging, fire, exotic plants, erosion, grazing/browsing, wind or ice damage,
hydrologic alterations, chestnut blight, dogwood anthracnose, southern pine
beetle, gypsy moth, and hemlock woolly adelgid, was recorded from each sample
site.
2.3.4 Sample Site Metadata:
Standard metadata, or information regarding the implementation of the sampling
protocol, were recorded at each plot. These included plot numbers, date(s) of
sampling, participants, geopolitical locality (county/city), survey site name, USGS
quadrangle, plot size and configuration, photographic documentation, and a
written description of the plot location. Plots were assigned unique alphanumeric
codes. A global positioning system (GPS) unit was routinely used to record
locational data with greater precision. For plots established prior to 2000, the
UTM (Universal Trans Mercator) coordinates of each plot location were
determined to 10 m (~ 33 ft) precision using either GPS or by using ArcView GIS
22
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
(Version 3.2; ESRI 1999), and all plot locations were mapped as precisely as
possible on USGS 7.5’ quadrangle maps. Plots established after 2000 were
mapped in the field using GPS receivers. Plot coordinates were either
differentially corrected or averaged from 30+ non-differentially corrected
positions. Accuracy of the post 2000 plot coordinates is estimate to be < 10 m (~
33 ft).
23
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Table 2.6. Topographic / hydrologic environmental indices recorded at each plot-sample
site (from protocol established by VANHP).
Topographic position: Soil Drainage Class:
A – plain / level A – very poorly drained
B – toe B – poorly drained
C – lower slope C – somewhat poorly drained
D – middle slope D – moderately well drained
E – upper slope E – well drained
H – crest F – rapidly drained
I – basin / depression
Inundation:
A - never
B - infrequently
C – regularly, for < 6 mos.
D – regularly, for > 6 mos.
Surface Substrate: % cover E – always submerged by shallow
decaying wood water (< 30 cm)
bedrock F – always submerged by deep
boulders and stones water (> 30 cm)
gravel and cobbles
mineral soil / sand Soil Moisture Regime:
litter / organic matter A – very xeric (moist for negligible time
water after precipitation)
other B – xeric (moist for brief time)
C – somewhat xeric (moist for short time)
Measured Slope (degrees) D – submesic (moist for moderately short
time)
E – mesic (moist for significant time)
F – subhygric (wet for significant part of
Slope shape growing season; mottles at < 20 cm)
Vertical G – hygric (wet for most of growing season;
C – concave permanent seepage / mottling)
X – convex H – subhydric (water table at or
S – straight near surface for most of the year)
Horizontal
C – concave I – hydric (water table at or above surface
X – convex year round)
S – straight Hydrologic Regime:
H – hummock and hollow microtopography Terrestrial (i.e., not a wetland)
I – irregular craggy/bouldery microtopography Non-Tidal
A – Permanently flooded
B – Semipermanently flooded
C – Seasonally flooded
Measured Aspect D – Intermittently flooded
______ degrees E – Temporarily flooded
F (flat) F – Saturated
V (variable)
24
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
2.4 Plot Data Analysis and Classification Methods
2.4.1 Data Preparation and Transformation:
Stem diameter measurements were used to compute density (stems/ha) and
basal area (m2/ha) for all woody plants at each plot. Basal area was calculated
by multiplying the geometric mean of each diameter class by the density of stems
within that class. Density and basal area were used to calculate importance
value, defined as the average of relative density and relative basal area for each
species.
Prior to analysis, most environmental variables were transformed, either to
normalize frequency distributions or to assign numeric values to categorical
variables. Topographic position, slope shape in vertical and horizontal directions,
and soil moisture regime were converted to ordinal variables (Table 2.7). While
the resulting absolute values of these variables are arbitrary, the rank orders of
values correspond to putative underlying environmental gradients. Aspect was
transformed using the cosine method of Beers et al. (1966), using the formula A'
= cos (45º - A) + 1, where A' = transformed aspect and A = aspect in degrees.
This transformation standardizes aspect to a linear variable from 0 (225º; SW,
dry, solar-exposed) to 2 (45º; NE, moist, sheltered), and can be used as a
surrogate variable for topographic moisture and solar exposure.
Surface substrate values were converted to decimals and arcsine transformed to
normalize their distributions. Since the values for all substrate classes sum to
100 and thus each can be defined as a linear combination of the others, a non-
vascular (bryophyte and lichen) substrate cover was added to eliminate
collinearity in surface substrate for most plots. Values for all soil variables except
pH were natural log-transformed to normalize their distributions and make the
values more biologically interpretable (Palmer 1993). A synthetic fertility index
(CEC x TBS/100) was also calculated for each plot.
Horizontal and vertical slope shape was also converted to a single ordinal
variable (scale = 0 to 10) using a modification of Parker (1982) (Table 2.7).
A synthetic Topographic Relative Moisture Index (TRMI) was calculated for each
plot using a procedure modified from Parker (1982). TRMI is a scalar ranging
from 0 (lowest moisture potential) to 60 (highest moisture potential) and
combining four topographic variables that potentially influence water runoff,
evapotranspiration, and soil moisture retention:
Slope inclination (10-point scale; per Parker [1982])
Slope shape (10-point scale; as above)
Aspect (20-point scale) = Beers-transformed aspect x 10
Topographic position (20-point scale) = (1-relative slope position) x 20
25
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Table 2.7. Ordinal variables used in analysis for scalar topographic and soil moisture variables
estimated in the field.
Topographic Position Slope Shape - Vertical and Horizontal
I - basin/depression = -1 C - concave = -1
A, J, K - plain/level, floodplain, stream bottom = 0 X - convex = +1
B - toe = 1 S - straight - 0
C - lower slope = 2
D, G = middle slope, ledge/terrace = 3 Slope Shape Index (SLSHI)
E, F = upper slope, escarpment/face = 4 vert. profile horiz. profile SLSHI
H = crest = 5 concave concave 10
concave straight 9
Soil Moisture Regime straight concave 7
A - very xeric = 1 straight straight 5
B - xeric = 2 straight convex 3
C - somewhat xeric = 3 convex straight 2
D - submesic = 4 convex convex 0
E - mesic = 5
F - subhygric = 6
G - hygric = 7
H - subhydric = 8
I - hydric = 9
Because mapped bedrock formations of Shenandoah National Park are
somewhat heterogeneous and contain similar lithologic units, each plot was
assigned to one of four aggregate geological classes (Table 2.8) based on the
prevalent surface rocks at the site; if no surface rocks were present, the
assignment was based on the mapped bedrock unit. Geologic substrate was
used in subsequent quantitative analyses by defining dummy (binary) variables
for classes 2, 3, and 4, with class 1 as the reference (ter Braak and Looman
1995).
Table 2.8. Aggregate geological classes used as dummy variables in data analysis.
No. Aggregate
Geological
Class:
Definition and relationship to formations mapped by Rader and
Evans (1993)1:
1 Alluvium heterogeneous, bouldery and cobbly stream-bottom alluvium derived
from and underlain by various formations
2 Acidic
Sedimentary
outcrops and debris of quartzite, metasandstone, metasiltstone, and
phyllite prevalent in Єch and Zsr
3 Granitic outcrops and debris of charnockite, charnockite gneiss, granite,
leucogranite, granulite, and related rocks in Yal, Yc, Ycm, Yor, and
Ypg
4 Mafic outcrops and debris of metabasalt prevalent in Czc
26
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
1 Names of formations:
Єch –Chilhowee Group (Antietam, Harpers, and Weaverton Formations)
ЄZc – Catoctin Formation
Yal – Leucogranite
Yc – Charnockite
Ycm – Chanockite Gneiss
Yor – Old Rag Granite
Ypg – Layered Pyroxene Granulite
Zsr – Swift Run Formation
Botanical nomenclature generally follows Kartesz (1999). As a rule, taxa were
treated at the highest level of resolution possible, but the identification of varieties
and subspecies was not always possible. A few taxa identified only at generic or
higher levels (e.g., “Carex sp.” or “unidentified woody seedling”) were deleted
prior to analysis.
2.4.2 Cluster Analysis:
Hierarchical, agglomerative cluster analysis, implemented in the software
program PC-ORD (version 4.17; McCune and Mefford 1999), was employed to
identify compositionally similar groups and generate a classification from the
combined 311 plot data set. During preliminary analyses, the Lance-Williams
Flexible-Beta linkage method (Lance and Williams 1966, 1967) was used in
conjunction with the Bray-Curtis coefficient of community (Bray and Curtis 1957)
to identify major groups in the data set. Based on these analyses, the full data
set was divided into six subsets containing, roughly, plots of 1) acidic forests, 2)
high-elevation forests, 3) low-elevation rich forests, 4) mesic and dry mesic
mixed forests, 5) rock outcrops, and 6) non-alluvial wetlands. Additionally, six
compositionally unique or heterogeneous plots were identified as outliers and
removed from the analysis.
Subsequent cluster analyses were conducted on each of the six groups using
three data treatments: 1) raw cover class scores, 2) cover class scores
relativized by site totals, and 3) cover class scores relatived by species maxima.
Moreover, analyses using each data treatment were run with two different
dissimilarity measures: the Bray-Curtis coefficient and Chord Distance
(relativized Euclidian distance). A beta setting of –0.5 was used in all analyses.
All six combinations of data treatments and clustering strategies performed
similarly in the analyses of each subset, producing dendrograms with similar
major divisions and plot groupings and a high percentage of plots with the same
finer-level group memberships. After examining the results from all six protocols,
the most ecologically interpretable dendrogram for each subset was accepted
(Appendix 1).
2.4.3 Compositional Summary Statistics:
27
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Compositional statistics were calculated to evaluate the adequacy of groups
recognized in cluster analysis and ultimately to assist in naming and describing
the community types. Initially, total mean cover and total frequency across all
311 plots were determined for every taxon. Cover class scores were converted
to the midpoints of their respective percent ranges, the midpoints were averaged,
and resulting values were back-transformed to cover class scores. For each
taxon in each group under consideration, the following summary statistics were
then calculated:
Frequency – the number of samples in a group in which a species
occurs.
Mean Cover – back-transformed cover class value corresponding to
mean percent cover calculated from midpoint values of cover class
ranges. All samples assigned to a group were considered when
calculating mean cover, not just those in which a taxon was present;
absences were assigned a cover value of 0.
Relative Cover – the arithmetic difference between mean cover (for a
given group of samples) and total mean cover (for the entire data set)
(= Mean Cover – Total Mean Cover). Expressed by plus or minus
symbols, this value provides a relative approximation of how much
more, or less, abundant a particular species is in a community type
compared to the overall data set.
Constancy – the proportion of samples in a group in which a species
occurs, expressed as a percentage (= [Frequency / Number of
samples in group] x 100). Because they are scaled to 100, constancy
values can be compared across community types with unequal
numbers of plots.
Fidelity – the degree to which a species is restricted to a group,
expressed as the proportion of total frequency that frequency in a
given group constitutes (= [Frequency / Total Frequency] x 100). An
accidental or exotic species can have maximal (100) fidelity to a type if
it occurs in only one sample in the entire data set. As a result, fidelity
alone can perform poorly as a criterion for identifying characteristic
species and distinguishing among types.
Indicator Value (IV) (= [Constancy x Fidelity] / 100). A synthetic value
indicating species that are both frequent within and relatively restricted
to a group of plots.
Indicator Value Adjusted by Cover, Scaled (Scaled Adj IV) (=
[Indicator Value x Mean Cover] / 9). By dividing IVxMC by 9, the
maximum possible cover value, this statistic synthesizes information
about frequency, diagnostic value, and mean abundance. A species
entirely restricted to a particularly community type, occurring in every
sample of that type, and attaining maximum mean cover will have a
Scaled Adjusted IV of 100 for that type. Empirically, taxa with Scaled
Adjusted IVs ≥ 15 are almost always those most characteristic of a
28
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
type, although the exact range of values in any given type or data set
may vary considerably.
Indicator Value Adjusted by Cover, Unscaled (Unscaled Adj IV) (=
Indicator Value x 2relative cover ). An alternative, unscaled synthetic
measure of adjusted IV, using relative cover as the modifier of IV.
Since cover classes form a logarithmic, rather than linear scale of
values, Unscaled Adjusted IV is a statistically more legitimate means of
incorporating information on cover, and has the advantage of not
favoring only dominant species and better identifying species that are
considerably more abundant within a given type than in the data set as
a whole. This statistic is sensitive, however, to vegetation types
containing few samples and to species with low overall frequency.
Additionally, the following statistics were generated for each group:
Mean Species Richness – the average number of species present
per plot (S ); only species rooted inside plot boundaries were included
in this calculation.
Homoteneity – the mean constancy of the S most constant species,
expressed as a fraction. This value (sensu Curtis 1959) can be
considered the constancy of the average species in a community type;
higher values for homoteneity indicate greater uniformity in species
composition among plots. Although homoteneity is not independent of
group size, often increasing as the number of group members
decreases, it can be used to evaluate whether community types have
been defined at an appropriate level.
These procedures were used to efficiently evaluate a sizeable number of groups
in the competing dendrograms generated by different cluster analysis protocols.
Several problematic plots which shifted among multiple groups depending on the
clustering protocol used were ultimately assigned to one group by evaluating the
statistical interpretability of each affected group with and without the questionable
plot, and by examining the position of the plot on the axes of non-metric
multidimensional scaling ordinations.
2.4.4 Community Type Structural Characterization:
The standard forestry statistics calculated for each plot (see section 2.3.2)
representing a community type were averaged to obtain a composite
characterization of woody vegetation for that type. In addition, the typical vertical
structure of each community type was determined by averaging cover class
scores of all woody species in each stratum across all plots representing the
type. Similarly, mean canopy height for a community type was obtained by
averaging the canopy height measurements from all representative plots.
29
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
2.4.5 Environmental Summary Statistics:
Mean values for continuous and ordinal environmental variables were calculated
for each group to aid in describing community types and identifying the
differences between them. These calculations were performed with raw
(untransformed) values, which were averaged across all plots representing a
given group. Mean aspect was calculated as the average position along an arc
defined by the range of aspect values.
2.4.6 Ordination:
The ordination method non-metric multidimensional scaling (NMDS; Kruskal
1964) was used to validate the classification, detect compositional variation and
trends that are obscured in cluster analysis, and aid in identifying the
environmental gradients along which vegetation classes and community types
are distributed. NMDS is a type of indirect gradient analysis that assigns
samples to coordinates in ordination space in a way that maximizes, to the extent
possible, the rank-order (i.e., non-parametric) correlation between inter-sample
distance in ordination space and inter-sample dissimilarity (i.e., ecological
distance; Minchin 1987). Ordination studies of each major compositional group
identified in cluster analysis, as well as of selected smaller groups of closely
related community types, were conducted (Appendix 1). NMDS was
implemented in PC-ORD (version 4.17; McCune and Mefford 1999). The Bray-
Curtis index was used to calculate dissimilarity and VARIMAX rotation was
employed to optimize axis placement in all ordination studies for this project.
Each ordination was computed using 40 random starting configurations, and
configurations with the lowest stress levels were used for interpretation.
Based on preliminary plots of stress vs. dimensionality, most ordinations were
extracted in three dimensions. Two-dimensional ordinations were used to
examine compositional variation within a few of the smaller groups. Pearson
correlations between environmental variables and sample coordinates on each
axis were calculated, and significant correlations were displayed through joint
plot overlays. Environmental variables used in ordination analyses were ordinal
variables for slope shape; continuous variables for arcsine-transformed surface
substrate values, Beers-transformed aspect, slope, elevation, raw and natural
log-transformed soil chemistry values; topographic relative moisture index
(TRMI), and dummy variables for geologic substrate. After preliminary studies,
the ordinal variable representing soil moisture regime was eliminated from the
analysis since it is redundant with, and less objective than, the synthetic TRMI
scalar.
2.4.7 Assignment of Classified Vegetation Types to the U.S. National Vegetation
Classification (USNVC) System:
30
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Once the classification was finalized, the classified vegetation types were
subjectively compared to existing units of the USNVC (Grossman et al. 1998,
Anderson et al. 1998, NatureServe 2002). All Shenandoah types were either
assigned to a conceptually similar USNVC type, or used as the basis for a new
USNVC unit. The global USNVC descriptions for existing types were edited, and
global descriptions for new types were written. Local park descriptions were
written for all classified types. During this process, the global and state
conservation ranks of each existing type were re-evaluated and modified if
needed, and all new types were ranked.
2.4.8 Development of Field Key to Shenandoah National Park Vegetation Types:
A draft dichotomous key for field identification of classified types was prepared by
NatureServe ecologists based on descriptions written by VANHP. It was
subsequently reviewed and modified by VANHP, and the final key was produced
after two days of field-testing in the park.
2.5 Image Processing and Classification
2.5.1 Hyperspectral Imagery:
Hyperspectral images from NASA’s AVIRIS platform were acquired for this
research. Two high altitude AVIRIS images were collected from NASA ER-2
aircraft, one on 14 May 2000 and another on 13 July 2001. High altitude AVIRIS
pixels have approximately 17 m spatial resolution and 224 bands at 10 nm
intervals between 400-2500 nm.
AVIRIS images were converted to reflectance, corrected for atmospheric effects,
and referenced to UTM map coordinates as described here. AVIRIS image
processing consisted of reading raw image data for four flight lines, two flight
strips per date to cover the extent of SNP. Each raw image strip showed a
cross-track illumination effect due to the bi-directional reflectance (BRDF)
properties of forest canopies and properties of the scanning AVIRIS sensor. This
brightness gradient was corrected using a tool in ENVI that fits a trend line to the
mean cross-track values for each band and then adds or subtracts the correction
factor at each cross-track pixel for each band. This method effectively removes
the dominant brightness gradient present in the raw image strips.
Cloud cover of different types was an issue in the hyperspectral image analysis,
as it can obscure or attenuate the reflectance of desired target areas. Both
AVIRIS images were affected by some cloud cover over parts of Shenandoah
National Park, though clouds were a larger problem for the summer image. The
image strips from 13 July 2001 had some coverage of cumulous clouds,
interfering with vegetation map creation. Unlike other types of clouds that
completely obscure the ground and create dark shadows, high cirrus clouds are
partially transparent and can contaminate image pixels by increasing the
31
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
brightness of the reflectance target in ways that are difficult to detect. Subtle
differences in cirrus cloud contaminated pixels could hamper modeling efforts by
adding atmospheric variation to pixel reflectance and obscuring variation due to
vegetative composition and structure. The May AVIRIS image for this analysis
had some high cirrus clouds that were affecting reflectance, predominantly at the
northern and southern ends of the image strips. These images were corrected
using a method developed by Gao et al. (1998) to remove high cirrus cloud
effects in AVIRIS imagery. The correction uses a relationship between apparent
reflectance at 664 nm (band 34) in the range of red visible light and reflectance
centered at 1374 nm (band 109), an atmospheric absorption window; the
relationship between these wavelength bands highlights the presence of high
cirrus clouds. A correction factor is derived by subtracting the slope of the
regression equation from the apparent reflectance at 1374 nm. This factor is
then subtracted from the apparent reflectance of each pixel at each wavelength
band. This method was developed to correct only bands from 400-1000 nm in
wavelength, and thus was used to correct only bands 1-41 in the spring AVIRIS
image only.
After removing the brightness gradient from images for both dates and the cirrus
cloud effects from the May 2000 image strips, each image strip was corrected for
atmospheric effects and converted to reflectance using the ACORN software
package. In order to transform the raw image strips into map coordinate space,
several hundred ground control points (GCP’s) were collected from the image
strips and corresponding digital USGS topographic quad grids for the entire park.
A total of 297 GCP’s were collected for the May 2000 AVIRIS images and 349
were collected for the July 2001 images. A triangulation method, also known as
rubber sheeting, was used to warp the image to UTM coordinates, zone17,
NAD83. Image irregularities in the raw imagery that resulted from the pitch and
yaw of the aircraft made a polynomial transformation method ineffective.
Triangulation is used when polynomial transformations are not possible and is
most accurate with a dense coverage of GCPs. Geocorrected image strips were
mosaicked into image mosaics for each date to produce the base image files for
the vegetation mapping analysis.
2.5.2 Landsat TM Imagery:
In addition to hyperspectral imagery, a time-series of Landsat TM images from
five dates was also used to test the ability of multi-temporal Landsat data to map
forest associations. The results of this analysis would be used to fill in small
portions of Shenandoah National Park that were not covered by AVIRIS image
data (see Figure 3.6). The time-series included three early spring images: 12
April 1984, 26 May 2000, and 24 May 2002, and two early fall images: 19
September 1984 and 5 September 2002. These dates were selected to take
advantage of the variable reflective properties of young leaves in spring and
older leaves late in the growing season. The images from 1984 were included
because they were taken before several major disturbances affected the park,
32
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
including gypsy moth defoliation, hemlock woolly adelgid defoliation, and large-
scale fires. Images from 2002 were used to incorporate the lasting effects of
these disturbances in the predictive model of vegetation distribution. It was left to
the statistical analysis to determine which predictive variables were most adept at
discriminating current vegetation conditions as seen in the field data, and to
weight those predictors accordingly. Each Landsat image was converted to
planetary reflectance and geometrically corrected to UTM map coordinates.
Landsat images were spatially accurate to within 30 m. Reflectance values from
TM bands 1-5 and 7 for each date (a total of 30 bands) were extracted as
described above and used in the canonical linear discriminant analysis.
2.5.3 Aerial Photography:
We acquired aerial photography as a reference source for interpreting vegetation
community types from satellite and hyperspectral imagery. We contracted with
an aerial photography firm (Air Photographics Inc., Martinsburg, WV) to acquire
1:24,000 scale color infrared photography over the park during leaf-on conditions
in late-August and early September 2001. Flights were flown with 60% overlap
and 30% sidelap resulting in 213 exposures delivered as color transparencies.
Of the 213 aerial photos imaged, only 156 are needed to completely cover the
park, the remaining 57 duplicate areas covered inside the park or are wholly
outside the park boundary. We had 156 images covering the park scanned at
800 dpi by a commercial firm (Spectrum Mapping LLC, Easton, MD) and
delivered as digital files in TIFF format. These images were orthorectified by
USGS-LSC to a topographic base using Erdas Imagine Orthobase software in a
block triangulation process. Output images are referenced to a Universal
Transverse Mercator (North American Datum of 1983) projection at 1-meter pixel
resolution. Average positional error of the orthorectified imagery is 2.4 meters
(root mean square error). Images were stored both individually and as collar-
clipped mosaics by 1:12,000 quadrangle.
We also compiled 1:12,000 digital orthophotography from the USGS Digial
Orthophoto Quarter Quardrangle (DOQQ) dataset as a reference source. The
data covering Shenandoah National Park were acquired by the USGS National
Mapping Division during leaf-off conditions in 1994 and 1997, and orthorectified
to 1-meter pixel resolution. Average stated positional error of this dataset is 2.3
meters RMSE.
2.5.4 Model Training Data:
Four reflectance spectra were extracted from the hyperspectral and multispectral
images for an area surrounding each of 305 vegetation plot locations using GPS
coordinates. These spectral samples covered an area of approximately 0.12 ha
(0.30 acres) around each plot for the AVIRIS analysis and 0.36 ha (0.89 acres)
for the Landsat analysis. The resulting spectral “bundles” were linked without
averaging to their respective forest plot data and used to map forest associations
33
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
developed from the field data. Figure 2.1 illustrates the spectral response
recorded by the AVIRIS sensor in May 2000 at 3 plots, with 4 bundles of samples
each. Due to AVIRIS image extents, not all of the plot data overlapped the image
data. Two hundred thirty-three (233) plots overlapped the 13 July 2001 AVIRIS
image mosaic and 265 plots overlapped the 14 May 2000 AVIRIS image mosaic
for a total of 932 and 1060 spectral samples, respectively. Some regions of
hyperspectral bands were excluded from the statistical analysis due to noise and
atmospheric interference. For AVIRIS, these excluded bands included bands 1-
6, 107-117, 153-170, and 217-224, leaving a total of 181 usable bands.
0
1000
2000
3000
4000
5000
6000
IMG6
IMG13
IMG20
IMG27
IMG34
IMG41
IMG48
IMG55
IMG62
IMG69
IMG76
IMG83
IMG90
IMG97
IMG104
IMG122
IMG129
IMG136
IMG143
IMG150
IMG175
IMG182
IMG189
IMG196
IMG203
IMG210
Band
Reflectance (*1000)
Figure 2.1. Example spectral response from AVIRIS hyperspectral sensor sampled at 3 plots
(red, orange, and maroon), with 4 spectral “bundles” each. A spectral bundle is the spectral
response recorded at the sample site and at 3 adjacent pixels to account for spatial registration
errors. Statistical analysis of spectral response in relation to vegetation type was conducted
using information from spectral bundles at sample sites.
The vegetation mapping models for each sensor employed the maximum number
of field plots intersecting with each image individually for a total of 265 plots for
the May AVIRIS mosaic, 233 plots for the July AVIRIS mosaic, and 299 plots for
the multitemporal Landsat TM imagery. Both the AVIRIS and Landsat data
covered the majority of the park and more than 98% of the vegetative inventory
plots sampled for a total of 34 community types.
2.5.5 Topographic Gradients:
Maps of digital topographic gradients were produced for Shenandoah National
Park to guide sample site selection, image processing, and vegetation
distribution modeling. Table 2.4 lists the topographic variables included in
34
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
statistical analysis and prediction of vegetation type. In addition to the
topographic based variables listed, maps of geologic type (from Morgan et al
2004) and image specific illumination were also used as inputs to statistical
analysis. Geologic types of alluvium, acidic sedimentary, granitic, mafic, and
calcerous rocks were input as dummy variables into statistical models. In lieu of
directly correcting for terrain related illumination effects in the imagery during
preprocessing, “Cosine I” images were created to approximate the illumination
angle of the sun that would be expected at the time of satellite image acquisition
(Townsend and Foster 2002). Inclusion of this data as predictive variables in
statistical analysis was meant to adjust for differential effects of solar illumination
on vegetation spectral reflectance.
The topographic gradient layers were initially created from a 15 meter resolution
digital elevation model (DEM) as described above. Topographic grids were
resampled to 17 meter resolution using nearest neighbor resampling for
equivalence with the AVIRIS hyperspectral data. Topographic gradients were
additionally resampled to 30 meter resolution using nearest neighbor resampling
for analysis with 30 meter Landsat TM data. Some topographic variables were
filtered using a low pass 3-by-3 filter before resampling to minimize noise in the
data analysis (e.g. planform curvature, profile curvature, relative moisture index,
relative slope position, topographic convergence index, topographic relative
moisture index, compound topographic index). Co-registered image and
environmental gradient models were used as input to statistical classification
models using specialized scripts written in IDL programming language for the
ENVI image processing system, and for the SAS statistical language.
2.5.6 Image classification (Canonical Linear Discriminant Analysis):
Canonical discriminant analysis (CDA) is a multivariate statistical method that
creates linear combinations of measured variables to optimize discrimination of
samples into predetermined classes. Unlike more generalized data
transformation and reduction techniques such as Principle Components Analysis
(PCA) and Minimum Noise Fraction (MNF), CDA requires a training sample
linked to predefined classes. When a training sample exists, CDA creates
transformed canonical variates or components that maximize the discrimination
between the classes of interest. Whereas PCA and MNF methods focus on
capturing most of the image-wide data variability within the first few transformed
dimensions, CDA focuses only on the data variability that is important to
discriminating between classes, which can produce unique results. Figure 2.2
illustrates the loading of outcrop classes along two canonical variates
constructed from image and topographic variables sampled at field plot locations,
and demonstrates the discrimination between these classes.
35
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 2.2. Illustration of loading of outcrop classes on two canonical axes constructed from
image and topographic variables using Canonical Discriminant Analysis (CDA). Colors and text
represent different classes of outcrop communities (e.g. O1, O2, etc.) and location of text labels
represent relative loading of sample plots in relation to two canonical axes. Plots from the same
vegetation community tend to cluster together, allowing prediction of vegetation class on the
basis of image and topographic information.
Canonical Linear Discriminant Analysis (CLDA) employs linear discriminant
analysis in conjunction with the canonical variates produced by the CDA to
predict class membership. The first step in the CLDA analysis was to run
stepwise discriminant analysis on all of the image and topographic variables
(sampled from the training data) resulting in a subset of variables useful for
discrimination between classes for each sensor. Table 2.9 lists the variables
selected from each sensor and topographic variable set from CLDA analysis.
Figures 2.3 and 2.4 illustrate the weighted importance of image and topographic
variables used as input to CLDA. The CDA for this study was done using the
CANDISC procedure in the SAS statistical software package. CANDISC
produced one less canonical variate than the number of classes for a total of 22
canonical components. Once these were generated, PROC DISCRIM was used
to create linear discriminant functions from combinations of the canonical
variates. The linear discriminant functions were then mathematically transformed
into probability values of class membership. The final models mapped the class
with the highest membership probability at the training data locations.
36
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Table 2.9. Image and topographic variables selected for use in canonical linear discriminant
analysis models. *Landsat multi-temporal imagery from April 1984, September 1984, May 2000,
May 2002, and September 2002.
AVIRIS May 2000 AVIRIS July 2001
Landsat Multi-
temporal*
Variable Name Variable Name Variable Name
IMGx 111 bands IMGx 98 bands IMGx 21 bands
TOPO0 Beers TOPO0 Beers TOPO0 Beers
TOPO1 Elevation TOPO1 Elevation TOPO1 Elevation
TOPO3 Plancurve TOPO3 Plancurve TOPO3 Plancurve
TOPO5 RMI TOPO5 RMI TOPO4 Procurve
TOPO6 RSP TOPO6 RSP TOPO5 RMI
TOPO7 Slope TOPO7 Slope TOPO6 RSP
TOPO8 Solar_ave TOPO8 Solar_ave TOPO7 Slope
TOPO9 TCI TOPO9 TCI TOPO8 Solar_ave
TOPO10 TRMI TOPO10 TRMI TOPO9 TCI
TOPO11 TSI TOPO11 TSI TOPO10 TRMI
TOPO13 COSI05 TOPO15 COSI07 TOPO11 TSI
TOPO19 Greenstone TOPO16 CTI TOPO12 COSI8404
TOPO20 Sandstone TOPO19 Greenstone TOPO13 COSI8409
TOPO21 Granite TOPO21 Granite TOPO14 COSI0205
TOPO22 Limestone TOPO22 Limestone TOPO16 CTI
TOPO19 Greenstone
TOPO20 Sandstone
TOPO21 Granite
TOPO22 Limestone
37
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
1260 nm
1023 nm
799 nm
655 nm
2041 nm
1603 nm
0
1000
2000
3000
4000
5000
6000
7000
8000
IMG6
IMG10
IMG14
IMG21
IMG25
IMG29
IMG33
IMG37
IMG41
IMG47
IMG54
IMG58
IMG61
IMG67
IMG75
IMG82
IMG86
IMG91
IMG94
IMG100
IMG117
IMG120
IMG133
IMG141
IMG146
IMG151
IMG172
IMG178
IMG183
IMG193
IMG200
IMG207
IMG214
TOPO1
TOPO6
TOPO9
TOPO15
TOPO21
Variables
Weighted Importance (W
i
)
Figure 2.3. Weighted importance of variables used in July 2001 AVIRIS image model for
discriminating between vegetation classes in CLDA analysis. Note the overwhelming importance
of hyperspectral image variables compared to topographic variables (refer to Table 2.9 for
explanation of variables).
38
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
COSI0205
COSI8409
COSI8404
TCI
RMI
0
200
400
600
800
1000
1200
1400
1600
1800
2000
IMG0
IMG1
IMG2
IMG3
IMG6
IMG9
IMG10
IMG11
IMG14
IMG15
IMG16
IMG 17
IMG 18
IMG 21
IMG 22
IMG 23
IMG25
IMG26
IMG28
IMG29
IMG30
TOPO0
TOPO1
TOPO3
TOPO4
TOPO5
TOPO6
TOPO7
TOPO8
TOPO9
TOPO10
TOPO11
TOPO12
TOPO13
TOPO14
TOPO16
TOPO19
TOPO20
TOPO21
TOPO22
Variables
Weighted Importance (W
i
)
Figure 2.4. Weighted importance of variables used in multi-temporal Landsat image model for
discriminating between vegetation classes in CLDA analysis. Note the importance of topographic
variables compared to image variables (refer to Table 2.9 for explanation of variables).
Coefficients from the CDA and the linear discriminant functions were then applied
to the combined image and topographic data to determine the most likely
vegetation class at each pixel based on statistical functions. Results of this
process are maps of predicted probabilities for each vegetation class. These
probability maps are then combined and the maximum class probability at each
pixel is determined and used to assign the vegetation type on the final map.
2.5.7 Post Classification Smoothing:
Statistical classification procedures used in our analysis (as well as other pixel-
based processing) predict vegetation class based on statistical properties at
individual pixels. Slight variations in properties between pixels (due to
measurement precision variability, or random errors) can lead to adjacent pixels
being assigned to different classes. The resulting map often has a “salt and
pepper” speckled pattern from the intermingling of pixels of different classes.
The goal of post-classification smoothing is to separate the “noise” pixels in this
speckling pattern from the underlying vegetation community “signal” to form
contiguous patches representing land cover classes on the ground. The typical
process for smoothing out noise pixels is to use a majority filter operation to re-
assign individual pixels to the class represented by a majority of surrounding
pixels, or to eliminate individual pixels and replace them with surrounding pixels
39
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
in a “sieve and clump” operation. However, these operations do not consider the
information content of the pixels being eliminated, just their adjacency. Finding
methods to consider both the information content and adjacency of pixels for
intelligent post classification smoothing is an active area of research (see Qian,
et al 2005). Due to the fact that maps of predicted probability for each class are
produced with the CLDA classification method, it is possible to use this
information to smooth out noise to produce contiguous patches, without losing
information content.
We used GIS to compute the mean prediction probability within a 40-meter
radius (~ 0.5 hectare) circular area around each pixel. This had the effect of
generalizing the probabilities within a circular area equivalent to 0.5 hectares (the
minimum mapping unit of the project). We generalized prediction probabilities in
this manner for each of the 34 vegetation communities mapped. We combined
the probabilities into a final generalized map by identifying the maximum
predicted probability for each pixel on the filtered map, and assigning the class
with the highest probability to the output map.
2.6 Accuracy Assessment
2.6.1 Generation of Random Sampling Points Stratified by Map Unit
We generated sample points for an accuracy assessment (AA) field survey using
a draft vegetation map produced from AVIRIS hyperspectral imagery and
topographic gradient models. We used a random point selection script in GIS to
select areas from the draft map that were at least 50 m (164’) from a road (to
minimize disturbance effects), but no more than 250 m (820.2’) from a road or
trail (to minimize travel time). We excluded from AA point selection areas with
canopy defoliation or disturbance occurring between 1984 and 2002 by
assessing changes between leaf-on 1984 and 2002 Landsat Thematic Mapper
satellite imagery (30 meter (84.2’) pixel resolution). We randomly selected AA
survey locations for each vegetation community type within remaining park areas
that were not in a disturbed or road influenced zone, but that were easily
accessible by roads or trails. The numbers of points allocated to each community
type were generated proportional to the area of the vegetation community within
the park as determined from a draft map of vegetation polygons. For 2004, 280
points were selected for field assessment of which 224 points were visited
representing 32 of 34 vegetation communities identified in the park.
For 2005, we generated additional accuracy assessment points in an attempt to
sample underrepresented types from the 2004 AA field survey. We evaluated
existing plot data (vegetation mapping plots, 2004 accuracy assessment plots,
long-term ecological monitoring plots, plant protection plots) to find areas that
were under-sampled, and we examined draft vegetation maps to find vegetation
communities that were underrepresented. We buffered existing plot data to 200
meters and removed these areas from consideration for new plot selection to
40
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
keep new AA plots away from previously sampled areas. We also examined plot
representation of vegetation communities, and randomly placed plots in
vegetation types that were under-represented on the basis of percent of park
area covered by the community, and number of previous plots placed in that
community. We targeted 199 random plots, and prioritized plots to insure that
those community types least represented by prior field sampling would be visited
first if all 199 plots could not be assessed in the field season. Of the 199 selected
points, 68 were visited in the field in 2005.
2.6.2 Accuracy Assessment Field Sampling Protocol
Lists of random plot coordinates and maps showing locations of plots to visit
were delivered to VANHP (2004) or SHEN (2005) for use by field personnel.
A set of Accuracy Assessment (AA) protocols and a customized AA field data
collection form (Appendix 4) consistent with the USGS-NPS Vegetation Mapping
Program were developed in consultation with NatureServe. Subsequently, data
were collected in the field by VANHP botanists and NPS biological technicians
with no prior involvement with the project. The general procedure for this task
was to navigate as closely as possible to the pre-selected point using a “wide
area augmentation system” (WAAS)-enabled Garmin GPS unit. At the point, a
new GPS waypoint was collected, a 0.5 ha (1.236 acre) circular plot (radius ~ 40
m, 131.2’) was established, and required environmental and floristic data were
collected from the plot.
Using the field key to park vegetation types (Appendix 3) and draft descriptions,
AA field crews identified the community type in each sample plot, and also
recorded information about vegetation in a 0.5 ha (1.24 acre) around the AA
location, to the extent that it was observed en route to and from the sample point.
Sufficient data on vegetation structure, floristic composition, and environmental
setting were recorded to evaluate the degree to which vegetation at sample sites
was representative of a classified type. Where an exact match was not found
between observed vegetation and the vegetation community described in the
key, the closest matching types were recorded. For 2004 surveys, the degree of
certainty the field crews had in the vegetation community type assignment was
recorded on a numeric scale from 1-5. For 2005 surveys, the degree to which
the observed vegetation matched the “archetype” of the community was recorded
on a numeric scale from 1-5.
2.6.3. Accuracy Assessment of Map Data
Internal Accuracy Assessment of CLDA Models:
Predictive accuracy of the CLDA models from each sensor were evaluated using
two separate metrics, resubstitution accuracy and cross validation accuracy.
Resubstitution accuracy reflects how many of the plots are accurately placed in
the correct class by the model. This is sometimes referred to as the percent
41
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
plots correct or PPC and is indicative of the model’s ability to accurately predict
the training data set. Cross validation accuracy is a more conservative estimate
calculated by creating the model n times, each time dropping out one of the n
observations when building the model and using the dropped observations to
validate the model. Cross validation is a more appropriate estimate of the
model’s ability to predict new data, though it is not as desirable as using an
entirely separate dataset for validation. Because the final map product for SHEN
was a merged version of results from different dates and sensors, the internal
cross validation accuracies for the individual models could not be used to
evaluate the final map
Incorporation of Field Sampling Data into Accuracy Assessment:
Field validation of the final vegetation mapping product is an integral part of
National Vegetation Mapping Program efforts. Field validation plot data was
incorporated into the accuracy assessment of the final map using a modified
fuzzy evaluation approach. Fuzzy accuracy assessment methods help address
the discrepancy between actual mixed forest composition that varies
continuously over the landscape and “hard” classes that assign a pixel or
polygon to a single discrete vegetation type. Fuzzy assessment generally
acknowledges that for a given point on a classified map, when choosing among
similar forest associations or classes, an on-the-ground observer may find two or
three classes within the key that would acceptably characterize the site. This
observation is understandable given the statistical clustering and ordination
methods used to derive vegetation classes from the field plot data, as there will
always be some overlap within the variance of each vegetation class. Within
these overlapping regions, plots could be characterized as multiple types and still
be considered accurately mapped.
This phenomena of multiple “right” answers for validation site class assignment
had to be addressed when comparing the field validation data to the final map
product for external accuracy assessment. The first issue that needed to be
addressed was the change from a preliminary map used to locate field validation
sites to the final map product. The preliminary map was the result of a different
statistical technique (Classification Trees) that did not prove as effective as the
CLDA modeling technique. Two issues in the validation data arose as a result of
the change in methods and maps. The first was that the original proportion of
map classes designed to be represented in the field validation sites was not
necessarily maintained. The second issue was that validation plots that were
placed within at least 0.5 ha (1.24 acre) polygons and away from polygon edges
on the preliminary map often fell haphazardly along vegetation transitions or
boundaries on the final map. This resulted in multiple map classes being
represented by a majority of the ~0.5 ha (1.24 acre) field validation sites, which
further required the use of fuzzy analysis in the accuracy assessment. The third
major issue with the field validation data was that the field crew often recorded
two and even three possible classes for a given validation site, perhaps
42
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
recognizing the fact that multiple classes were acceptable or that the .5 ha (1.24
acre) area included multiple types. Their notes also indicated that sites with
varying amounts of disturbance were more difficult to place within the
classification scheme. This is expected as the original sampling was designed to
avoid disturbed areas and instead focused on undisturbed forest types.
An accuracy assessment scheme was designed to account for a majority of the
issues discussed above. In order to compare map classes to field validation
observations, a 4x4 pixel analysis window (.46 ha, 1.14 acres) was placed
around field validation sites to approximate the 0.5 ha (1.24 acre) area observed
by field validation teams. All the classes occurring within each site window were
extracted using GIS. Each window was then visually assessed and up to four
possible correct classes were recorded. In most cases, the first two classes
included the center pixel and majority class (most common class occurring in the
window) if they were different. Additional possible classes were included when
the window was evenly divided between 3 or 4 different types or when a type
occurred in the window and was clearly very common just outside the window’s
extents. Using this methodology, 90% of the validation sites had at least 2
reasonable map classes within the analysis window. Only 23% of the sites had
up to 4 possible correct classes. In comparison, 37% of the sites had at least 2
possible classes assigned to them in the field and 7% had 3 classes recorded in
the field validation.
In an effort to increase sample size of accuracy assessment plots, collaborators
at Virginia Natural Heritage (G. Fleming) examined vegetation composition at
plots previously sampled in the park for the Long-term Ecological Monitoring
(LTEM) program (1991-2001) and plant protection activities (2002-2004). A total
of 277 plots were examined for vegetation composition. Of these plots, only 249
represented unique locations as more than one LTEM plot may have been
recorded for each LTEM site. Plots were assigned geologic parent material type,
elevation, aspect, slope, slope shape, topographic position, and landform codes
to assist in assignment into vegetation community type. Each plot record was
examined for vegetation composition and environmental attributes, and was
assigned to the closest matching vegetation community type (e.g. map code)
based on expert judgment. Where assignment to one type was not obvious, a
second community type was listed. A “confidence level” of “confident” (80-100%
certain), “probable” (50-80% certain), or “uncertain” (10-50% certain) was also
assigned to each plot to denote how well the plot fit into the vegetation
community class. Extenuating circumstances influencing the classification, such
as disturbance or succession, were noted.
3. RESULTS
3.1 Landforms and Ecological Land Units
43
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Ecological land unit mapping resulted in the creation of two new GIS layers:
landforms and ecological land units (ELU) based on the 15 meter resolution
digital elevation model. Other layers used as inputs to the ELU mapping effort
(geology and elevation) were simply recoded from existing maps. Fourteen
distinct landforms were mapped in the park. Most abundant landforms by area
are sideslopes and steep slopes, but significant areas of cove and upper slope
exist in the park as well (Figure 3.1). Landform types are illustrated in Figure
3.2.
828.09
20812.32
1277.06
9311.67
541.62
17008.54
6242.40
13281.55
4627.67
2349.36
494.46
1458.61
12.80 0.20
0
5000
10000
15000
20000
25000
Cliff
Steep slope
Slope crest
Upper slope
Flat summit
Sideslope N/NE
Cove N/NE
Sideslope S/SW
Cove S/SW
Dry flat
Slope bottom
Stream
Wetland
Lake
Hectares
Figure 3.1. Landforms of Shenandoah National Park by area (ha).
Of the possible 210 combinations of elevation, geology, and landform types, 163
distinct ecological land units were mapped in the park. However, 40 ELUs
represent less than 10 hectares each, and 15 of those comprise less than 1
hectare each. Twenty-eight ELU types represent over 1000 hectares, with the
largest types being in mid elevation-basaltic-side slopes (8,583 ha), mid
elevation-granitic-steep slopes (7,916 ha), mid elevation-siliciclastic-steep slopes
(7,763 ha) and mid elevation-basaltic-N/NE facing side slopes (7,012 ha).
Overall, the composition of ELU’s in the park is quite diverse, indicating the
potential of numerous microhabitats available to plants.
44
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.2. Example landform mapping results, Jeremy’s Run area, North District, Shenandoah
National Park
45
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
3.2 Vegetation and Accuracy Assessment Plots
As described in section 2.3, 311 total plots were sampled and used for
classification in this project. Of the 311 total plots used for classification, 208
plots were sampled during the growing seasons of 2001, 2002, and 2003.
Additionally, 93 plots were previously sampled using VANHP protocols during the
period 1999-2001, and 10 additional plots were sampled prior to 1999. Figure
3.3a displays locations of sampled plots. Appendix 5 lists the plot coordinates of
plots sampled and used for classification (n=305), as well as for plots dropped
from statistical consideration as outliers (n=6). Figure 3.3b displays locations of
2004 AA plots. Figure 3.3c displays locations of 2005 AA plots.
46
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.3a Location of vegetation field sampling plots (n=305) used for classification scheme
development and as training sites for mapping, Shenandoah National Park.
47
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.3b Location of accuracy assessment (AA) field sampling plots (n=224)
collected in 2004, Shenandoah National Park.
48
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.3c Location of accuracy assessment (AA) field sampling plots (n=68)
collected in 2005, Shenandoah National Park.
49
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
3.3 Vegetation Classification Scheme
Based on the combined results of cluster analysis, summary statistical analysis,
and ordination studies, 34 natural community types were recognized in the
classification of Shenandoah National Park data. Three hundred and five (305)
sampling sites and 762 vascular plant taxa were used in this classification. Six
(6) sampling sites of the original 311 sites field sampled were dropped due to
missing or incomplete data, or because they were outliers in the resulting
classification and did not represent existing or unique vegetation associations.
Membership in community types ranges from one to 31 plots. Thirty one (31)
community types were assigned to existing USNVC units, while 3 new types
were described and ranked.
The basic unit of classification is equivalent to the “association” recognized in
traditional vegetation studies (Barbour, Burk, and Pitts 1987; Mueller-Dombois
and Ellenberg 1974) and the U.S. National Vegetation Classification (USNVC;
Grossman et al. 1998), representing stands of vegetation of relatively
homogeneous composition that share a set of characteristic species and recur on
the landscape under similar environmental conditions. Protocols of the USNVC
were followed in naming the community types, using the scientific names of up to
six characteristic species, with distinct vertical strata indicated. As a rule,
species are listed by descending order of importance and structural position, i.e.,
canopy species are listed first, followed by understory species, shrubs, and
herbs. Nominal species in the same stratum are separated by a dash (-), while
different strata are separated by a slash (/). Redundant varietal and subspecific
epithets (e.g., Quercus rubra var. rubra) are not used in community names. The
characteristic physiognomy (i.e., forest, woodland, shrubland, etc.) of a type is
listed at the end of its name.
For convenience, community types are aggregated into the ecological groups
and ecological classes of Fleming et al. (2006) representing natural groups of
vegetation types sharing gross climatic, topographic, edaphic, physiognomic, and
floristic similarities. The hierarchical arrangement of community types within the
formal USNVC physiognomic-floristic hierarchy is presented in the index to
Appendix 2.
Despite encompassing an inevitable degree of variation, the community types
defined in this study are generally recognizable in the field and potentially
mappable. However, because the classification is based on composition in all
layers, not just the tallest, these community types differ considerably from “cover
types” (sensu Eyre 1980) used in forestry and large-scale vegetation mapping.
Since our purpose is to define ecological units, all plants at a site are considered.
In forests, for instance, shrubs and herbs often respond to more subtle
environmental gradients and may reveal more about local site conditions and
associated animal species than do trees, which tend to be more broadly
distributed and exhibit less environmental specificity. Likewise, herbaceous
50
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
species occurring with low cover may be more restricted to certain site conditions
and thus far more diagnostic of a community type than more widespread,
dominant shrubs and trees. The species used as nominals may be characteristic
of a type because of their abundance, constancy, or relative restriction to the
type. Although they can never be surrogates for descriptions, the names of
communities are constructed so that one can distinguish among types, identify
types readily in the field, and assign new stands to previously classified types. In
order to meet the first objective, an emphasis has been placed on diagnostic
species (e.g., those with high adjusted IV values). However, in the prevailing
mixed forests of this area and other regions in the eastern United States,
characteristic canopy trees are usually not restricted to a particular type. Many
forested types, despite having distinctive total floristic compositions, have
variable overstories composed of wide-ranging tree species with low fidelity and
indicator value. Exclusion of such species from a community name altogether is
not desirable and obfuscates the ready identification of the type in the field.
Hence the approach typically taken by VANHP ecologists in naming forest
community types involves the combined use of indicator, constant, and dominant
species, with nominal species of the tree strata often common to multiple types,
and nominal species of the shrub and herb strata contributing more diagnostic
value.
51
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
3.3.1 Hierarchical Classification of Ecological Groups and Associations:
The hierarchical classification and 34 natural communities of Shenandoah
National Park are summarized in this section. Detailed descriptions of each
community type are provided in Appendix 2.
I. HIGH-ELEVATION COMMUNITIES
1. Central Appalachian Northern Hardwood Forests
a. Betula alleghaniensis – Quercus rubra / Acer
(pensylvanicum, spicatum) / Dryopteris intermedia –
Oclemena acuminata Forest
Central Appalachian Northern Hardwood Forest (Yellow
Birch – Northern Red Oak Type)
(USNVC CEGL008502, map code F7, 8 plots)
This association is limited to relatively small patches at the
highest elevations (range 960 to 1225 m, 3150 to 4019 ft),
particularly on rocky, west to north-facing slopes. Soils are
extremely acidic with low base status and very high organic
matter content. Dominant overstory trees are typically
somewhat stunted and gnarled from repeated ice and wind
damage. The type is somewhat transitional between high-
elevation northern red oak forests (CEGL008506) and high-
elevation boulderfield woodlands dominated by Betula
alleghaniensis (CEGL008504), intergrading with both along
topographic gradients.
2. Northern Red Oak Forests
a. Quercus rubra – Quercus alba / Ilex montana /
Dennstaedtia punctilobula – Carex pensylvanica –
Deschampsia flexuosa Forest
Northern Red Oak Forest (Pennsylvania Sedge – Wavy
Hairgrass Type)
(USNVC CEGL008506, map code F9, 25 plots)
In the Park, this association is limited to gentle, mostly
convex slopes and crests on the highest metabasalt and
granitic ridges. It forms an extensive, almost continuous
patch in the central district from the vicinity of Big Meadows
N to the vicinity of The Pinnacle and Marys Rock. Smaller,
outlying patches occur on Hightop, Stony Mt, The Sag,
Mount Marshall, Hogback and other high-elevation ridges.
Mean elevation of plots samples is 1060 m (3478’). Soils
are extremely acidic and infertile, sometimes bouldery but
usually with relatively low surface cover of rocks. The
52
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
vegetation is typically an open, stunted forest dominated by
somewhat gnarled Quercus rubra and containing moderately
diverse understory and herbaceous layers.
3. High-Elevation Boulderfield Forests and Woodlands
a. Betula alleghaniensis / Sorbus americana – Acer
spicatum / Polypodium appalachianum Forest
Central Appalachian High-Elevation Boulderfield Forest
(USNVC CEGL008504, map code O4, 4 plots)
This very distinctive association is restricted to high-
elevation (mean = 1070 m, 3510’), mostly west to north-
facing boulderfields of both metabasalt and granitic rubble.
The physiognomy is mostly that of a woodland, and
overstory trees are typically stunted and gnarled from
frequent ice and wind damage. Betula alleghaniensis is the
overwhelming canopy dominant, and community floristics
are characterized by northern and high-elevation species.
High cover of diverse lichens and bryophytes is also typical.
Surface substrate averages 76% cover of loose boulders
and stones; mineral soil could not be extracted from
sampling sites. Large, outstanding examples occur on the
north flanks of Hawksbill and Stony Man. This vegetation
type intergrades with the Central Appalachian Northern
Hardwood Forest (CEGL008502), which occupies adjacent
sites with somewhat lower boulder cover and greater soil
development.
4. High-Elevation Outcrop Barrens
a. Diervilla lonicera – Solidago simplex var. randii –
Deschampsia flexuosa – Hylotelephium telephioides –
Saxifraga michauxii Herbaceous Vegetation
High-Elevation Greenstone Barren
(USNVC CEGL008536, map code O1, 9 plots)
This community type represents high-elevation metabasalt
outcrop barrens occurring at Franklin Cliffs, Hawksbill,
Crescent Rock, Stony Man, and Mount Marshall. The
vegetation occupies massive, wind- and ice-blasted
metabasalt exposures on upper, west- to north-facing ridge
flanks. Mean elevation of sampled stands is 1090 m (3576’),
and exposed rock cover averages 67%. Soils (very limited)
are extremely acidic and infertile. The vegetation is
characterized by a patchwork of shrub thickets (typically <
25% cover in plot samples), herbaceous mats (typically <
40% cover), and crustose lichen colonies on exposed rock
53
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
surfaces. Northern and high-elevation species predominate
and ten state-rare plant species are associated, including
several long-range boreal disjuncts. This community type is
endemic to six sites in Shenandoah National Park. There
are probably less than 20 discrete outcrops supporting it,
and the extreme rarity and small-patch size merits the
assigned global rank of "G1."
b. Photinia melanocarpa – Gaylussacia baccata / Carex
pensylvanica Shrubland
High-Elevation Outcrop Barren (Black Chokeberry
Igneous/Metamorphic Type)
(USNVC CEGL008508, map code O3, 3 plots)
In the Park, this association occupies high-elevation
metabasalt outcrops similar to those of the preceding.
Habitats are more west-facing (vs northwest-facing), and
have a lower mean elevation (975 m, 3199’). Soils (very
sparse) are even less fertile than those of CEGL008536.
The vegetation is characterized by shrub thickets
(particularly of Photinia melanocarpa and/or Gaylussacia
baccata), sparse herbaceous patches, and lichens.
Although the known occurrences are on metabasalt,
additional examples of this type could occur on granitic rocks
in the Park. This shrubland occurs on outcrops of several
plutonic and metavolcanic formations along the length of the
Blue Ridge in VA. Most known examples are south of the
Park, and some extend to elevations as low as 730 m
(2395’).
c. Kalmia latifolia – Vaccinium pallidum Shrubland
High-Elevation Heath Barren / Pavement
(USNVC CEGL008538, map code O2, 3 plots)
Middle- to high-elevation cliffs and outcrops of granitic rocks
at Old Rag and Millers Head support this shrubland in the
Park. Additional occurrences may be present on other
granitic outcrops, and possibly on some higher quartzite
exposures. Exposed rock cover averages 52% and soil
(extremely acidic and infertile) is limited. Total vegetation
cover is typically < 25% in a plot sample and is characterized
by dwarfed ericaceous shrub thickets with scattered,
severely stunted Pinus spp. and Quercus spp.
54
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
II. LOW- TO MIDDLE-ELEVATION MESIC FORESTS
1. Rich Cove and Slope Forests
a. Liriodendron tulipifera – Aesculus flava – (Fraxinus
americana, Tilia americana) / Actaea racemosa –
Laportea canadensis Forest
Southern Appalachian Cove Forest (Typic Montane Type)
(USNVC CEGL007710, map code F10, 17 plots)
This is a lush mesophytic forest community of lower
elevations on substrates weathered from metabasalt and
pyroxene-bearing granites. Many or all sites supporting this
vegetation were cleared or cut-over in the past. Elevation
range of plot samples is 311 to 985 m (1020 to 3232 ft),
mean of 629 m (2064 ft), with lower-slope topographic
positions and easterly aspects prevalent. Slopes are
concave in one or both directions and sites have relatively
high moisture potential (TRMI). Soil samples are moderately
acidic with moderately high Ca, Mg, Mn, and TBS levels.
The herb layer of this association is very lush with patch-
clonal forbs such as Laportea canadensis and Caulophyllum
thalictroides. Species characteristic of higher elevations are
mostly lacking from this type. The current vegetation is more
or less the result of secondary succession; the abundance of
understory Acer saccharum in some stands is probably an
indicator of ongoing compositional changes (see
CEGL006237 below). The assignment of this vegetation to
the primarily southern CEGL007710 is a bit problematic, but
it seems to fit fairly well, if a gradual shift in species
composition and elevation is accepted. In the park, this
community intergrades with nearly monospecific
successional forests of Liriodendron tulipifera (CEGL007220,
see below) along a seral gradient. It also may intergrade with
the Park's other rich cove forest (CEGL006237), which
typically occurs at higher mean elevation.
b. Acer saccharum – Fraxinus americana – Tilia americana
– Liriodendron tulipifera / Actaea racemosa Forest
Central Appalachian Rich Cove Forest
(USNVC CEGL006237, map code F15, 5 plots)
Similar to the preceding but occurring at somewhat higher
elevations. Sampled or documented in accuracy
assessment primarily in coves of the central section of the
Park (both flanks), and near Loft Mountain in the southern
section. Occupies middle to upper-slope ravines and
northerly slopes (often very bouldery) underlain by
55
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
metabasalt at ~640 to 1070 m (2100 to 3510 ft) elevation.
Soils are extremely acidic, but have moderately high base
cation levels. In the Park, this association is characterized
by dominance or co-dominance of Acer saccharum and a
suite of characteristic, nutrient-demanding mesophytic forbs,
including Aconitum reclinatum, Viola canadensis, and
Angelica triquinata. Liriodendron tulipifera is less
characteristic of this type than the preceding, and is absent
from higher-elevation stands. However, the type may
intergrade with the Park’s lower-elevation rich cove forest
(CEGL007710) at intermediate elevations, or along a seral
gradient.
2. Acidic Cove Forests
a. Pinus strobus – Quercus (rubra, alba) – Liriodendron
tulipifera Forest
Central Appalachian Acidic Cove Forest (White Pine – Mixed
Hardwoods Type)
(USNVC CEGL006304, map code F12, 7 plots)
This mixed hardwood-white pine forest community was
documented from low-elevation coves on the western side of
the Park. It occupies mesic lower slopes and flats at
elevations < 600 m (1968’). Sites are underlain by
metabasalt and charnockite but may have significant regolith
of acidic colluvium from upslope Chilhowee group
metasedimentary rocks. Soils are intermediate in base
status. Expressions of this vegetation in the Park are
typically well-developed, moderately diverse forests. Pinus
strobus varies in abundance from widely scattered to
dominant over small areas. Although not plot-sampled on
the eastern flank of the Park, observations indicate that this
type forms patches at the lower elevations on that flank as
well.
3. Mesic Mixed Hardwood Forests
a. Fagus grandifolia – Quercus (alba, rubra) – Liriodendron
tulipifera / Polystichum acrostichoides Forest
Mid-Atlantic Mesic Mixed Hardwood Forest
(USNVC CEGL006075, map code F6, 1 plot)
This distinctive forest community, dominated by Fagus
grandifolia, was plot-sampled at a single site in a low-
elevation ravine at the foot of the western slope bordering
the Shenandoah Valley. Although bedrock is shale and
limestone of the Waynesboro Formation, the surface regolith
56
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
probably consists of acidic colluvium from upslope
Chilhowee rocks. Soils are intermediate in base status.
This stand appears to represent an outlier of vegetation that
is characteristic and widespread on mesic uplands of the
Piedmont Plateau to the east of the Park. Several additional
patches of this vegetation have been observed or
documented in accuracy assessment; most are in low-
elevation hollows underlain by granitic rocks on the eastern
periphery of the Park.
4. Eastern Hemlock-Hardwood Forests
a. Tsuga canadensis – Betula alleghaniensis Lower New
England / Northern Piedmont Forest
Hemlock – Northern Hardwood Forest
(USNVC CEGL006109, map code F8, 7 plots)
This hemlock-hardwood forest occupies high-elevation
ravines and, less frequently, cool sheltered sites at lower
elevations. With one exception, plot-sampled sites are
situated on gentle toe slopes and bottoms with abundant
moisture; other examples in steep, rocky ravines are known
in the Park but were not sampled. Elevation ranges from
378 to 995 m (1240 to 3264 ft), mean ~ 800 m (2625’).
Surface substrate is often characterized by abundant
boulder cover and soils are extremely acidic, but with
moderately high Ca and Mg levels, probably reflecting the
influence of materials weathered from Catoctin metabasalt or
pyroxene-rich granites. Sites supporting this vegetation
often have a mixed hydrology, i.e., they are essentially
"uplands" with small seep or stream inclusions. This type
represents hemlock-northern hardwood vegetation with
northern affinities. It co-occurs with the High-Elevation
Seepage Swamp (CEGL008533) type, and may grade into it
as site conditions become more generally saturated. Stands
of this community are undergoing phyiognomic and
compositional changes due to extensive, adelgid-related
mortality of Tsuga canadensis.
5. Early-Successional Mesic Forests
a. Liriodendron tulipifera / (Cercis canadensis) / (Lindera
benzoin) Forest
Successional Tuliptree Forest (Circumneutral Type)
(USNVC CEGL007220, map code F13, 4 plots)
This association represents nearly monospecific
successional forests that grew up on fields abandoned in the
57
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
early 20th century. All sample plots occur on mesic, lower or
middle slopes at < 450 m (1476’) elevation, although the
type was documented up to 823 m (2700’) elevation during
accuracy assessment. Underlying bedrock is metabasalt or
pyroxene-rich granites and soils are relatively fertile. In
cluster analysis, the plots forming this type were not
separable from plots of the Montane Alluvial Forest (F11)
and had to be segregated using other analytical methods.
The environmental context of these two communities is
distinctly different, although both share a diversity of weedy
species and have similar histories of extensive
anthropogenic disturbance. In the Park, this type appears
to be a precursor of, and often transitional with, the lower-
elevation rich cove type (CEGL007710) above, and is
consistently characterized by almost monospecific overstory
dominance by Liriodendron tulipifera and dense Lindera
benzoin shrub layers. The sites supporting this type may
have been disturbed more heavily or for a longer periods
than those supporting rich cove forests.
b. Robinia pseudoacacia Forest
Black Locust Successional Forest
(USNVC CEGL007279, map code F21, 6 plots)
An early-successional forest community associated with
abandoned fields and areas around old home sites. This
type is scattered throughout the park but extends to higher
topographic positions and higher elevations than does the
ecologically similar Liriodendron tulipifera / (Cercis
canadensis) / (Lindera benzoin) Forest (CEGL007220).
Underlying bedrock is mostly metabasalt or granitic, and
soils are intermediate in fertility. Although the six plots are
basically from just two sites, similar vegetation has been
observed or documented during accuracy assessment in
many places in the Park, e.g., Milam Gap, Big Meadows,
South River picnic area, Loft Mountain, etc. Current stands
in the Park probably represent vegetation successionally
transitional between pioneering forests once dominated by
Robinia pseudoacacia and one or more of the montane or
basic oak-hickory forests. This vegetation has a prominent
component of invasive exotics such as Alliaria petiolata and
often lacks Liriodendron tulipifera. It is treated here as part
of the broad, USNVC Robinia pseudoacacia successional
forest type.
58
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
III. LOW- TO MIDDLE-ELEVATION DRY AND DRY MESIC FORESTS AND
WOODLANDS
1. Basic Oak-Hickory Forests
a. Quercus rubra – Quercus prinus – Carya ovalis / Cercis
canadensis / Solidago caesia Forest
Central Appalachian Basic Oak-Hickory Forest
(Submontane/Foothills Type)
(USNVC CEGL008514, map code F19, 7 plots)
In the Park, this forest community is primarily associated
with low- to middle-elevation (< 815 m, < 2674’) slopes
underlain by metabasalt. Habitats are more mesic than
those of other low-elevation oak-hickory forests (see
CEGL006216 and CEGL008515 below). Sites are typically
on rocky, middle and upper slopes with intermediate soil
fertility. Southwesterly aspects prevail among plots, which
may represent an artifact of limited sampling. This type is
most common in the western Piedmont and is elevation-
limited on the main Blue Ridge in Virginia. This community
has a very diverse overstory with several Carya spp.,
Quercus spp., Liriodendron tulipifera, and Fraxinus
americana prominent in variable proportions; it is almost
impossible to name the type in a way that conveys this
diversity. Cercis canadensis was very infrequent in plot
samples from the Park, but was more common in AA points
tagged to this type.
b. Quercus alba – Carya glabra – Fraxinus americana /
Cercis canadensis / Muhlenbergia sobolifera – Elymus
hystrix Forest
Northern Hardpan Basic Oak-Hickory Forest
(USNVC CEGL006216, map code F20, 3 plots)
In the Park, this association is known only from upper-slope
benches on Dickey Ridge at elevations from 550 to 690 m
(1804 to 2264 ft). Soils are shallow to bedrock and sites
have substantial cover of loose boulders and stones, along
with patches of exposed mineral soil. Soils extracted from
plots have the highest mean pH, Ca, and Fertility Index
(CEC x TBS/100) among classified types. This community
type is endemic to the Piedmont and Blue Ridge foothills of
northern Virginia and Maryland. It is most characteristic of
diabase flatwoods of Culpeper Basin, where it is associated
with mafic, hardpan soils. Examples also occur on slope
benches of metabasalt foothills such as the Watery
Mountains, where base-rich soils are shallow to bedrock.
59
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Although known in the Park only from Dickey Ridge, it could
also occur on low-elevation metabasalt slopes elsewhere.
Quercus prinus is a common oak in the overstory mix in
these Blue Ridge/foothill stands, but is absent from the
distributional centrum on the Culpeper Basin.
2. Acidic Oak-Hickory Forests
a. Quercus alba – Quercus prinus – Carya glabra / Cornus
florida / Vaccinium pallidum / Carex pensylvanica Forest
Central Appalachian Acidic Oak-Hickory Forest
(USNVC CEGL008515, map code F18, 5 plots)
This forest community is apparently restricted (or nearly so)
in the Park to low-elevation slopes of the metasedimentary
terrain on the Park's western flank. In most cases, geologic
substrate is presumed to be metasiltstone or phyllite of the
Harpers Formation. Strong compositional differences
between this and the two oak/heath types cannot be
explained by topography or soil chemistry, and are assumed
to be related to soil texture, depth, and moisture-holding
capacity. This vegetation probably occupies less fertile sites
on the same shaley soils that support a more montane oak-
hickory forest (see CEGL008516 below) at higher elevations.
The Park represents the eastern margin of its distribution,
which is centered in the shale districts of the Ridge and
Valley province to the west.
3. Montane Mixed Oak and Oak-Hickory Forests
a. Quercus prinus – Quercus rubra / Hamamelis virginiana
Forest
Central Appalachian Dry-Mesic Chestnut Oak – Northern
Red Oak Forest
(USNVC CEGL006057, map code F5, 31 plots)
This association is one of the most extensive vegetation
types in the Park, and is widespread at lower and middle
elevations on all geological substrates. The elevation range
of plot samples is 256 to 1018 m (840 to 3340 ft), mean of
564 m (1850’). It occupies various slope positions, as well
as infertile, slope-base floodplains filled with bouldery
quartzite alluvium. Sites are generally submesic and
frequently very bouldery or stony. Soils have slightly higher
base status than those of the area's oak / heath forests.
Because it is compositionally variable and lies near the
centrum of compositional and environmental gradients, this
association is hard to characterize and has characteristics of
60
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
oak/heath, oak-hickory, and low-elevation boulderfield
forests. It intergrades with many of the other major forest
communities along topographic, moisture, and soil fertility
gradients.
b. Quercus rubra – Quercus alba – Fraxinus americana -
Carya (ovata, ovalis) / Actaea racemosa Forest
Central Appalachian Montane Oak-Hickory Forest (Basic
Type)
(USNVC CEGL008518, map code F16, 18 plots)
This association comprises “rich" oak-hickory forests of
middle to high-elevation ridge crests and gentle upper
slopes, mostly over metabasaltic substrates (one plot each
was located on charnockite and metasiltstone / phyllite).
Elevation range of plot samples is 808 to 1067 m (2651 to
3501 ft), mean of 969 m (3179’). Based on AA data, the
type also appears to occur locally in the 700-800 m elevation
range, where it probably intergrades with the Central
Appalachian Basic Oak-Hickory Forest (CEGL008514; see
above). Soils are apparently deep, usually lack substantial
rock cover, and are intermediate in fertility. This very
distinctive type has an overstory of oaks, hickories, and
white ash, along with a lush, forb-rich herb layer that
resembles that of a rich cove forest. It covers fairly
extensive areas in the Park and grades into Northern Red
Oak Forests (F9) at higher elevations and into other oak and
oak-hickory types at middle elevations.
c. Quercus prinus – Quercus rubra – Carya ovalis /
Solidago (ulmifolia, arguta) – Galium latifolium Forest
Central Appalachian Montane Oak-Hickory Forest (Acidic
Type)
(USNVC CEGL008516, map code F17, 12 plots)
This small- to medium-patch, rather dry but diverse oak-
hickory forest occurs primarily on shaley units (metasiltstone,
phyllite) of the Chilhowee Group at middle elevations in the
southern section. It occurs most frequently on narrow,
stony, convex crests and upper east-facing slopes. The
elevation range of plot samples is 539 to 969 m (1768 to
3179 ft), mean of 800 m (2625’). Surface substrate is
usually somewhat stony and contains locally large areas of
exposed mineral soil. Soil samples had low to intermediate
base status, except for high mean Mn. Overstory
composition of this type is similar to CEGL008514 (in Basic
Oak-Hickory Forests group) but the herbaceous vegetation
61
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
is very different and reflects the drier, more montane
habitats.
4. Oak / Heath Forests
a. Quercus prinus – (Quercus coccinea, Quercus velutina)
/ Kalmia latifolia / Vaccinium pallidum Forest
Central Appalachian / Northern Piedmont Low-Elevation
Chestnut Oak Forest
(USNVC CEGL006299, map code F3, 22 plots)
This association occurs on dry, rocky, infertile slopes at
lower to middle elevations throughout the Park, but is
particularly extensive on the dry, acidic metasedimentary
substrates of the western Blue Ridge flank in the south
district. Soils collected from plots are extremely acidic and
infertile, with high Fe levels. This is the Park’s characteristic
dry-site forest dominated by Quercus prinus and typically an
evergreen shrub layer of Kalmia latifolia.
b. Quercus coccinea – Quercus velutina – Quercus alba /
Amelanchier arborea /
Gaylussacia baccata Forest
Mixed Oak / Heath Forest (Low-Elevation White Oak –
Scarlet Oak – Black Oak Type)
(USNVC CEGL008521, map code F4, 10 plots)
This association forms the principal forest cover on the low-
elevation alluvial fan terrain at the western foot of the Park,
mostly in the south district. The co-dominance of Quercus
alba with other Quercus spp. is distinctive among the two
oak/heath types in the Park, and the shrub layer is typically
dominated by Gaylussacia baccata and other deciduous
ericads. Sites are gentle (0 to 10-degree slope) lower slopes
and flats at very low elevations (< 500 m, < 1640’) at the foot
of the western Blue Ridge flank. Underlying bedrock
(principally shale and limestone of the Waynesboro
Formation) is well covered by deep colluvial and alluvial fan
deposits weathered from upslope Chilhowee Group
quartzite. Soil chemistry is similar to the preceding type.
This community type is most abundant on similar, rolling
terrain of the Piedmont Plateau east of the mountains.
5. Pine-Oak / Heath Woodlands
a. Pinus (pungens, rigida) / Quercus ilicifolia / Gaylussacia
baccata Woodland
Central Appalachian Pine – Oak / Heath Woodland
62
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
(USNVC CEGL004996, map code F1, 8 plots)
This association covers large areas on the metasedimentary
substrates in the southern section of the Park, but is much
less common on the granitic suite and metabasalt
elsewhere. Stands typically occur on south- to west-facing,
convex, upper slopes, ridge crests, and cliff-tops. These
sites are among the most xeric and infertile habitats in the
Park, and most have a demonstrable history of fires. The
most characteristic physiognomic expression is an open
woodland of stunted Quercus prinus, Pinus rigida, and/or
Pinus pungens, with dense shrub thickets of Quercus
ilicifolia and ericads. To a great extent, the dominant pines
of this vegetation require occasional burning for
regeneration, and some stands from which fire has been
absent for long periods have become nearly closed forests.
Because of recent depredations by the southern pine beetle,
existing plot samples from the Park have rather low pine
cover, even though they clearly represent the type.
6. Mountain / Piedmont Basic Woodlands
a. Fraxinus americana – Carya glabra / Muhlenbergia
sobolifera – Helianthus divaricatus – Solidago ulmifolia
Woodland
Central Appalachian Basic Woodland
(USNVC CEGL003683, map code O5, 14 plots)
This low-elevation (475 to 750 m, 1558 to 2461 ft) dry
woodland occurs on metabasalt and other base-rich
substrates of the Central Appalachians, but is known only
from metabasalt in the Park. It is typical of rocky, south- to
west-facing slopes, forming a woodland "matrix" around
exposed cliffs and outcrops. Bedrock and boulder cover in
plot samples averages > 40%. Levels of pH, Ca, Mg, and
TBS in soil samples from plots are among the highest in the
Park. Open, stunted overstories tend to be dominated by
Fraxinus americana and Carya spp.; a diversity of dry-
mesophytic and xerophytic shrubs and herbs is associated.
7. Low-Elevation Boulderfield Forests and Woodlands
a. Quercus prinus – Betula lenta / Parthenocissus
quinquefolia Talus Woodland
Chestnut Oak – Black Birch Wooded Talus Slope
(USNVC CEGL006565, map code F2, 7 plots)
This community is widespread in the Park on boulderfields
and bouldery colluvial slopes weathered from resistant
63
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
quartzites of the Chilhowee Group in the southern section.
Scattered occurrences occupy similar habitats on granitic
terrain and, rarely, metabasalt. This association ranges from
the lowest elevations to ~ 975 m (3200’). Northerly slopes
prevail among plot samples, but this is probably an artifact of
limited sampling. Soils could not be extracted from all plots;
those that could were extremely acidic and infertile, with high
Fe levels. These habitats are extremely difficult to plot-
sample, which is why the type is under-sampled in the Park.
Physiognomy is quite variable, ranging from very open
woodlands of gnarled Betula lenta to more closed stands of
mixed Betula and Quercus prinus. Betula lenta is generally
a pioneer woody invader of open boulderfield edges. As
boulderfields weather and soil material fills the intersticial
spaces, oaks and other species become established. This
type grades into other forest communities (particularly the
chestnut oak - red oak forest [CEGL006057]) along
topographic gradients.
b. Tilia americana – Fraxinus americana / Acer
pensylvanicum – Ostrya virginiana / Parthenocissus
quinquefolia – Impatiens pallida Woodland
Central Appalachian Basic Boulderfield Forest (Montane
Basswood – White Ash Type)
(USNVC CEGL008528, map code F14, 14 plots)
This association is widespread in the Park on boulderfields
and bouldery colluvial slopes weathered from Catoctin
metabasalt and, less frequently, pyroxene-bearing granitic
rocks. It usually occupies steep middle slopes and is
especially extensive in the elevation zone from 762 to 914 m
(2500 to 3000 ft), less commonly extending to 500 m (1640’)
and 1040 m (3412’). Boulder cover in plot samples
averages > 50% and soils have moderately high Ca and Mg.
Distribution of this type is centered in the elevation zone
where Liriodendron tulipifera drops out as a dominant tree of
coves and mesic slopes. Because of constraints imposed by
the Alliance level of the USNVC, this type is formally
classified (in the USNVC) as a woodland, but is better
characterized as an open forest. Tilia americana, Fraxinus
americana, and Quercus rubra are the usual overstory
dominants in varying proportions; shrub and herb-layer
densities, as well as overall species-richness, vary
considerably with the relative abundances of rock cover and
intersticial soil material.
64
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
8. Early Successional Dry and Dry-Mesic Forests
a. Pinus virginiana Successional Forest
Virginia Pine Succesional Forest
(USNVC CEGL002591, *[not mapped], 0 plots)
This association occurs very locally in the Park at lower and
middle elevations. It is a pioneering forest on dry, eroded,
and/or depleted soils of old fields and pastures Most of
these areas are underlain by the more base-rich
metabasaltic and granitic rocks, and are located both on
broad crests and relatively gentle side slopes. This type
accommodates both monospecific stands of Pinus virginiana
and decadent stands in which P. virginiana is co-dominant
with emergent hardwoods. Various Quercus spp., Carya
spp., Liriodendron tulipifera, and Acer rubrum appear to be
the most frequent tree associates; understory and
herbaceous species vary widely with site conditions and
land-use history. This community type was not plot-
sampled, but was documented in several locations by
observations and accuracy assessment procedures.
IV. LOW- TO MIDDLE-ELEVATION ROCK OUTCROPS AND BARRENS
1. Low-Elevation Basic Outcrop Barrens
a. Juniperus virginiana – Fraxinus americana / Carex
pensylvanica – Cheilanthes lanosa Wooded Herbaceous
Vegetation
Central Appalachian Circumneutral Barren
(USNVC CEGL006037, map code O6, 8 plots)
This community type comprises low-elevation (< 580 m, <
1903’) outcrop barrens on Catoctin metabasalt and is
documented in the Park from Dickey Ridge, lower Overall
Run, Cedar Run, and Goat Ridge. Sites are typically on
steep (~27 degree) middle slopes with south or west
aspects. Surface cover of bedrock and loose rocks in plots
averages > 60%. Mean pH, Ca, Mg, and TBS levels in soil
samples were among the highest of classified vegetation
types in the Park. Stands usually occur in small patches and
have dense patches of graminoids (e.g., Schizachyrium
scoparium, Carex pensylvanica, Bouteloua curtipendula) and
xerophytic forbs; widely scattered, stunted trees and shrubs
are intermingled. Although confined to metabasalt in the
Park, this association also occupies sites underlain by
65
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
calcareous shales and sandstones in the adjacent Ridge and
Valley province.
b. Fraxinus americana / Physocarpus opulifolius / Carex
pensylvanica – Allium cernuum – (Phacelia dubia)
Wooded Herbaceous Vegetation
Central Appalachian Mafic Barren (Ninebark / Pennsylvania
Sedge Type)
(USNVC CEGL008529, map code O7, 7 plots)
This is the "middle-elevation" (~550 to 1036 m, 1804 to 3400
ft) rock outcrop barren of metabasalt and pyroxene-bearing
granites in SHNP. Most sites are on steep (mean = 30-
degree), westerly, middle to upper-slope, convex outcrops.
Surface cover of bedrock and loose rocks in plots averages
> 60%. Soil samples have intermediate base status.
Physiognomy of stands is similar to the preceding type.
Compositionally, this association lacks many typical low-
elevation species present in CEGL006037 and contains a
number of distinctly montane species.
V. ALLUVIAL FLOODPLAIN COMMUNITIES
1. Piedmont / Low Mountain Alluvial Forests
a. Liriodendron tulipifera – Platanus occidentalis – Betula
lenta / Lindera benzoin / Circaea lutetiana ssp.
canadensis Forest
Northern Blue Ridge Montane Alluvial Forest
(USNVC CEGL006255, map code F11, 13 plots)
This forest community is apparently confined to the larger,
mountain-foot floodplains with relatively fertile alluvial
deposits. Habitats are nearly flat, bouldery, and well-
drained, with moderately fertile soils derived from
metabasalt, pyroxene-rich granites, or metasiltstone/phyllite.
Many of these sites were probably cleared and subjected to
multiple disturbances during the historical period in which the
Park area was heavily populated. This type does not occur
on sterile, acidic alluvium derived from the Chilhowee Group
and deposited in floodplains at the foot of the western flank.
CEGL006255 is a new USNVC type based on plot data from
the Park and other qualitative data.
66
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
IV. NON-ALLUVIAL WETLANDS
1. Mafic Fens and Seeps
a. Spiraea alba var. latifolia – Cornus racemosa /
Calamagrostis canadensis – Sanguisorba canadensis –
Carex scoparia Shrub Herbaceous Vegetation
Northern Blue Ridge Mafic Fen
(USNVC CEGL006249, map code W1, 4 plots)
This community appears to be endemic to Shenandoah
National Park, where it is confined to groundwater-saturated,
high-elevation stream-head wetlands in the vicinity of Big
Meadows on both sides of Skyline Drive. All stands have
been disturbed by hydrologic alterations, excessive deer
grazing, and probably fire exclusion. It is similar to another
shrubland of mafic seeps in the southern Blue Ridge of VA
and NC. CEGL006249 is a new USNVC type defined to
cover this vegetation. It is a demonstrably rare vegetation
type of considerable conservation concern.
2. Woodland Seeps
a. Caltha palustris – Impatiens capensis – Viola cucullata
Herbaceous Vegetation [Provisional]
Central Appalachian Woodland Seep
(USNVC CEGL006258, map code W3, 2 plots) This
provisional vegetation type is represented by two plot
samples located in narrow, groundwater-saturated seeps
that are closely bordered by upland vegetation. The two
samples are quite different and do not form a strongly
homogeneous group. Similar habitats and vegetation are
scattered throughout the Park. CEGL006258 is a new
USNVC type defined to cover this vegetation. More
inventory and data collection in the Central Appalachians is
needed to make this classification unit more robust. Similar
communities from the southern Blue Ridge have been
classified. At least one of the occurrences in the Park is
much smaller than the minimum mapping unit size (0.5 ha)
3. Mountain / Piedmont Acidic Seepage Swamps
a. Acer rubrum – Nyssa sylvatica / Ilex verticillata –
Vaccinium fuscatum / Osmunda cinnamomea Forest
Central Appalachian Acidic Seepage Swamp
(USNVC CEGL007853, map code W2, 3 plots)
67
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
This saturated forest community occurs along headwaters
streams on the acidic, metasedimentary terrain of the
western flank. All known examples are at very low
elevations (< 500 m, < 1640’) on ancient alluvial fans
bordering the Shenandoah Valley. Habitats typically feature
braided streams with Sphagnum-covered hummocks. Soils
are extremely acidic and infertile, with high Fe levels. The
few known stands in the Park conform closely to the USNVC
description, although plot SHNP632 is a marginal, somewhat
disturbed example.
4. Mountain / Piedmont Basic Seepage Swamps
a. Acer rubrum – Fraxinus americana – Fraxinus nigra –
Liriodendron tulipifera / Carex bromoides – Caltha
palustris Forest
Central Appalachian Basic Seepage Swamp
(USNVC CEGL008416, map code W4, 8 plots)
Lower- to middle-elevation forests occurring in linear patches
along groundwater-saturated bottoms of streams and in
headwaters seepage areas. Plot-sampled sites range from
421 to 939 m (1381 to 3081 ft) elevation, mean of 832 m
(2730’), and are confined to substrates weathered from
metabasalt and base-rich granites. Habitats are generally
very bouldery and gravelly, with pronounced hummock-and-
hollow microtopography and braided streams. Soils
collected from plots have relatively high pH, Ca, Mg, Fe, and
TBS levels. This type clearly represents a basic forested
seepage wetland type conceptually similar to CEGL008416,
but exhibits compositional variation related to topography
(particularly increased importance of Betula alleghaniensis
and Tsuga canadensis as elevation increases). At middle
elevations, it grades into the High-Elevation Seepage
Swamp type (CEGL008533) below, and several plots could
be assigned almost equally well to either type. Fraxinus
nigra, which is considered "diagnostic" of these wetlands,
reaches its southern limits in VA and is quite sporadic in the
Park (it is present in only half of the plots). Similar
topographic gradation, as well as gradation apparently
related to soil chemistry, has been noted on a regional level.
68
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
5. High-Elevation Seepage Swamps
a. Tsuga canadensis – Betula alleghaniensis / Veratrum
viride – Carex scabrata – Oclemena acuminata Forest
High-Elevation Hemlock – Yellow Birch Seepage Swamp
(USNVC CEGL008533, map code W5, 8 plots)
Middle- to high-elevation forests occurring in linear patches
along groundwater-saturated bottoms of streams and in
headwaters seepage areas. Plot-sampled sites range from
670 to 1036 m (2198 to 3400 ft) elevation (most are > 900 m,
> 2953’) and occur on all major substrate types. Habitats
are generally less rocky than those of the Mountain /
Piedmont basic seepage swamp (CEGL008416) above, but
have similar hummock-and-hollow microtopography and
braided streams. Soils have low to intermediate base status.
Stands of this community are undergoing physiognomic and
compositional changes due to extensive, adelgid-related
mortality of Tsuga canadensis.
6. Shenandoah Valley Sinkhole Ponds
a. Quercus palustris / Panicum rigidulum var. rigidulum –
Panicum verrucosum – Eleocharis acicularis
Herbaceous Vegetation
Shenandoah Valley Sinkhole Pond (Typic Type)
(USNVC CEGL007858, map code W6, 1 plot)
The Park boundary runs through a single small but
representative example of a Shenandoah Valley sinkhole
pond. The habitat is seasonally flooded, with aluminum-rich
clay soils. Many stands of this community type occur just
outside the western Park boundary in Augusta, Rockingham,
and Page Counties. They occupy seasonal ponds
developed on the massive alluvial fans deposited along the
base of the Blue Ridge over former karst terrain. This
association is endemic to a three-county area of the
Shenandoah Valley in Virginia.
69
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
3.3.2 Relationship of Vegetation to Field-Measured Environmental Variables:
The relationship between compositional groups and environmental gradients was
examined in a series of NMDS ordinations (Appendix 1). Important topographic /
hydrologic and soil chemistry gradients are identified by joint plot vector overlays
on each ordination diagram. Each environmental factor with a Pearson product-
moment correlation of r >|0.45| with stand scores on any of the axes is plotted as
a vector, the direction of which indicated the direction of maximum correlation
through ordination space. Vector line lengths are determined by the strength of
the correlation. This critical value for the correlation coefficient is provided in the
caption for each ordination diagram. Significance levels are uncorrected for
multiple comparisons. Joint plots differ from biplots in that vectors emanate from
the centroid of ordination space rather than the origin of the axes, and vectors
are based on correlations instead of least-squares regression equations
(McCune and Mefford 1999). Because vectors in PC-ORD joint plots are not
scaled, the strengths of environmental gradients are not comparable between
ordination diagrams.
When stand distributions in the various ordination diagrams are examined, the
disposition of major vegetation groups and community types generally
corresponds well with clusters identified by the Lance-Williams flexible beta
method. As a whole, the ordinations indicate that bedrock parent material, soil
fertility, elevation, and topographic position are the most important, interrelated
environmental factors influencing major vegetation patterns in Shenandoah
National Park. TRMI (site moisture potential), slope shape, slope inclination,
Beer’s-transformed aspect, and surface substrate characteristics are less
important correlates of major vegetation patterns, although they frequently
characterize differences between community types within the major groups
(Appendix 2).
The results of these ordination studies suggest that bedrock stratigraphy in
Shenandoah National Park exerts strong topographic control that results in more
or less regularly recurring landforms, as well as somewhat predictable variation
in site moisture and soil chemistry.
3.3.3 Summary of Riparian and Wetland Zone Vegetation:
Riparian and wetland vegetation is very limited in Shenandoah National Park,
covering approximately 1632 hectares (4,032 acres), or 2% of the park.
Vegetation generally falls into three hydrologic classes of Cowardin et.al (1979):
temporarily flooded forests, saturated forests and shrublands, and seasonally
flooded herbaceous vegetation.
Narrow but definite alluvial floodplains along larger streams at the lowest
elevations support both palustrine and terrestrial forests. Hydrologic regime is
probably similar to that reported for Passage and Mill Creeks in the Massanutten
70
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Mountains (Shenandoah County), where small-scale alluvial landforms greatly
influence species distributions (Hupp 1982, Hupp 1986, Olson and Hupp 1986).
The lowest terraces of these narrow bottoms are probably flooded briefly once a
year, while higher terraces are inundated only during rare, catastrophic floods.
Microtopography is complex and usually includes multiple floodplain terraces,
fans at tributary hollow mouths, various bank features, bouldery and cobbly
depositional bars, and coarse woody debris transported by floods. Soils are
highly variable, lack profile development, and range from well-drained to poorly
drained. Along larger streams that drain large areas of granitic and greenstone
terrain, soil fertility is typically high and at least some obligate wetland species
are present. The Northern Blue Ridge Montane Alluvial Forest (map unit F11) is
the principal community type on these more fertile, moisture-holding floodplains,
which are most extensive at the foot of the eastern slope, and the foot of the
western slope in the North District.
Similar floodplains on the western flank of the park that drain large areas of
Chilhowee Group metasedimentary terrain have extremely acidic, nutrient-poor
soils that appear to be much more drought-prone despite their low topographic
position. These sites typically support variants of upland forests in the park,
principally the Central Appalachian Dry-Mesic Chestnut Oak – Northern Red Oak
Forest (map unit F5) or the Central Appalachian Acidic Cove Forest (White Pine
– Mixed Hardwoods Type; map unit F12), and obligate wetland species are
lacking.
The park has a great diversity of seepage wetland vegetation that also exhibits a
strong correlation with substrate conditions. Narrow spring runs, mostly shaded
by trees rooted in adjacent upland forests, are scattered throughout the park on
granitic and metabasaltic substrates, supporting the herbaceous Central
Appalachian Woodland Seep community type. Along larger headwaters streams
where spring runs coalesce in larger, braided bottoms or where lateral
groundwater seepage is abundant, forested wetlands are characteristic. Three
associations, separated by soil preferences and elevation, are present in the
Park: the Central Appalachian Acidic Seepage Swamp (map unit W3) on low-
elevation metasedimentary substrates; the Central Appalachian Basic Seepage
Swamp (map unit W4) on low- to middle-elevation granitic and metabasaltic
substrates; and the High-Elevation Hemlock-Yellow Birch Seepage Swamp (map
unit W5) at elevations mostly above 900 m (occasionally lower in sheltered
habitats).
The most noteworthy seepage wetland is the Northern Blue Ridge Mafic Fen
(map unit W1), which appears to be endemic to high elevations of the Central
District near Big Meadows. Physiognomically, this vegetation type is a mosaic of
shrubs and herbaceous openings, and several state-rare plant species are
characteristic. The type is strongly affiliated with unconsolidated soils, washed
from Catoctin metabasalt, that are highly acidic and rather low in base status
71
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
excepting high Mg, Fe, and Al. Complete descriptions of each wetland
community type are provided in Appendix 2.
3.3.4 Relationship of Vegetation to Ecological Land Units
Plots sampled fell in 83 different ecological land units as defined by GIS and
described in sections 2.1.1 and 3.1. Of the 305 field sampling plots used in
classification, 111 (36%) were placed in just 8 ELU types; mid elevation-basaltic-
steep slopes (n=29), mid elevation-basaltic-S/SW facing side slopes (n=16), high
elevation-basaltic- upper slopes (n=14), high elevation-basaltic-S/SW facing side
slopes (n=11), mid elevation-siliciclastic-steep slopes (n=11), high elevation-
basaltic-flats (n=10), high elevation-basaltic-N/NE facing side slopes (n=10), and
low elevation-basaltic-N/NE facing side slopes (n=10). By contrast, 194 (64%) of
field plots were placed in the remaining 75 ecological land units. A total of 20
field plots were placed in riparian or wetland ELU types (e.g. stream side, slope
bottom, or wetland landform types). Of the 163 ELU types potentially occurring
in the park (section 3.1), 83 types (51%) were sampled in the field.
Vegetation communities occurred in as few as one ELU type or across as many
as 31 ELU types (Figure 3.4). Average number of ELUs per vegetation
community was 8.9. In general, vegetation communities occurred across a
variable number of ecological land unit types as expected from community
analysis. Communities described from vegetation plot compositional analysis as
being generalist (e.g. F5, Central Appalachian Dry-Mesic Chestnut Oak -
Northern Red Oak Forest) occurred across numerous ecological land units
(n=31), while environmentally restricted communities (e.g. F20, Northern
Hardpan Basic Oak - Hickory Forest) occurred across only 3 ecological land
units.
72
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
0
5
10
15
20
25
30
35
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
O1
O2
O3
O4
O5
O6
O7
W1
W2
W3
W4
W5
W6
Community type
Count of ELU
Figure 3.4 Number of ecological land units (ELU) per vegetation community type, from vegetation
training data set (n=305).
3.4 Vegetation Map
The output from the CLDA vegetation mapping analysis included vegetation
maps for Shenandoah National Park derived from three image sets; the spring
AVIRIS image mosaic, the summer AVIRIS image mosaic, and the multitemporal
Landsat dataset. In order to create the most accurate final map with complete
spatial coverage for Shenandoah National Park, results from these three maps
were merged using a rule-based algorithm and the posterior probability maps.
Before merging, a cloud mask was developed and applied to the summer AVIRIS
image mosaic map to remove areas of bright cloud and dark shadow. Once
these areas were excluded, the merging algorithm selected the class at each
pixel that had the highest posterior probability from between the two AVIRIS
maps. In the case of a tie between the two posterior probabilities, the spring
AVIRIS map class was selected, as that image model had a higher overall
accuracy and it was acquired prior to a large fire that is evident in the summer
image. The less accurate Landsat map was used to fill in small areas at the
edge of the park boundary that were not covered by either AVIRIS image mosaic.
73
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
The final merged vegetation class map used results from about half of each
AVIRIS vegetation map. The final merged image products include: a final
vegetation class map (Figures 3.5a, 3.5b, and 3.5c), an image number map that
indicates source image for final class assigned (1=spring AV, 2=summer AV,
3=Landsat)(Figure 3.6), and a map showing maximum probability assignments
(Figure 3.7). All final map products are in ArcInfo GRID format, 17 m pixel
resolution, and projected in Universal Transverse Mercator, Zone 17,North
American Datum of 1983 projection.
Figure 3.8 and Table 3.1 list vegetation mapping results by area. Predominant
map units by area are Liriodendron tulipifera-Fraxinus americana-Tilia Americana
/ Lindera benzoin/Cimicifuga racemosa / Laportea canadensis “Rich Cove and
Slope Forests” (Map code F10), Quercus montana / Kalmia latifolia / Vaccinium
pallidum (Map code F3), and Quercus Montana – Quercus rubra / Cornus florida
/ Viburnum acerifolium (Map code F5) “Oak/Heath Forests”. Other prominent
types are Tilia americana - Fraxinus americana / Acer pensylvanicum - Ostrya
virginiana / Parthenocissus quinquefolia - Impatiens pallida (Map code F14) “Low
Elevation Boulderfield Forests and Woodlands”, Quercus rubra - Quercus alba -
Carya (ovata, ovalis) / Ageratina altissima - Cimicifuga racemosa (Map code
F16) “Montane Oak-Hickory Forests”, and Fraxinus americana - Carya glabra /
Muhlenbergia sobolifera - Helianthus divaricatus - Phacelia dubia Map code O5)
“Mountain/Piedmont Basic Woodlands”.
74
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.5a Final vegetation map result, version 1.1, North District, Shenandoah National Park,
with location of training plots. Refer to figure 3.5d for legend to vegetation community types.
75
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.5b Final vegetation map result, version 1.1, Central District, Shenandoah
National Park, with location of training plots. Refer to figure 3.5d for legend to vegetation
community types.
76
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.5c Final vegetation map result, version 1.1, South District, Shenandoah National Park,
with location of training plots. Refer to figure 3.5d for legend to vegetation community types.
77
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Legend
F1 - Central Appalachian Pine - Oak / Heath Woodland
F2 - Chestnut Oak - Black Birch Wooded Talus Slope
F3 - Central Appalachian / Northern Piedmont Low-Elevation Chestnut Oak Forest
F4 - Mixed Oak / Heath Forest (Piedmont / Central Appalachian Low-Elevation Type)
F5 - Central Appalachian Dry-Mesic Chestnut Oak - Northern Red Oak Forest
F6 - Mid-Atlantic Mesic Mixed Hardwood Forest
F7 - Central Appalachian Northern Hardwood Forest (Yellow Birch - Northern Red Oak Type
)
F8 - Hemlock - Northern Hardwood Forest
F9 - Northern Red Oak Forest (Pennsylvania Sedge - Wavy Hairgrass Type)
F10 - Southern Appalachian Cove Forest (Typic Montane Type)
F11 - Northern Blue Ridge Montane Alluvial Forest
F12 - Central Appalachian Acidic Cove Forest (White Pine - Mixed Hardwoods Type)
F13 - Successional Tuliptree Forest (Circumneutral Type)
F14 - Central Appalachian Basic Boulderfield Forest (Montane Basswood - White Ash Type)
F15 - Central Appalachian Rich Cove Forest
F16 - Central Appalachian Montane Oak - Hickory Forest (Basic Type)
F17 - Central Appalachian Montane Oak - Hickory Forest (Acidic Type)
F18 - Central Appalachian Acidic Oak - Hickory Forest
F19 - Central Appalachian Basic Oak - Hickory Forest (Submontane / Foothills Type)
F20 - Northern Hardpan Basic Oak - Hickory Forest
F21 - Black Locust Successional Forest
O1 - High-Elevation Greenstone Barren
O2 - High-Elevation Acidic Heath Barren / Pavement
O3 - High-Elevation Outcrop Barren (Black Chokeberry Igneous / Metamorphic Type)
O4 - Central Appalachian High-Elevation Boulderfield Forest
O5 - Central Appalachian Basic Woodland
O6 - Central Appalachian Circumneutral Barren
O7 - Central Appalachian Mafic Barren (Ninebark / Pennsylvania Sedge Type)
W1 - Northern Blue Ridge Mafic Fen
W2 - Central Appalachian Acidic Seepage Swamp
W3 - Central Appalachian Woodland Seep
W4 - Central Appalachian Basic Seepage Swamp
W5 - High-Elevation Hemlock - Yellow Birch Seepage Swamp
W6 - Shenandoah Valley Sinkhole Pond (Typic Type)
Figure 3.5d. Map legend for final vegetation map showing vegetation community groups. Refer
to pp. 37-48 for community descriptions.
78
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.6 Image sources used as inputs to final vegetation mapping result.
79
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Figure 3.7 Final class probabilities by pixel for vegetation mapping result. Higher probabilities
(blue) indicate strong class associations.
80
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
0
2,000
4,000
6,000
8,000
10,000
12,000
14,000
16,000
F1
F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14
F15
F16
F17
F18
F19
F20
F21
O1
O2
O3
O4
O5
O6
O7
W1
W2
W3
W4
W5
W6
Community type
Hectares
Figure 3.8. Mapped vegetation communities of Shenandoah National Park, by area.
81
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Table 3.1. Mapped vegetation communities of Shenandoah National Park, by area.
Map
Code
Common Association Name Hectares Percent
Area
Cumulative
Area
F3 Central Appalachian / Northern Piedmont Low-Elevation
Chestnut Oak Forest
14,895 18.80% 18.8%
F10 Southern Appalachian Cove Forest (Typic Montane Type) 14,188 17.90% 36.7%
F5 Central Appalachian Dry-Mesic Chestnut Oak - Northern Red
Oak Forest
10,405 13.13% 49.8%
F14 Central Appalachian Basic Boulderfield Forest (Montane
Basswood - White Ash Type)
6,237 7.87% 57.7%
F16 Central Appalachian Montane Oak - Hickory Forest (Basic
Type)
5,883 7.42% 65.1%
F11 Northern Blue Ridge Montane Alluvial Forest 4,162 5.25% 70.4%
F9 Northern Red Oak Forest (Pennsylvania Sedge - Wavy
Hairgrass Type)
3,671 4.63% 75.0%
O5 Central Appalachian Basic Woodland 2,608 3.29% 78.3%
F21 Black Locust Successional Forest 2,039 2.57% 80.9%
F2 Chestnut Oak - Black Birch Wooded Talus Slope 1,925 2.43% 83.3%
F19 Central Appalachian Basic Oak - Hickory Forest (Submontane
/ Foothills Type)
1,862 2.35% 85.7%
F1 Central Appalachian Pine - Oak / Heath Woodland 1,749 2.21% 87.9%
F4 Mixed Oak / Heath Forest (Piedmont / Central Appalachian
Low-Elevation Type)
1,700 2.15% 90.0%
F17 Central Appalachian Montane Oak - Hickory Forest (Acidic
Type)
1,342 1.69% 91.7%
F13 Successional Tuliptree Forest (Circumneutral Type) 1,008 1.27% 93.0%
F12 Central Appalachian Acidic Cove Forest (White Pine - Mixed
Hardwoods Type)
968 1.22% 94.2%
F8 Hemlock - Northern Hardwood Forest 935 1.18% 95.4%
O5 Central Appalachian Basic Woodland 567 0.72% 96.1%
O1 High-Elevation Greenstone Barren 541 0.68% 96.8%
O7 Central Appalachian Mafic Barren (Ninebark / Pennsylvania
Sedge Type)
420 0.53% 97.3%
F15 Central Appalachian Rich Cove Forest 370 0.47% 97.8%
O2 High-Elevation Acidic Heath Barren / Pavement 359 0.45% 98.2%
O6 Central Appalachian Circumneutral Barren 300 0.38% 98.6%
F18 Central Appalachian Acidic Oak - Hickory Forest 256 0.32% 98.9%
W1 Northern Blue Ridge Mafic Fen 201 0.25% 99.2%
F20 Northern Hardpan Basic Oak - Hickory Forest 182 0.23% 99.4%
F7 Central Appalachian Northern Hardwood Forest (Yellow Birch -
Northern Red Oak Type)
139 0.18% 99.6%
W4 Central Appalachian Basic Seepage Swamp 117 0.15% 99.7%
W2 Central Appalachian Acidic Seepage Swamp 113 0.14% 99.9%
O4 Central Appalachian High-Elevation Boulderfield Forest 43 0.05% 99.9%
W3 Central Appalachian Woodland Seep 29 0.04% 100.0%
F6 Mid-Atlantic Mesic Mixed Hardwood Forest 12 0.02% 100.0%
O3 High-Elevation Outcrop Barren (Black Chokeberry Igneous /
Metamorphic Type)
11 0.01% 100.0%
W6 Shenandoah Valley Sinkhole Pond (Typic Type) 5 0.01% 100.0%
82
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
3.5 Accuracy Assessment
3.5.1 Internal Accuracy Assessment:
The trio of CLDA vegetation mapping models from the AVIRIS and Landsat data
produced variable success at predicting the assigned vegetation class of the
training set. Spring AVIRIS predicted the correct vegetation type out of 33
classes for 83% (88% percent plots correct, or PPC) of 1060 spectral samples.
Summer AVIRIS predicted the correct vegetation association out of 32 classes
for 82% (87% PPC) of the 932 training sample cases. Multitemporal Landsat TM
predicted the correct vegetation type 57% of the time (64% PPC) from 1196
samples. Overall accuracies for the two AVIRIS maps are remarkably similar
and surpass the desired 80% threshold, though different classes are best
predicted on different maps. The Landsat model accuracy around 60% is similar
to accuracy achieved for Landsat data in the literature for such specific and
detailed vegetation classes. Landsat is probably less powerful at discriminating
detailed vegetation classes because it has fewer, broader spectral bands than
the AVIRIS hyperspectral imagery and its pixel size of 30m is less compatible
with the field sampling protocol used in this study than the 17m AVIRIS image
pixel size.
The accuracy of the final merged map for Shenandoah National Park was
assessed internally by sampling the final map with the training sample locations
and computing the percent plots correct (PPC) for each class and the entire map
(Appendix 6, Table A.6.1). The overall PPC accuracy of the merged map was
88%. This accuracy was higher than the overall accuracies of the individual
image models because only the best predicted areas from each map were
incorporated into the final merged product. Five of the association level classes
were mapped with a training set PPC lower than 80% accuracy.
3.5.2 Field validated accuracy assessment:
Once all the possible map classes were recorded from the final map, all possible
map classes were compared to the first (presumably most likely) recorded field
class for each validation site. We tabulated results for accuracy assessment
separately for 2004 and 2005 field surveys since the protocols and field crews
varied between campaigns. We also report results from a “paper” accuracy
assessment analysis using other vegetation field plots collected for long-term
ecological monitoring (LTEM) and plant protection activities. The results of this
analysis are shown in Appendix 6. Overall field validation accuracy for the 2004
accuracy assessment campaign was 64% (Appendix 6, Table A.6.2). By-class
accuracy ranged from a high of 100% to a low of 0%. However, the number of
validation plots in some classes was low. Overall field validation accuracy from
the 2005 accuracy assessment campaign was 51% (Appendix 6, Table A.6.3),
again reflecting low sample sizes. By-class accuracy ranged from a high of 100%
to a low of 0%. Overall accuracy reported from a “paper” accuracy assessment
83
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
using previously collected plot data was 51% (Appendix 6, Table A.6.4). Once
again, by-class accuracy ranged from a high of 100% to a low of 0%.
While overall field validated accuracy appears low, a tabulation of accuracy by
area using the 2004 accuracy assessment campaign data reveals that 80% of
the park by area is mapped at an 80% accuracy level, and 95% of the park by
area is mapped at 64% accuracy level (Table 3.2).
84
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
Table 3.2 Cumulative accuracy by cumulative park area. Vegetation associations mapped in
Shenandoah National Park in this effort are arranged in descending order by park area.
Accuracy figures are from 2004 field accuracy assessment.
Map
Code
Common Association Name Hectares Cumulative
Area
Accuracy
2004 AA
Cumulative
Accuracy
F3 Central Appalachian / Northern Piedmont Low-
Elevation Chestnut Oak Forest
14,895 18.8% 83.0%
F10 Southern Appalachian Cove Forest (Typic Montane
Type)
14,188 36.7% 89.0%
F5 Central Appalachian Dry-Mesic Chestnut Oak -
Northern Red Oak Forest
10,405 49.8% 70.0%
F14 Central Appalachian Basic Boulderfield Forest
(Montane Basswood - White Ash Type)
6,237 57.7% 60.0%
F16 Central Appalachian Montane Oak - Hickory Forest
(Basic Type)
5,883 65.1% 82.0%
F11 Northern Blue Ridge Montane Alluvial Forest 4,162 70.4% 100.0%
F9 Northern Red Oak Forest (Pennsylvania Sedge -
Wavy Hairgrass Type)
3,671 75.0% 100.0% 83.4%
O5 Central Appalachian Basic Woodland 2,608 78.3%
F21 Black Locust Successional Forest 2,039 80.9% 56.0% 80.0%
F2 Chestnut Oak - Black Birch Wooded Talus Slope 1,925 83.3% 0.0%
F19 Central Appalachian Basic Oak - Hickory Forest
(Submontane / Foothills Type)
1,862 85.7% 55.0% 69.5%
F1 Central Appalachian Pine - Oak / Heath Woodland 1,749 87.9% 68.0%
F4 Mixed Oak / Heath Forest (Piedmont / Central
Appalachian Low-Elevation Type)
1,700 90.0% 50.0% 67.8%
F17 Central Appalachian Montane Oak - Hickory Forest
(Acidic Type)
1,342 91.7% 60.0%
F13 Successional Tuliptree Forest (Circumneutral Type) 1,008 93.0% 32.0%
F12 Central Appalachian Acidic Cove Forest (White
Pine - Mixed Hardwoods Type)
968 94.2% 14.0%
F8 Hemlock - Northern Hardwood Forest 935 95.4% 100.0% 63.7%
W5 Central Appalachian Basic Woodland 567 96.1%
O1 High-Elevation Greenstone Barren 541 96.8%
O7 Central Appalachian Mafic Barren (Ninebark /
Pennsylvania Sedge Type)
420 97.3%
F15 Central Appalachian Rich Cove Forest 370 97.8% 50.0%
O2 High-Elevation Acidic Heath Barren / Pavement 359 98.2%
O6 Central Appalachian Circumneutral Barren 300 98.6%
F18 Central Appalachian Acidic Oak - Hickory Forest 256 98.9% 100.0%
W1 Northern Blue Ridge Mafic Fen 201 99.2%
F20 Northern Hardpan Basic Oak - Hickory Forest 182 99.4%
F7 Central Appalachian Northern Hardwood Forest
(Yellow Birch - Northern Red Oak Type)
139 99.6% 40.0%
W4 Central Appalachian Basic Seepage Swamp 117 99.7% 0.0%
W2 Central Appalachian Acidic Seepage Swamp 113 99.9%
O4 Central Appalachian High-Elevation Boulderfield
Forest
43 99.9%
W3 Central Appalachian Woodland Seep 29 100.0%
F6 Mid-Atlantic Mesic Mixed Hardwood Forest 12 100.0%
O3 High-Elevation Outcrop Barren (Black Chokeberry
Igneous / Metamorphic Type)
11 100.0%
W6 Shenandoah Valley Sinkhole Pond (Typic Type) 5 100.0%
85
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
4. Discussion/Conclusion
The results of this project demonstrate innovative application of the latest
techniques in vegetation mapping using hyperspectral imaging and landscape
modeling. Mapping results are generally consistent with previous mapping
efforts (by area) and show that the USNVC can be reliably mapped with
techniques other than manual interpretation of aerial photography. In fact, by
exploiting the fine spectral specificity of hyperspectral imaging and the spatial
heterogeneity of environmental gradient models, much more information can
potentially be extracted on growing environments than can be visually
interpreted. The fact that Association-level classes can be mapped using these
techniques is an important finding.
However, the modeling techniques listed here do have limitations, and the field
accuracy assessment demonstrates that these approaches may not always be
amenable to the same type of error assessment as when using traditional aerial
photography and polygon-based photo interpretation. For example, the internal
accuracy assessment of the vegetation models agrees extremely well with the
field data classified into vegetation communities. In other words the model fits
the training data. However, the field validation results do not fit nearly as well to
the model results. This discrepancy can be interpreted in several ways. One
possibility is that the training data are biased towards natural communities, and
these are over represented in the resulting models at the cost of accurately
representing the amount of disturbed or successional forests in the park. For
instance, class F21 was sampled somewhat rarely in the original vegetation
sampling (i.e. we did not sample a lot of disturbed Ash-Black Locust forest), but
this class turned out to be much more common in the validation sample. The
map reflects the original field data, so this class is probably underrepresented,
and is obviously a lot more common than the original field sample would suggest.
A second possibility is that the accuracy assessment field data do not accurately
account for the intergraded nature of the forest communities, and may miss the
“correct” community type in a nearby pixel. Although we attempted to account for
this in the accuracy assessment evaluation by looking at the vegetation class of
the majority of pixels in a neighborhood around the sampled pixel, the fact that
multiple possible vegetation community types were recorded for many of the field
plots made this difficult to properly address. In addition, the sample size of the
accuracy assessment field set is small relative to the size of the park and the
number of classes mapped. The distribution of accuracy assessment points was
also based on a draft map using an earlier and less accurate set of assumptions
on vegetation community distribution. Therefore, some classes were under-
sampled in the final accuracy assessment, and some classes were not sampled
at all.
Potential solutions to the low accuracy reported for some classes are to collapse
classes into higher levels of the USNVC hierarchy, add additional accuracy
86
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
assessment sample points, or to revise mapping methodologies. First, collapsing
classes from association into alliance or higher level categories has appeal in
that classes of similar floristic composition can be combined to minimize
commission errors (e.g. incorrect assignment of classes when compared to
ground truth). However, the classification scheme derived for SHEN (Appendix
2) defines 34 associations, 31 alliances, and 14 formations. Collapsing classes
to the alliance level does not result in significant reduction in specificity between
classes, and collapsing classes to the formation level may be so general that
valuable information is lost. Clearly some intermediate level of reclassification is
warranted.
Secondly, it would be possible to increase statistical confidence in accuracy
assessment by increasing the number of samples for each class. Guidelines for
the USGS-NPS vegetation mapping program call for a minimum of 30 samples
per class (Mike Story, NPS, pers. comm.). For the 34 vegetation communities
mapped in Shenandoah NP, this would mean sampling 1020 plots for accuracy
assessment. While a noble goal, we found plot sampling to the detail required to
assess association-level communities in this park to be exceedingly difficult. Two
years of effort at field accuracy assessment yielded only 292 sampled plots.
Logistical difficulties in navigating to plots often resulted in sampling only 1-2
plots per day, and additional time was required at the plots for the field crews to
key vegetation into their closest association. Additional samples could be
gleaned from air-photo interpretation, but confidently interpreting associations
from aerial images is very difficult. This problem needs additional investigation
for a satisfactory result in heavily vegetated (and heavily disturbed) mixed-
deciduous forest dominated park units of the eastern U.S.
Lastly, mapping methods could be revised to re-assess and re-combine
individual class probability maps for better class representation. Some
communities classified from plot vegetation sampling are small patch types or are
exceedingly rare in the park. While we attempted to map all classes present in
the training data set (i.e. all types classified from vegetation plot sampling), some
types may need to be eliminated as too small to map. For these classes,
notations can be made of their presence in the surrounding matrix forest
community. Additionally, it may be possible to restrict distributions of over-
mapped communities by manipulating maps of predicted class probability using
ecological land unit maps and knowledge of vegetation occurrence. For
example, alluvial forest communities should not occur on upper hill slopes, and
therefore it may be appropriate to minimize probabilities for this class using GIS
operations within hill slope ecological land unit types.
Environmental gradient models helped to place the vegetation communities in
context of the physical environment, although not always with the desired
specificity. Figure 4.1 illustrates that a number of community types occur across
a broad range of ecological land unit types. This result may mean that certain
communities (e.g. the Oak/Heath types F5 and F3) are generalists, or that the
87
USGS-NPS Vegetation Mapping Program
Shenandoah National Park
models are not incorporating the fine scale information on micro climate or soil
properties necessary to distinguish the habitat preferences of certain
communities. However, environmental gradient models (incorporated as
topographic variables) were selected as significant discriminating variables in the
CLDA models in both the AVIRIS and Landsat-based classifications. The
information content of the image bands clearly overwhelmed the topographic
variables in relative importance in the AVIRIS-based models (Figure 2.3), but the
topographic variables added significant information to the Landsat-based multi-
temporal models (Figure 2.4).
The greatest benefit of this approach, however, is the nature of the processing
itself. Since the models are quantitative and digital, the input parameters can be
reset and reprocessed, allowing for fine tuning of the output products. This ability
to quickly tweak and re-run the models will allow for future revision of the map
products to meet the needs of the park and national map accuracy standards, or
to incorporate new variables in the modeling process. It is our hope that the
results of this project will be revised and updated as new information becomes
available, or as issues with the initial mapping are discovered. We att