A healthy coastline is essential for the economic and ecological
well being of the United States and its territories, and conserving
and managing coastal resources is a key priority. One critical
aspect of this work is understanding the changes that affect coastal
landscapes.
Many factors affect coastal ecosystems, especially the land cover
and use of the watershed draining into oceans and lakes. Water that
runs off the land carries pollutants that affect the receiving water
bodies, and urban coastal development changes the extent and quality
of coastal habitats. Additionally, the natural processes of erosion
and deposition change the land forms around the coastline. Land
cover change affects coastal communities, the coastline and the
benthic environment adjacent to the coast.
Quantifying land cover change in coastal communities provides vital
information that helps local decision makers make better land
management choices. Understanding the rate and types of change in a
consistent and cost-effective way has been the responsibility of the
National Oceanic and Atmospheric Administration (NOAA) Coastal
Services Center’s Coastal Change Analysis Program (C-CAP), based in
Charleston, S.C.
Coastal Land Use Monitoring
C-CAP, originally conceived in the 1980s, conducted coastal analyses
on a project-by-project basis in the 1990s through grants and
cooperative agreements. The projects provided valuable regional
coastal data, but weren’t well suited for standard, comprehensive
national coverage. The program now takes a more holistic approach by
monitoring the entire U.S. coastline for land cover and land cover
change.
To achieve this objective, CSC teamed with the U.S. Geological
Survey (USGS) and the interagency Multi-Resolution Land
Characteristics (MRLC) consortium to develop the National Land Cover
Database (NLCD), a national database of land cover components based
on Landsat Thematic Mapper imagery collected in 2001. By joining
forces, these federal agencies have minimized data duplication and
lowered costs while providing a consistent baseline for the country
with a unified classification scheme. NLCD 2001 datasets are land
cover, percent impervious and percent tree canopy density, and
eventually will be available for the whole country. Their status can
be monitored at
http://www.mrlc.gov/nlcd_overall_status.asp. In addition to the
base NLCD layers, all the C-CAP zones have estuarine and palustrine
wetland classes, a 5-year retrospective land cover change map and an
accuracy assessment of the land cover map. For more NLCD 2001
details, see
www.mrlc.gov/mrlc2k_nlcd.asp, and for more details on the C-CAP
program see www.csc.noaa.gov/landcover.
A Unique Classification Approach
The approach used by USGS and NOAA to produce their maps was
implemented by Sanborn. The approach differs from traditional
cluster analysis used throughout the remote sensing industry, as it
applies a regression tree analysis methodology (Figure 1).
Training sites are gathered through site visits and
photointerpretation of aerial or high-resolution satellite imagery.
The sites are labeled according to a predefined classification
system, and the spectral information reflecting from these areas is
combined with other data such as relief, soils, Landsat scene number
and date. Once the statistical relationships between the training
site class and these independent variables are calculated, the
equation is reapplied to all pixels to create the basis for the land
cover classification.
Creating the automated map is only the first part of the process.
Next the map is reviewed for errors, which are corrected through
modeling using ancillary data and reclassifying specific areas or
classes. Finally, when all automated methods are exhausted, maps are
edited by hand to ensure accuracy. Figure 2a shows examples of the
final map for Oregon. Once the base date map is established, Sanborn
produces a 5-year retrospective change map. The method maps change
in the imagery between 1995 and 2001, rather than the classified
map. A significant advantage of this approach vs. comparing two land
cover maps using different methods—e.g., comparing NLCD 2001 to NLCD
1992—is that the user doesn’t compare the accumulation of errors of
both maps. In fact, he or she compares the spectral change that
relates to the land cover change within the pixel.
Obviously not all spectral changes represent a class change. For
example, agricultural changes resulting from crop rotation show
spectral change but not class change. The first stage of change
detection is to create a mask that encompasses all changes between
the two images. Once this has been done, areas that have changed can
be mapped into their respective classes. Thus, any class change can
be identified (Figure 2b) from this map, and the base date map
(Figure 2c) and the 5-year retrospective map can be created (Figure
2a). The final land cover product is assessed for accuracy based on
a minimum mapping unit of two acres, requiring an overall accuracy
of greater than 85 percent for approval.
Mapped Areas
The effort already undertaken to create a mapping baseline will
produce a complete baseline dataset for the U.S. coast. Figure 3
shows the status of the coastline mapped with the C-CAP
specifications.
The data have been or are being produced by Sanborn and other
contractors, the USGS Southeast Gap Analysis program and NOAA staff.
Once approved, the data are available at
www.csc.noaa.gov/landcover. These maps have been used for many
applications, including rare species protection, wetland monitoring,
runoff control, watershed management, open space preservation and
water resource management. Examples of these applications can be
seen at
www.csc.noaa.gov/crs/lca/apps.html.
Future Mapping
Although current land cover products have many uses, higher
resolution products are in demand for more operational decision
making. In addition, Landsat Thematic Mapper is becoming an
unreliable data source on which to base future mapping efforts.
Landsat 7 isn’t delivering imagery that is reliable enough for land
cover mapping, and Landsat 5, although still producing after a long
and fruitful life, is well past its expected life span and is thus a
risky basis for a mapping program.
Based on these realities there are really two needs in the coastal
land cover mapping world:
1. Maintain the information on the overall state of the coastline at
a national and state level for federal and state decision makers.
2. Develop much higher resolution land cover maps to help local
managers make decisions.
There’s also a synergy that can be gained by working at a higher
resolution in collaboration with state and local governments for
land cover mapping. An example of this is a partnership that Sanborn
is coordinating with the state of Maine, USGS and NOAA that is
already under way.
Land Cover Monitoring Program
Moving forward, there is a need to regularly update the coastal land
cover coverages nationwide and where possible increase map
resolution to less than 10 meters. This requires a cheap source of
higher resolution data than is currently provided by Landsat. The
project also requires a low-cost processing methodology that lends
itself to large area collection efforts and can leverage existing
C-CAP datasets.
The two commercial options for medium-resolution imagery are
France’s SPOT 5 and India’s IRS ResourceSat 1 (IRS-D) satellites.
Both can collect medium- and high-resolution imagery and are
designed to collect imagery over large areas (Figures 4a and 4b).
Using a medium-resolution sensor will allow researchers to conduct
change-detection analysis based on existing and new Landsat imagery.
High-resolution panchromatic imagery then can be used to create
detailed boundaries (segments) that can be labeled using the
medium-resolution map (Figure 5). Using software packages like
Definiens Imaging’s eCognition (www.definiens-imaging.com),
image segments can be created in a hierarchical fashion to produce
5-meter boundaries. The updated medium-resolution land cover dataset
can be used to label image segments. The methods for labeling
segments can use the medium-resolution land cover map and other
ancillary data such as slope, soils, roads, hydrology and other data
elements that would complement the classification.
The advantages of this proposed coastal change methodology is that
it leverages existing data with 30-meter land cover data and change
products. It produces a higher resolution product by using 5- to
6.25-meter imagery to define landscape boundaries, which results in
reduced speckle, a more readable land cover map (more like polygons
than rasters) and the identification of smaller land cover features.
The process also reduces cost to the user, allowing more extensive
areas to be mapped for the same price. The approach is being used to
map land cover for the state of Maine and in a pilot project in
Florida’s panhandle.
High-Resolution Options
Although the aforementioned approach may work for nationwide monitoring,
many projects need more detailed site-specific data. Coastal communities
require high-resolution land cover maps that can be derived from
high-resolution digital imagery. There are a wide variety of
high-resolution sensors now available from aircrafts (Figure 4c) and
satellites (Figure 4d), and many of these sensors provide imagery
suitable for automated land cover classification .
Four-band (blue, green, red and near-infrared), 12-bit imagery generally
is preferred for land cover classification. The timing of the collection
window can be set depending on the type of land cover classification
required; impervious classifications are best conducted with leaf-off
imagery, forest types with early senescence and wetland mapping with
spring green up. For general land cover classifications, often spring
green up allows the best differentiations of land cover types.
Sanborn takes a two-stage approach to high-resolution land cover
mapping. The first stage is to create an impervious map that contains
all manmade structures in the image. This is often the most important
aspect for the map user. The second stage is to map the vegetated lands.
This is done by conducting segmentation on the imagery, then labeling
the segments based on spectral properties, shape, texture and ancillary
datasets. Vegetated segments then are aggregated by land cover to the
predetermined minimum mapping unit (Figure 6).
Ongoing Development
The C-CAP program has created a consistent, quality-controlled dataset
for the nation’s coastline. The data created from Landsat Thematic
Mapper imagery have 30-meter resolution and a 2001 base date. Although
they have proved useful for many applications, thought must be given to
how coastal land cover mapping will progress in the future. Two
methodologies suggested by Sanborn will introduce high-resolution
mapping to coastal monitoring.
The coastal change methodology leverages previously generated land cover
datasets, reducing processing time and cost and maintaining consistency
among current and future coastal monitoring products. The coastal change
methodology uses sensors that can collect imagery at multiple ground
sampling distances. Then it uses the multispectral bands for change
detection and classification and the higher resolution panchromatic
imagery for the land cover boundaries.
Site-specific land cover data needed for localized areas can be created
from the variety of digital sensors now on the market. Creating the land
cover map by first classifying the impervious and then the vegetated
land with segmentation will lead to usable, accurate land cover maps
that will help identify areas for protection and restoration, determine
land cover change, and enhance the planner’s toolbox by improving the
information available for local decision making.
