banner literature about projects tools meetings search links BBS GAP home GAP home
home

 

 

Comparison of Satellite Imagery Interpretation Using Known Association/Alliance-Level Data

VICKIE J. SMITH,
Department of Wildlife and Fisheries Sciences, South Dakota State University, Brookings

CHAD J. KOPPLIN
Department of Wildlife and Fisheries Sciences, South Dakota State University, Brookings

JONATHAN A. JENKS
Department of Wildlife and Fisheries Sciences, South Dakota State University, Brookings

BRUCE K. WYLIE
U.S. Geological Survey, EROS Data Center, Sioux Falls, South Dakota

JAMES E. VOGELMANN
U.S. Geological Survey, EROS Data Center, Sioux Falls, South Dakota

and ROBERT W. KLAVER
Department of Wildlife and Fisheries Sciences, South Dakota State University, Brookings, and U.S. Geological Survey, EROS Data Center, Sioux Falls, South Dakota

Introduction

The U.S. Geological Survey/National Park Service (USGS-NPS)
Vegetation Mapping Program was initiated in 1991 to provide na-
tional parks with information on their natural resources. Parks are
classified based on the association level of the National Vegetation
Classification System (NVCS) with a goal of 80% map accuracy.
Classification has been completed for three national parks in South
Dakota: Mt. Rushmore, Jewel Cave, and Wind Cave (Cogan et al. 1999).

These USGS-NPS maps offer a unique opportunity to evaluate the
extent to which satellite imagery can be used to classify land cover
based on the NVCS vegetation classification system (Table 1). Our
objectives were to determine the degree to which association-level
classifications could be interpreted from satellite imagery and the
accuracy of these interpretations. We hypothesized that with de-
tailed training data for association level or community type, accu-
racy of land cover maps would be at least 80%.

Table 1. Natural land cover classification system (Jennings 1996).

Category Example
Class Woodland
Subclass Mainly Evergreen Woodlands
Group Evergreen Needle-leaved Woodlands
Formation Evergreen Coniferous Woodlands with Rounded Crowns

Physiognomic

Floristic

Community Alliance ....... Juniperus occidentalis

Community Type ..............Juniperus occidentalis/Artemisia tridentata

Methods

Study Area - Wind Cave National Park (WCNP) is an 11,450 ha
parcel in the southern Black Hills of western South Dakota. The
park is a mosaic of ponderosa pine (Pinus ponderosa) stands and
mixed- and shortgrass prairies. The vegetation of the park, includ-
ing an approximately 2 km buffer, was classified by USGS-NPS to
an association level using aerial photography (Table 2; Figure 1 -
see Web version of the Bulletin at http://www.gap.uidaho.edu/Bulletins/8).

Classification Procedure - Landsat Thematic Mapper (TM) leaf-off
satellite scenes for path 33/row 29 and path 33/row 30 for 1992
were acquired through the Multi-Resolution Land Characteristics
Consortium. These two scenes, which encompass the Black Hills
of South Dakota (including WCNP), were mosaicked and processed
through an unsupervised classification with 20 iterations resulting
in 255 clusters and a 30 m cell size (Lauver and Whistler 1993,
Scott et al. 1993, Stoms 1996, Vogelmann et al. 1998). GIS Summary Interpretation - GIS summary techniques were used
in this study. A GIS summary calculates cross-tabulation statistics
between input files (known data entered on zeroed image and clus-
tered satellite scene) and creates an output report stating the agree-
ment between the known data and individual clusters. GIS summa-
ries were generated on the clustered image clipped to the WCNP
boundary (GIS Summary I) and on the mosaicked scenes, contain-
ing the entire Black Hills (GIS Summary II). Methods followed
Vogelmann et al. (1998).

 

Cluster Selection Interpretation - Using Imagine software (ERDAS),
the USGS-NPS vegetation coverage was overlaid on the clipped,
mosaicked satellite image. Each of the 255 clusters was selected to
determine the vegetation type it represented. Labels were given to
designate the present vegetation types. Once all 255 clusters had
been interpreted, any cluster that was designated by only one asso-
ciation was assigned to that type. Clusters containing more than
one association were compared and assigned to the association with
the highest percent composition.

Accuracy Assessment - Stratified, random points (908) were se-
lected for accuracy assessment based on the number of pixels in a
given category (Krebs 1989, Johnson et al. 1999). A minimum of
two points was assessed for small classes. Nine classes in the USGS-
NPS coverage could not be assessed for accuracy because of size,
accessibility, or lack of field information (Cogan et al. 1999). An
error matrix was generated for the remaining classes (Table 3).

Table 2. Categories included in original USGS-NPS coverage and comparison to three methods of interpretation.
Category USGS-NPS Coverage (ha)Cluster Selection (ha)GIS Summary I (ha)GIS Summary II (ha)
Purple-3-awn ­ Fetid Marigold Herbaceous Vegetation861.05382.32307.8320.76
Ponderosa Pine Limestone Cliff Sparse Vegetation129.77N/AN/AN/A
Redbeds Sparse Vegetation68.5342.12N/A6.48
Black Hills Rock Outcrop Sparse Vegetation1,394.62N/AN/AN/A
Shale Barren Slope Sparse Vegetation370.87359.6429.2632.4
White Sedimentary Rock Outcrop Sparse Vegetation269.59N/A22.6822.68
Bison Wallows1,24.35N/AN/AN/A
Burned Pine776.89942.84239.76317.52
Emergent Wetland Herbaceous Complex154.35N/AN/AN/A
Little Bluestem ­ Grama Grass / Threadleaf Sedge Herbaceous Vegetation4259.703304.81150.21140.48
Western Wheatgrass ­ Kentucky Bluegrass Complex10746.3613011.8417933.417677.44
Introduced Weedy Graminoid1,339.80N/AN/AN/A
Needle-and-thread ­ Blue Grama / Threadleaf Sedge Herbaceous Vegetation286.33243.0N/AN/A
Mountain Mahogany ­ Sideoats Grama I578.48197.6451.84129.6
Mountain Mahogany ­ Sideoats Grama II1,3187.97N/AN/AN/A
Lead Plant162.67155.52N/AN/A
Chokecherry Shrubland700.93165.24N/AN/A
Beaked Willow Shrubland1,37.77N/AN/AN/A
Western Snowberry Shrubland757.72758.16N/AN/A
Creeping Juniper ­ Little Bluestem Shrubland1,20.11N/AN/AN/A
Plains Cottonwood ­ Western Snowberry Forest352.7242.12N/AN/A
Boxelder/Chokecherry Forest1149.64N/AN/AN/A
Bur Oak Stand1,314.41N/AN/AN/A
Green Ash ­ American Elm / Western Snowberry Forest311.5893.96N/AN/A
Birch ­ Aspen Stand1,344.50N/AN/AN/A
Ponderosa Pine Woodland Complex I2713.995495.046207.846207.84
Ponderosa Pine ­ Little Bluestem Woodland4224.281707.485469.125469.12
Ponderosa Pine ­ Chokecherry Forest1001.03N/A3.243.24
Ponderosa Pine Woodland Complex II4506.535585.762993.762993.76
Young Ponderosa Pine Dense Cover1995.48N/AN/AN/A
Transportation, Communications, and Utilities331.31N/A100.4487.48
Crop and Pasture1744.391574.641244.161302.48
Other Agricultural Lands112.06155.52N/A32.4
Open Water28.1816.216.219.44
Strip Mines and Gravel Pits41.49N/A19.446.48

1 Not interpreted in any of the three classifications

2 Removed from coverage because of size when converted from vector to raster format

3No accuracy assessment performed

Results

Two categories were removed from the GIS summary analysis by
the gridding procedure. These classes, Creeping Juniper/Little
Bluestem Shrubland and Bison Wallows, contained less than 5 ha
in long, narrow corridors. A 30 m grid cell contains 0.09 ha, but
neighboring spectral classes can mask small areas and remove them
during the gridding process (Congalton 1997). Neither of these
categories could be interpreted using the Cluster Selection method
(Table 2).

Ten additional categories (less than 190 ha) were not interpreted by
any of the three methods (GIS Summary I, GIS Summary II, and
Cluster Selection Interpretations) listed in Table 2. Although some
vegetation types less than 190 ha were interpreted, polygon size
could be the factor determining whether or not the vegetation type
could be distinguished.

GIS Summary Interpretations - Fifteen of the 35 NPS categories
were present using GIS Summary I on WCNP (Figure 2 - see Web
version of the Bulletin at http://www.gap.uidaho.edu/Bulletins/8).
Seventeen of the 35 NPS categories were present using GIS Sum-
mary II on the Black Hills (Table 2; Figure 3 - see Web version of
the Bulletin).

Cluster Selection Interpretation - Nineteen of the original 35 NPS
categories were classified using Cluster Selection (Figure 4 - see
Web version of the Bulletin at http://www.gap.uidaho.edu/Bulletins/.
The Cluster Selection method resulted in classification of
six classes not found in the GIS Summaries of WCNP and the Black
Hills (Table 2).

Accuracy Assessment - Based on preliminary assessment of the
techniques used to classify the imagery, the Cluster Selection Inter-
pretation resulted in the highest number of classes and most closely
resembled the USGS-NPS coverage from aerial photography (Table
2). We restricted our accuracy assessment to this methodology be-
cause others would be less accurate by default. Because the USGS-
NPS coverage was only 73.0% accurate, we only determined simi-
larity between the classified satellite image and the USGS-NPS
coverage.

Four hundred and twenty-nine of 908 pixels were correctly classi-
fied to association using this interpretation, resulting in accuracy of
47%. User’s accuracy ranged from 0-84%, and producer’s accu-
racy ranged from 0-63% (Table 3 - see Web version of the Bulletin
at http://www.gap.uidaho.edu/Bulletins/8). However, using this
association-level information does accurately separate conifer from
grassland categories. When categories were grouped to the forma-
tion level, overall accuracy increased to 76% (Table 4).


Table 4. Formation-level accuracy assessment.

Item Purple-3-awn-

Fetid Marigold

 Redbeds SV Burned Pine Grassland Shrubland Evergreen Total
Purple-3-awn-Fetid Marigold 8 0 3 12 0 0 23
Redbeds SV 0 0 0 4 0 0 4
Burned Pine 6 0 14 1 0 18 39
Grassland 10 0 18 364 11 18 421
Shrubland 1 0 3 39 2 17 62
Evergreen 3 0 12 30 11 303 359
Total 28 0 50 450 24 356 908

 

Discussion

Three categories were overestimated in both of the GIS Summary
interpretations; these classes were Western Wheatgrass/Kentucky
Bluegrass Complex, Ponderosa Pine Complex I, and Ponderosa Pine
Complex II. These three categories comprised 49.2% of all of the
land classified in the USGS-NPS coverage. From these results, we
conclude that using a GIS summary to interpret land cover overes-
timated the dominant vegetation types.

During processing, the GIS summary showed confusion between
similar categories in a number of clusters. Confusion occurred
within herbaceous, pine, sparse vegetation, man-made, and agri-
cultural categories. Although recommended accuracy requirements

for GAP are 80% at the alliance level, overall accuracy approached
80% only after aggregating categories to the formation level. Ac-
curacy assessment included a comparison to the NPS classified
coverage, not to actual ground-truth information; hence, accuracy
of the original coverage (73.0%) may have interfered with the in-
terpretation of satellite imagery obtained from ground vegetation.
An accuracy assessment with ground-truth data from vegetation in
WCNP may increase accuracy of the classified image; this infor-
mation was not available at the time of the study.

Although techniques discussed were not able to distinguish accu-
rate alliance information, additional limitations exist with these interpretations. 
The Black Hills contain white spruce at elevations higher than 1,625 m. WCNP does not contain white spruce; there-
fore, we were not able to evaluate this vegetation type using any of
the three methods. In this case, white spruce would have been clas-
sified as Ponderosa Pine in the Black Hills, resulting in additional
inaccuracies in the land cover map. WCNP also is relatively small
in size when compared to the mosaicked scenes of the Black Hills.
Using a small area, such as WCNP (11,450 ha), as the sole training
data may cause inaccuracies when applied to a large area, such as
the two mosaicked Black Hills scenes (9,232,736 ha). Because of
these limitations, more known vegetation types would need to be
included in the Cluster Selection and GIS Summary methods to
adequately represent all vegetation types present in the Black Hills.

Recommendations and Further Research

While this study found an average accuracy of only 47% using sat-
ellite interpretations to predict alliance-level information, other
methods may have improved results. Ancillary data and
multitemporal satellite scenes can be used to assist in the interpre-
tation of categories that are small or that have specific habitat re-
quirements. For example, digital soil survey information may help
narrow the classification of large categories that dominate an inter-
pretation. However, these data are not available for the Black Hills. A second alternative that may enhance accuracy is the use of newer
satellite imagery. The MRLC acquisition of satellite imagery for
South Dakota dates to 1992. The ground-truth information was
collected in 1997, and aerial photography was taken in 1998 to be
used as ground-truth information. These temporal differences may
cause inaccuracies in classification. A Landsat 7 scene from 14
October 1999 for southwestern South Dakota is currently being in-
terpreted. Classification of this scene will likely compensate for
temporal changes and will allow the use of 15 m panchromatic reso-
lution to aid in interpretation of vegetation pattern.

Acknowledgments

We would like to thank Dan Roddy, Wind Cave National Park, for
access to the USGS-NPS coverage for the park. We thank the EROS
Data Center for help with methodology used in this study and for
support. We thank the National GAP Office for funding.

Literature Cited

Cogan, D., H. Marriott, J. Von Loh, and M.J. Pucherelli. 1999.

USGS-NPS Vegetation Mapping Program: Wind Cave National
Park, South Dakota. Technical Service Center, Denver, Colo-
rado. Technical Memorandum No. 8260-99-03.

Congalton, R.G. 1997. Exploring and evaluating the consequences
of vector-to-raster and raster-to-vector conversion. Photogram-
metric Engineering and Remote Sensing 63:425–434.

Jennings, M.D. 1996. Mapping units: Their classification and
nomenclature for gap analysis land cover data. Pages 71-78 in
J. M. Scott, T. H. Tear, and F. W. Davis, editors. Gap Analysis:
a landscape approach to biodiversity planning. American Soci-
ety for Photogrammetry and Remote Sensing, Bethesda, Mary-
land.

Johnson, R.R., K.F. Higgins, D.E. Naugle, and J.A. Jenks. 1999. 
A comparison of sampling techniques to estimate number of
wetlands. Wildlife Society Bulletin 27:103-108.

Krebs, C.J. 1989. Ecological methodology. Harper & Row, New York, 
New York. 654 pp.

Lauver, C.L., and J.L. Whistler. 1993. A hierarchical classifica-
tion of Landsat TM imagery to identify natural grassland areas
and rare species habitat. Photogrammetric Engineering and Re-
mote Sensing 59:627-634.

Scott, J.M., F. Davis, B. Csuti, R. Noss, B. Butterfield, C. Groves,

H. Anderson, S. Caicco, F. D’Erchia, T.C. Edwards, Jr., J.
Ulliman, and R.G. Wright, 1993. Gap analysis: A geographic
approach to protection of biological diversity. Wildlife Mono-
graphs 123:1–41.

Stoms, D.M. 1996. Actual vegetation layer. National Gap Analy-
sis Program Handbook. http://www.gap.uidaho.edu/handbook/
LandCoverMapping.

Vogelmann, J.E., T. Sohl, and S.M. Howard. 1998. Regional char-
acterization of land cover using multiple sources of data. Pho-
togrammetric Engineering and Remote Sensing 64:45–47.