When you connect to the NBII Metadata Clearinghouse you will be able to search through metadata-based descriptions of biological data sets and information products from many different sources to identify those that meet your particular search criteria.
The NBII Metadata Clearinghouse: http://metadata.nbii.gov/
The NBII Home Page: http://www.nbii.gov/
Powered by Mercury
Image-Based Vegetation Classification and Mapping
- National Park Service Pacific Northwest Region
Vegetation and Landform Database Development Study
Region of the National Park Service contracted
with Pacific Meridian Resources to develop and
produce a comprehensive GIS vegetation land cover
and geomorphic landform database for Olympic
National Park. The study was designed to develop a
comprehensive, consistent inventory and mapping of
the vegetation and landform characteristics for
the park using digital Landsat Thematic Mapper
(TM) satellite imagery and field collected data as
the primary information bases.
vegetation and landform databases for the project
areas and demonstrate the effectiveness of Pacific
Meridian Resources' procedures for producing those
databases.
-- Pacific Northwest Region Vegetation and
Landform Database Development Final Report" from
which this metadata was developed can be obtained
online at
Suite 850
Brock, Bruce Freet, Roger Hoffman, Ed Schriener,
Darin Swinney
steps: 1) sample selection 2) labeling of
reference and map sites and 3) generation of error
matrices and estimates of accuracy. A population
of vegetation polygons 5 acres or greater from
which to sample was generated from the
unclassified Landsat TM imagery using Image
Segmentation. The resulting polygons were then
overlaid on the image-based raster classification
data. Summaries of the raster classification data
for size/structure, species and crown cover data
layers were generated for each of the segmentation
polygons. Labeling rules were developed to assign
a map label to each polygon based on the summaries
of the pixel data falling within the polygon. The
relative proportion of each class labels in the
classified pixel maps was used to determine the
number of samples to select in each class from the
polygon coverage. The three pixel layers were
overlain and the number of acres of each
crown-cover/size-structure/ species combination
was determined. The relative proportion was then
applied to each strata in the polygon coverage and
a random number generator was used to select the
proper number of samples from each strata.
The location of the accuracy assessment sites were
displayed over the imagery on screen and were then
manually delineated on aerial photographs. Each of
these reference sites were photo-interpreted and
labeled with the appropriate size/structure,
species, and crown class label. To account for
variation in human photo interpretation, fuzzy
logic was incorporated in the reference call.
After photo interpreting the site, the interpreter
assigned a "best", "good", "acceptable", "poor" or
"unacceptable" call to all possible labels.
Approximately 4000 field based accuracy
assessment sites were established across the four
parks in the study. An accuracy assessment field
form was completed for each site. Each of these
sites were re-interpreted based on the field
collected data and notes and a fuzzy-logic matrix
was completed. These field-based accuracy
assessment sites were then integrated with the
office-based photo-interpreted accuracy assessment
sites to comprise the final reference data set.
The dates of the imagery and the aerial
photographs differed, so some land cover changes
had occurred. If the vegetation on the site had
changed (e.g., fire, vegetation regrowth, etc.)
between the dates of the imagery and the photos,
the site was eliminated from further analysis.
Difference matrices were employed to analyze
the differences between the reference data and the
map data in this study.
Due to the complex forest species and size diversity and
complexity present throughout the park, the
Olympic National Park presented a tremendous
challenge in the image classification process.
Nevertheless, the final classification accuracies
for each layer were all well in excess of 80%.
The high quantitative accuracy for the species
layer on Olympic National Park was due in large
part to the extensive draft map reviews and
comments by Pacific Meridian and NPS personnel as
well as the significant time devoted by NPS and
Pacific Meridian personnel to developing,
refining, and applying ecological species models
to the forest species mapping process. These
models corrected many spectrally confused species
classes while identifying potential areas for
improvement in the species classification map
layer.
No one species class notably contributed to the
species classification map error. More often, the
tree density of the site and the shadowing present
influenced the site accuracy.
Imagery - Olympic National Park
the Landsat Thematic Mapper data (bands 1,2,3,4,5,
and 7) were utilized for this mapping project. In
addition, an additional imagery band was created
and utilized by ratioing original wavelength band
3 (red visible) with band 4 (near infra-red).
This particular band combination has been shown to
assist in minimizing the impact of shadowing in
satellite image classification. This new ratio
band was added to the original six reflective
bands to create a new seven-band imagery data set
that was used for all project landcover
classification. Once a complete image data
set was built for each park, the satellite image
data was further subset into "eco-regions" for
Olympic and North Cascades parks. Field
data collection for the purpose of supporting the
image classification efforts was completed
simultaneously with the vegetation inventory field
data collection. While the primary objective of
field data collection for this study was to
complete a comprehensive vegetation inventory for
the park, valuable information was also gathered
to assist project image analysts in the
classification of the Landsat TM satellite image
data. This information was primarily in the form
of detailed field notes for specific land areas,
field descriptions of ecological trends, and
identification of anomalies and/or ecotones. Image
analysts also made extensive use of the specific
plot data and notes collected as part of the
vegetation inventory. Upon completion of the
band ratioing described in the preprocessing
section above, an unsupervised classification was
performed on the imagery set for each park.
Seventy-five to one-hundred spectral classes were
identified in the classification. Clustering
analysis performed on these classes identified
several classes which were spectrally very similar
and represented very general landcover types
within the study area. For example, water, snow,
and obvious bare ground were each represented by
multiple spectral classes. In cases where a
spectral class could reliably be found to
represent a single land cover type, i.e. water,
snow, etc., the spectral classes were simply
relabeled to that land cover type. The
remainder of the spectral classes were then given
a unique color so that it could be easily
distinguished from the other spectrally different
classes. This newly colored spectral variation
map was used to identify areas that may represent
acceptable training sites for subsequent image
supervised classification based on their spectral
homogeneity. An image training site is an area of
consistent tree crown cover, tree species or
species mixes, tree size or size mixes, forest
structure, or non-forest type that is utilized by
imagery analysts as an example or representation
of an area possessing those landcover
characteristics. Training sites are defined as
having evenly distributed vegetation or other land
cover throughout the entire training site polygon.
This is evidenced by uniform texture, color, and
tone throughout the polygon on both aerial
photography and satellite imagery. Individual
pixels within a single training site polygon
should all have similar spectral reflectance in a
single satellite data band. Following field data
collection, 550 training sites were delineated.
All training site polygons were
digitized directly on the digital satellite
imagery. These polygons encompassed the imagery
pixels that contain the spectral reflectance
values associated with the vegetation described by
the training site. Various statistical parameters
describing the digital numbers of the pixels from
all seven bands of imagery were generated for each
training site. In addition, identical statistical
parameters were generated from each of the initial
unsupervised classes produced prior to field data
collection. Utilizing a multivariate cluster
analysis statistical technique, training
statistics generated from both supervised and
unsupervised approaches were grouped together.
The results of the cluster analysis was a
final set of supervised training sites that
represents all the vegetation types to be mapped
in the study areas as well as representing the
range of spectral variation in the satellite
imagery within the study area. A series of
supervised classifications was run on the study
area image data set utilizing the spectral
statistics of the digitized training sites. After
each classification, the map was evaluated for
accuracy and consistency. Areas that were
consistently classified correctly were
subsequently set aside in an evolving raster GIS
data set and removed from any further
classification processing. Another iteration of
spectral analysis, classification, review, and
masking was performed. Draft plots of
all three raster map layers were produced at a
scale of 1:24,000 and reviewed for accuracy by the
image analysts. These hardcopy maps were also
presented to each specific park headquarters for
National Park Service (NPS) review and comments.
In conjunction with the image analysts, about 30
NPS employees were estimated to have reviewed and
commented on the maps during multiple weeks. Park
personnel used aerial photography,
orthophotography, existing ancillary GIS and plot
data, and most importantly personal knowledge to
evaluate the maps and comment on their accuracy
and consistency. Image analysts and NPS field
personal field verified the many of the maps for
most parks over the course of several months.
After comments were submitted, maps were edited
and reprocessed. During and following the
image classification process, ecological modeling
was utilized for the park as a tool for enhancing
the identification and classification of forest
species throughout the park. Two distinct types of
ecological species models were executed: 1)
park-wide model; and 2) site/region specific
model. The park-wide model was developed and
executed for the entire park as a single unit.
This model was designed to simply identify and
"flag" potential species misclassifications for
various elevation/aspect zones throughout the park
where particular species occurrence could be
fairly reliably predicted. The site
specific model of Olympic National Park consisted
of the following: The Olympic National Park
imagery was classified using three independent
eco-zones: Coastal Zone, Western Inland Zone, and
Eastern Inland Zone. For each zone, a specific
species model similar to the park-wide model
described above was developed and executed
providing a more eco-zone specific series of
elevation/aspect breaks for forest species which
more precisely evaluated the species
classification for that area. Numerous other
species-specific models were developed and
executed that not only incorporated information
from elevation and aspect data, but also utilized
image training site classification data to refine
and enhance the species classification. In total,
approximately 150 site-specific models to correct
classification problems were developed for Olympic
National Park.
Suite 850
space in a particular land area
are described in this coverage.
Suite 850
reliability or suitability of this information for
a particular purpose. Original data elements were
compiled from various sources. Spatial
information may not meet National Mapping Accuracy
Standards. This information may be updated,
corrected, or otherwise modified without
notification. For additional information about
this data contact the National Park Service.
Information Infrastructure Metadata