IKONOS Imagery for Resource Management: Tree Cover, Impervious Surfaces, and Riparian Buffer Analyses in the Mid-Atlantic Region

Abstract

High-resolution imagery from the IKONOS satellite may be useful for many resource management applications. We assessed the utility of IKONOS imagery for applications in the mid-Atlantic region, including mapping of tree cover, impervious surface areas, and riparian buffer zone variables in relation to stream health ratings. We focused on a 1313-km2 area in central Maryland using precision-georeferenced IKONOS products. We found the IKONOS imagery to be a valuable resource for these applications, and were able to achieve map accuracies comparable to manual aerial photo interpretation. We were also able to use derived data sets for consistent assessments over areas that would be difficult to accomplish with traditional photographic mapping methods. For example, we found that a stream health rating of excellent required no more than 6% impervious cover in the watershed, and at least 65% tree cover in the riparian zone. A rating of good required less than 10% impervious and 60% tree cover. A number of issues associated with application of the IKONOS data arose, however, including logistics of image acquisition related to phenological and atmospheric conditions, shadowing within canopies and between scene elements, and limited spectral discrimination of cover types. Cost per unit area was also a nontrivial consideration for the image data products we used, but allowed us to provide valuable derived products to agencies in support of their planning and regulatory decision-making processes. We report on both the capabilities and limitations of IKONOS imagery for these varied applications.

Full text

IKONOS imagery for resource management: Tree cover, impervious
surfaces, and riparian buffer analyses in the mid-Atlantic region
Scott J. Goetz
a,b,
*, Robb K. Wright
b
, Andrew J. Smith
b
, Elizabeth Zinecker
b
, Erika Schaub
b
a
The Woods Hole Research Center, PO Box 296, Woods Hole, MA 02543-0296, USA
b
Department of Geography, University of Maryland, College Park, MD 20742-8225, USA
Received 9 August 2002; received in revised form 4 June 2003; accepted 4 July 2003
Abstract
High-resolution imagery from the IKONOS satellite may be useful for many resource management applications. We assessed the utility of
IKONOS imagery for applications in the mid-Atlantic region, including mapping of tree cover, impervious surface areas, and riparian buffer
zone variables in relation to stream health ratings. We focused on a 1313-km
2
area in central Maryland using precision-georeferenced
IKONOS products. We found the IKONOS imagery to be a valuable resource for these applications, and were able to achieve map accuracies
comparable to manual aerial photo interpretation. We were also able to use derived data sets for consistent assessments over areas that would
be difficult to accomplish with traditional photographic mapping methods. For example, we found that a stream health rating of excellent
required no more than 6% impervious cover in the watershed, and at least 65% tree cover in the riparian zone. A rating of good required less
than 10% impervious and 60% tree cover. A number of issues associated with application of the IKONOS data arose, however, including
logistics of image acquisition related to phenological and atmospheric conditions, shadowing within canopies and between scene elements,
and limited spectral discrimination of cover types. Cost per unit area was also a nontrivial consideration for the image data products we used,
but allowed us to provide valuable derived products to agencies in support of their planning and regulatory decision-making processes. We
report on both the capabilities and limitations of IKONOS imagery for these varied applications.
D2003 Elsevier Inc. All rights reserved.
Keywords: IKONOS imagery; Resource management; Mid-Atlantic region
1. Introduction
Information on land cover has become an integral part of
the environmental and developmental planning process. It
has aided in the advancement of more effective land use
planning, habitat assessments, and hydrological applica-
tions. Historically, land cover and land use information
was obtained by a combination of field measurements and
aerial photo interpretation. This approach typically required
intensive interpretation by expert analysts, and cross vali-
dation methods to ensure that analyst interpretations were
consistent. Recently, satellite imagery has become available
at spatial resolution nearly comparable to aerial photo-
graphs, with the added advantage of digital multispectral
information more complete even than those provided by
digital orthophotographs (DOQs). The platform stability of
high-resolution (1– 5 m) imagery acquired from Earth-
observing satellites provides another advantage to aerial
photographs acquired from aircraft, which roll, pitch, and
yaw during flight and require corrections for those effects.
Because of these advances and advantages, as well as
commercial potential, several companies have launched or
plan to launch high-resolution satellites in the near future
(Stoney, 2001).
High-resolution multispectral imagery has many poten-
tial benefits to government organizations, nonprofit agen-
cies, and a wide array of mapping and related commercial
ventures (Dial, Bowen, Gerlach, Grodecki, & Oleszczuk,
this issue; Sawaya et al., this issue; Tanaka & Toshiro,
2001). Applications of these data can aid and assist the
monitoring and management of resource lands, parks, wet-
lands, and other protected areas, as well assess the effects of
natural disasters or complement protective measures in areas
0034-4257/$ - see front matter D2003 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2003.07.010
* Corresponding author. The Woods Hole Research Center, PO Box
296, Woods Hole, MA 02543-0296, USA. Tel.: +1-508-540-9900; fax: +1-
508-540-9700.
E-mail address: sgoetz@whrc.org (S.J. Goetz).
www.elsevier.com/locate/rse
Remote Sensing of Environment 88 (2003) 195 – 208

that have the potential to burn or flood, to name just a few.
High spatial resolution imagery is not a panacea for these
applications, however, owing to a number of issues that
arise with their use. For example, the increased textural
information available in fine-resolution imagery allows for
improved interpretation based on the shape and texture of
ground features, but techniques that have been developed to
process and analyze current satellite data, including vegeta-
tion indices or multitemporal classification techniques that
utilize mid-infrared or thermal channels (e.g., Varlyguin,
Wright, Goetz, & Prince, 2001), may not be applicable to
the additional information provided by high-resolution sat-
ellites. Other issues include difficulties in acquiring consis-
tent sequential acquisition dates, intensive computer
processing time and disk space required to store larger
image data sets, and considerations of economic efficiencies
(Fisher & Goetz, 2001).
The objective of this paper was to assess the practicality
of using high-resolution imagery, in this case provided by the
IKONOS satellite operated by Space Imaging, for resource
management applications in Montgomery County in the
mid-Atlantic region of the United States. We sought to utilize
the IKONOS imagery as an alternative to air photo interpre-
tation for updating the forested lands map, as well as to map
changes that had occurred in land use, particularly residential
development and intensification of impervious surface areas.
Additional specific goals were to assess the utility of the
IKONOS imagery for production of tree cover and impervi-
ous surface area maps, and to examine those map products
and derivatives, including riparian buffer zone land cover,
relative to stream water quality. Related applications include
the use of IKONOS to train subpixel algorithms of tree cover
and impervious surfaces using coarser resolution imagery
(e.g., Landsat). The tree and impervious area maps, when
combined, provide an improved ability to meet resource
management goals, particularly with respect to comprehen-
sive planning, rural land protection, and goals for improved
water quality. We identify and address some of the benefits
and limitations of high-resolution imagery for these research
applications.
2. Study area and data sets
The study area encompasses all of Montgomery County,
MD (1313 km
2
)(Fig. 1), an area comprised of a mixture of
forest and farms interspersed with a range of residential
developments and industrial/commercial zones. The county
was selected for this research because it seeks to develop
improved geospatial information, capabilities, and technol-
ogies to assess impacts of environmental change, and a
range of resource management decisions. The Maryland
National Capital Parks and Planning Commission (M-
NCPPC) and the Montgomery Department of Environmen-
tal Protection (DEP) are well advanced in their utilization of
geographic information systems (GIS) and have been na-
tionally recognized and awarded for their countywide forest
preservation and stream protection strategies. The Planning
Board of Montgomery County has established master plans
for each region to plan future development, and has been at
the forefront of many of Maryland’s ‘‘Smart Growth’’
programs. One of these includes the ‘‘Legacy Open Space’’
program to fund the purchase of culturally and environmen-
tally unique lands and to ‘‘protect the county’s surface water
supply’’ (M-NCPPC, 2000). To date, more than 12,000 ha
have been protected as parks and open space, at a cost of
over US$137 million, and additional acquisitions are
planned within a 39,000-ha agricultural preserve. There is
a desire by the DEP, M-NCPPC, and others to have these
lands connected to other parcels reserved under the
counties’ parks system and related land conservation pro-
grams, providing the county a ‘‘green infrastructure’’ of
natural areas to promote recreation opportunities, protect
water resources, and maximize biodiversity (Weber & Wolf,
2000). In the 10 years since the last aerial photo-based forest
land cover classification, however, there had been consid-
erable residential development and expanded transportation
projects, with associated predictions of rapid future rates of
land conversion (Jantz, Goetz, & Shelley, in press). These
factors provided an impetus for the applications research
presented here.
2.1. IKONOS imagery
IKONOS is the first commercially owned satellite
providing 1-m resolution panchromatic image data and 4-
m multispectral imagery (Dial et al., this issue).The
multispectral image data include three visible and one
infrared channel (Table 1). Data are collected in 11-bit
radiometric resolution and provided in a format compatible
with image analysis software. The tile size for each
individual scene is 11.3 11.3 km. Because of its pointable
off-nadir viewing capability, the satellite revisit interval is
as little as 3 – 4 days.
The IKONOS images of Montgomery County (Fig. 2)
were obtained through the NASA Scientific Data Purchase, a
program designed to make remote sensing data sets available
for research and practical applications (Birk, Stanley, &
Snyder, this issue). The acquired imagery was equivalent
to Space Imaging’s Carterra Precision product, which is a
‘‘precision-georeferenced’’ image data product with an ab-
solute horizontal geometric accuracy of 5 m. The NASA
Scientific Data Purchase contract specification allowed for
mislocation errors of 250 m in standard products (see Helder,
Coan, Patrick, & Gaska, this issue), thus a precision product
was required for the analyses we wished to conduct. The
orthocorrected product we acquired had an RMS error of
approximately 1.9 m. Eleven IKONOS image tiles were
acquired in six swaths to provide complete coverage of the
entire county (Table 2). The nominal cost of the imagery was
US$141 km
2
. Although the IKONOS platform has a short
repeat schedule, a cloud-free image could not be acquired for
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208196

Panel 7 during a 12-month acquisition window. For this
panel, a cloud and cloud shadow mask were created manu-
ally, through which all processings were filtered. Note that
several of the panels were acquired in late spring during leaf-
off conditions, while others were acquired in early summer
leaf-on conditions (Fig. 2). We also note that the off-nadir
view of the county for all panels, in conjunction with the
collection azimuth (Table 2), affected the radiance detected
at the sensor from the combined atmosphere and surface
constituents.
2.2. Planimetric and natural resource data layers
The sole source of reference (i.e., training and validation)
data came from planimetric GIS data sets provided by the
M-NCPPC. The primary data set for tree cover mapping was
a ‘‘natural features’’ database that contained countywide
polygon data for forested lands, agricultural lands, and
water features. The database was visually interpreted by a
commercial vendor contracted by a multiagency consor-
tium, based on use of 1992 aerial photography. The derived
forest cover map had a large minimum mapping unit to
represent forest areas, rather than individual or small clus-
ters of trees. The agricultural areas contained a wide range
of cover types, including row crops, pasture land, fallow
fields, and grassy regions.
Additional reference data were provided by the plani-
metric data of the built environment, which included highly
detailed polygons of road and building footprints. This
provided the primary data set used for impervious area
mapping. Because the planimetric and natural features
coverages were developed using photography flown some
10 years ago (1992), changes had occurred in the county up
Table 1
IKONOS spectral band widths specified as full width at half maximum
Band Spectral wavelength (nm)
1 445 – 516 Blue
2 506 – 595 Green
3 632 – 698 Red
4 757 – 853 Near-infrared
Pan 450 – 900 Panchromatic
Fig. 1. Study area map of Montgomery County, MD, within the mid-Atlantic region and Chesapeake Bay watershed.
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208 197

to the time of the IKONOS acquisition. This required
evaluation of the accuracy of the impervious and natural
features coverages, and screening for data sets suitable for
algorithm training and map validation.
3. Methods
3.1. IKONOS image preprocessing
The four multispectral IKONOS bands were radiometri-
cally corrected to at-sensor reflectance utilizing the methods
outlined by Goetz (1997) and the calibration parameters
provided by Space Imaging in the imagery metadata. These
corrections to top-of-atmosphere reflectance minimize the
errors associated with solar elevation and allow for between-
band comparisons of data values.
Preliminary supervised and unsupervised forest/nonfor-
est classifications were done using the IKONOS imagery.
An unsupervised classification was run using an ISO-
DATA algorithm, specifying 255 clusters, 10 iterations,
and a confidence interval of 0.96. The resultant clusters
were assigned to three categories (forest, nonforest, or
unknown) using manual interpretation of the imagery
itself. Clusters assigned to the ‘‘unknown’’ class were
located in both the forested and nonforested areas. These
were used as a mask to create a new image, which was
missing the areas that were earlier assigned to either
forest or nonforest. The unsupervised classification was
then run again on the revised image, consisting entirely of
unknown areas, using the same set of clustering param-
eters. This process was repeated five times, until it was
difficult to identify any new unknown areas. A supervised
maximum likelihood classification (MLC) was also de-
veloped using visually interpreted training areas for each
Table 2
IKONOS acquisition metadata
Panel Date Local
time
Sun
elevation
Sun
azimuth
Collection
elevation
Collection
azimuth
0 5/1/2001 10:59 62.08 143.46 78.25 24.83
1 5/1/2001 11:00 62.16 143.80 88.49 95.46
3 5/23/2001 11:02 67.36 139.09 85.83 276.78
4 4/6/2000 10:35 51.28 139.97 76.46 119.51
5 6/26/2000 10:53 66.98 126.80 79.06 301.54
7 4/20/2000 10:45 57.25 140.47 77.35 349.31
Fig. 2. Location and acquisition dates of IKONOS image panels of the study area. Note the cloud cover remaining in Panel 7 despite a 14-month acquisition
window. (Includes material from Space Imagingn.)
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208198

of four land cover types (tree/forest, agriculture, urban/
built, and water). A total of 100 training areas was
identified. Classifications produced using these two
approaches were assessed for accuracy using 1000 ran-
domly selected samples that were visually interpreted to
the appropriate class, of which 100 were visited on the
ground to verify the interpretations.
Based on other work using Landsat Thematic Mapper
(TM) imagery, we suspect that the modest classification
accuracies we found (75 – 85%) using these classification
approaches with IKONOS imagery were due to a com-
bination of leaf-on/leaf-off imagery, limited image spec-
tral bands and spatial– spectral variability within and
between image components, as well as limitations in
the use of hand-delineated training areas or statistically
defined clusters of similar objects. As a result, we
decided to take a different approach to the classification,
focused on the use of a decision tree classifier and a
large number of training samples (described below). We
also derived four additional image variables to aid the
spectral discrimination process including: (i) Normalized
Difference Vegetation Index (NDVI), (ii) Atmospherically
Resistant Vegetation Index (ARVI), (iii) NIR/red (simple
ratio), and (iv) NIR/blue. The NDVI was added to aid in
the discrimination of sediment-laden water from wet
fields. ARVI (Kaufman & Tanre, 1992) was included to
help remove artifacts within the canopy itself, such as
canopy shadow. The simple ratio and NIR/blue ratio were
included to aid the discrimination of agricultural fields
and shrubs from trees.
3.2. Reference data preprocessing
In the case of tree cover mapping and accuracy assess-
ment, natural features reference data screening was done
by overlaying the IKONOS imagery with the reference
coverage and by deleting polygons whose cover type did
not match. The incorrectly attributed polygons included
forested and agricultural areas that had been converted to
residential or commercial (developed) land use, as well as
forested or agricultural areas that had some development
within the polygon to the extent that it could skew the
training and validation data. The water polygons were also
evaluated for accuracy with errors occurring mostly in
farm ponds that had been drained or developed, as well
as portions of water bodies occurring in rapids and over
rock that were subject to changes associated with flow
volume. These screening processes permitted the use of the
natural features coverage to develop independent training
and cross-validation data sets for the decision tree classi-
fier, and subsequent development of a tree cover map
rather than forest cover per se (where forest would imply,
e.g., >60% tree cover).
In order to minimize edge pixels, the forest, grass, and
water polygons were internally buffered by 20 m (five
IKONOS pixels). Areas converted from a natural to a
built environment since the creation of the data set were
not added to the training data set used for image
classification since the existing data included enough
samples for the spectral responses of buildings and roads.
The planimetric data were rasterized to 4-m pixels
utilizing the same pixel footprint as the IKONOS imag-
ery, so that there was exact pixel coregistration. The final
reference data set consisted of a coregistered raster file
containing forest, crops/grasses, water, buildings, and
roads.
Following data quality assessment, the reference data sets
were used as the basis for accuracy assessment. Tree cover
was assessed using the natural features data set and imper-
vious cover with the planimetric data. A combination of
random selection and reserved sample sets was used, as well
as independent interpretation of imagery and field visits.
These were derived differently for tree and impervious
cover, as described below.
3.3. Tree cover mapping
Having the natural features data set available for train-
ing the classification algorithm allowed for 1.7 million
unique randomly generated samples to be used in a land
cover stratification sampling scheme from the training
data, and retention of a proportional number of samples
in relation to the training data. Of the 1.7 million points,
600,000 were randomly removed from the data set to be
used as mutually exclusive validation points. The number
of training samples allocated to each panel depended on
the respective area of the panel, with the smallest receiving
70,000 and the largest receiving 250,000 samples. The
sample size also varied between forests, agricultural lands,
water, and urban features. Virtually no urban or suburban
trees or grasses were used in the training data because
these managed areas have high vegetation indices that
create statistical confusion with, for example, deciduous
forest spectra.
We used a decision tree classifier, or more specifically
a classification tree algorithm, within the statistical soft-
ware package S-PLUS, developed by Insightful. This
univariate decision tree algorithm, based on Breiman,
Freidman, Olshend, and Stone (1984), recursively thresh-
olds the training data into increasingly homogeneous
partitions using nonparametric rules. Decision trees have
become popular for land cover mapping (Friedl & Brod-
ley, 1997; Hansen, Dubayah, & DeFries, 1996; McCauley
& Goetz, in press) because the tree output is intuitive,
with each IKONOS band or variable threshold listed for
each successive partition. The decision tree classifier in S-
PLUS works by evaluating individual sample points, so
the values of the IKONOS pixels for the original bands,
and the derived indices and ratios were used.
Because temporal differences in the IKONOS scene
acquisition dates (Table 2) caused type confusion in early
classifications, leaf-on and leaf-off imagery required the
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208 199

development of more scene-specific decision trees based
on sample points from each individual scene only. To aid
in the development of a robust classification, each of the
six IKONOS panels was classified individually, and the
resulting decision trees were used to classify its respective
IKONOS image by applying the threshold breakpoints
listed in the classification tree to the eight image variables
described above. The panels were then mosaicked back
together for the final map of the entire county. An
accuracy assessment was completed on the final classifi-
cation using the data points reserved from the initial
sampling of the natural features coverage, without dupli-
cation of sample selections.
3.4. Impervious area mapping
The impervious surface mapping was also done using a
classification tree approach, but made use of the planimetric
map of buildings and roads rather than the natural features
database. Upon inspection, it was apparent that the building
features had few errors but the transportation features
contained numerous errors of omission, particularly parking
lots. In order to evaluate the completeness of the data, the
county was divided into 689 blocks, each of which consists
of many thousands of IKONOS pixels. Each block was
checked for errors against recent DOQs and the imagery
itself. Areas developed since the creation of the planimetric
data were not included in the training data set. Of the 689
original blocks, 197 were deemed sufficient for algorithm
development and validation-based tests of spectral discrim-
ination with sample size. Approximately 600,000 data
samples (4-m pixels) resulted.
Samples within each block were coded either impervi-
ous or nonimpervious based on interpretation of the plani-
metric data. The sample was then divided into training and
validation data sets based on each swath of IKONOS
imagery, resulting in about one-fourth (140,366) being
reserved for validation. This site-based approach reduced
the influence of spatial autocorrelation and artificially high
validation statistics. Additional training data were derived
for nonimpervious features such as water bodies and wet-
lands based on the natural features coverage described
above.
Predictor variables were chosen based on an analysis of
impervious feature signatures and their vegetation indices.
As with the tree cover mapping, the set of IKONOS spectral
variables used for prediction consisted of the four optical
bands, the NDVI, and the simple ratio (NIR/red). Two
additional spectral ratios that showed promise in discrimi-
nating open soil plots from impervious features were also
used (NIR/blue and NIR/green). The spectral information
and impervious class for each of these data sets were used to
grow impervious surface classification trees. The resulting
tree-based algorithms were then applied to the corresponding
IKONOS image tile, and thematic accuracy statistics were
calculated.
3.5. Stream health analyses
Statistical analyses were performed on the impervious
and tree cover maps to determine whether there were
statistically significant differences between small water-
sheds (approximately equivalent to hydrological unit code
HUC 14), as rated by water quality experts into one of four
possible stream resource condition categories. These were
based on countywide baseline stream monitoring data,
including pH, dissolved oxygen, and temperature, among
others (Van Ness, Brown, Haddaway, Marshall, & Jordahl,
1997). This multiagency effort also assessed a number of
biological indicators based on fish and benthic macroinver-
tebrate surveys, and were used to define an index of
biological integrity (IBI) (Stribling, Jessup, White, Boward,
& Hurd, 1998). Of the 296 small watersheds in the county,
245 had been analyzed and represented by a categorical
stream resource quality rating of excellent, good, fair, or
poor. We refer to these as stream health ratings. About one-
third of these (93) was indicated as preliminary assess-
ments, which were analyzed as separate categories from the
completed assessments.
We examined the amount of tree cover within the
entire area of each watershed, as well as within 100-ft
riparian buffers (a size determined by independent buffer
monitoring and restoration activities) to explore whether
there were links to the stream health ratings. The stream
buffer was calculated using a hydrology vector layer
provided by the county. The amount of impervious
surface area within each watershed was also calculated
and expressed, as with tree cover, as a proportion of total
watershed area.
One-way analysis of variance (ANOVA) was used to
assess the relation of each variable to the stream health
ratings. Logistic regression was also used to explore the
relationship of stream health with the mapped variables
(i.e., impervious cover, watershed tree cover, and riparian
buffer tree cover). Logistic regression is a nonlinear and
nonparametric technique that uses maximum likelihood
estimation to predict the odds probability of categorical
dependent variables (in this case stream health rating).
This is unlike ordinary least squares regression estimation
of changes in a numerical dependent variable. The tech-
nique also has the advantage of not assuming normal
distributions or homoskedastic independent variables
(Menard, 1995). It is often used to test binary variables,
Table 3
Comparison of county area classified as tree or nontree using supervised
and unsupervised approaches
Classification type Type Percent of area
Unsupervised Tree 44.6
Nontree 55.4
Supervised Tree 49.7
Nontree 50.3
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208200

but also can be applied to multinomial dependents with
more classes, as in our case with ordinal ratings of stream
health.
4. Results
4.1. Tree cover
The extent of the tree cover classifications derived using
the supervised and unsupervised approaches is summarized
in Table 3. The unsupervised classification provided an
estimate of f5% (67 km
2
) more forest than the supervised
classification. Accuracy statistics generated for the initial
supervised and unsupervised classifications (Table 4) indi-
cate that both classifications were significantly better than
random, but the differences between the classified maps,
while not statistically significant ( p= 0.07), were substan-
tial enough to convince us to take a different approach (i.e.,
decision trees trained with a large sample of reference
data).
The tree cover map derived from the decision tree
approach (Fig. 3) was spatially detailed and boundaries
between scenes are virtually nonexistent, despite the range
of dates in the IKONOS acquisitions in both years and
seasons. This was due, in large part, to the accuracy of the
decision tree algorithm and the abundance of training data.
Trees occupied nearly 38% of the county (Table 5). Areas
obscured by cloud or cloud shadow, which could not be
correctly classified, comprised 7% of the total county area.
The original forest classification done ca. 1992, and
subsequently manipulated for use as training data for our
mapping, depicted just 28% of the county as forested
(36,849 ha). The primary reason for the difference was the
depiction of a forest/nonforest classification in the natural
Fig. 3. IKONOS tree cover map of Montgomery County, MD. Note the forest cover data (from natural features GIS coverage; Fig. 4) used in the areas obscured
by clouds.
Table 4
Accuracy assessment of supervised and unsupervised classifications
Classification Overall
accuracy
jstatistic Zstatistic
Supervised 0.86 0.72 15.7
Unsupervised 0.83 0.67 13.3
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208 201

features layer, whereas use of IKONOS permitted us to
capture, in many cases, individual trees. A subset of the
IKONOS tree cover map, along with the reference data
(natural features coverage), shows differences between the
two mapping approaches (Fig. 4), where tree cover is
shown in lighter green compared to the reference forest
coverage (darker green). Note the missing trees in the
subdivisions and between the links of the golf course in
the lower left. It would be possible, if desired, to filter the
tree cover map in such a way as to approximate the forest
cover map (e.g., by retaining only those areas with at least
five of the surrounding eight pixels in a 3 3 window
classified as tree).
Accuracy statistics based on the independent 600,000
reference data samples reserved for validation show very
high map accuracies, with omission errors of 2.1% and
commission errors of 4.1% (Table 6). Overall classification
accuracy was 97.3% (j= 0.95).
Fig. 4. IKONOS image detail with classification results. (a) IKONOS imagery; (b) 1992 natural features coverage of forest areas; (c) tree cover derived from
imagery using a decision tree classifier; and (d) IKONOS image with natural features coverage and derived tree cover maps overlaid. Note the small patches of
trees in (c) compared to the forest cover in (b) and the substantial amount of tree cover missing from the forest area coverage (d). (Includes material from Space
Imagingn.)
Table 5
Classification of county using decision tree classifier
Classified type Area (ha) Percent of county
Tree 49,522 37.7
Nontree 72,268 55.1
Cloud 9478 7.2
Total 131,268 100
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208202

4.2. Impervious surfaces
The impervious surface map showed a level of detail
sufficient to map individual houses within subdivisions,
as well as clearly depicting areas where residential
development had occurred since the original planimetric
coverage was produced (Fig. 5). This map of impervious
areas depicts a wide range of materials, some of which
have very different spectral properties (e.g., pavement,
concrete, roof tiles, etc.). Other mapping methods based
on the use of regression trees, rather than classification
trees, permit estimation of imperviousness as a continu-
ous variable (i.e., as a percentage between 0 and 100).
We have used the planimetric reference data to train an
algorithm of continuous subpixel impervious maps based
on Landsat imagery (Smith et al., fourthcoming), but the
IKONOS impervious map could be used in place of
planimetric data.
Internal accuracy assessment checks are provided by
the decision tree software, including estimates of misclas-
sification rates for each intermediate and terminal node,
as well as an overall algorithm assessment. Additional
checks of misclassification error were done by cross-
validating the algorithm using 10% of the data samples
in an iterative procedure, cycling through the full data set
(after Friedl, Brodley, & Strahler, 1999). Analysis of
these accuracy measurements for various tree-growing
scenarios strengthened the final algorithm selection and
the overall classification accuracies. Independent statisti-
cal validation using the 140,366 samples reserved from
the image sampling approach demonstrated the ability of
the algorithm to discriminate impervious from nonimper-
vious component materials in the imagery (Table 7).
Omission errors in the impervious map were 9.4% and
commission errors were 10.8%. Overall classification
accuracy was 84.2% (j= 0.36). Despite our efforts to
screen for mislabeled and omitted features (e.g., drive-
ways and sidewalks) in the planimetric data used for
training, some remained present in both the training and
Table 6
Accuracy assessment of tree cover map
Validation
sample
(natural
features)
Classified
(IKONOS)
Number
correct
Producer’s
accuracy
User’s
accuracy
Tree 212,703 217,224 208,351 98.0 95.9
Nontree 287,297 282,577 278,236 96.9 98.5
Fig. 5. Impervious surface areas (yellow) for a portion of Montgomery County derived from IKONOS imagery (right) compared to planimetric coverage
derived from aerial photographs (left), showing changes in residential development since the area was initially mapped (ca. 1993). IKONOS tree cover (green)
is also shown, with the IKONOS panchromatic image as a backdrop. (Includes material from Space Imagingn.)
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208 203

validation data samples, which produced somewhat lower
accuracy statistics than expected. We believe, based in
part on results such as those shown in Fig. 5, that the
accuracy of the map is actually higher than that sug-
gested by the interpretation of planimetric data.
4.3. Riparian buffers and stream health
Average tree cover within riparian buffers ranged from
as little as 6% to over 98% by watershed, impervious cover
within the watersheds ranged from 0% to 43%, and total
tree cover (both inside and outside buffers) ranged between
6% and 94%. This range of observations was more than
sufficient to detect the sensitivity of the stream health to
buffer quality. Differences between watersheds were also
visually evident in map comparisons for these same
variables (Fig. 6), as well as in the statistics derived for
each rating category (Table 8 and Fig. 7). ANOVA showed
significant differences between watersheds with different
Fig. 6. Maps of small watershed aggregated values for (a) stream health rankings, (b) impervious surface cover (%), (c) tree cover (%), and (d) riparian buffer
tree cover (%). (b) – (d) were calculated using the IKONOS-derived maps of these land cover variables.
Table 7
Accuracy assessment of impervious surfaces map
Validation
sample
(planimetric)
Classified
(IKONOS)
Number
correct
Producer’s
accuracy
User’s
accuracy
Impervious 108,730 110,466 98,513 90.6 89.2
Nonimpervious 31,636 29,900 19,683 62.2 65.8
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208204

amounts of impervious surface area, tree cover, and ripar-
ian tree buffers. Variance ratios ( Fstatistics) were 42.1,
16.3, and 15.0, respectively, all significant at p< 0.001
(N= 245). Streams with excellent ratings had significantly
lower impervious cover and higher tree cover, both within
the watershed and riparian buffers, than streams rated
good, fair, or poor. This trend continued progressively
through each rating category, with increasing average
impervious area and decreasing tree cover as stream health
decreased. All pairwise comparisons confirmed statistically
significant differences ( p< 0.001) between stream rating
categories. There were no substantial differences between
those streams with preliminary assessments and those with
completed assessments, except in the case of those rated
excellent (n= 32) versus preliminary excellent (n= 6), but
this was likely a result of the small sample size of the
latter.
The goodness of fit test of Hosmer and Lemeshow
(1989) failed to reject the null hypothesis of no difference
between the observed and model-predicted values of stream
health; thus, the multivariate model was an adequate pre-
dictor of stream health rating. There is no variance-
explained statistic in logistic regression when the dependent
variable is categorical, but reduction in residual deviance
from the null model was highly significant (v
2
= 33.1,
p< 0.001). There was, however, multicollinearity among
variables in the regression. Across watersheds, the extent of
tree cover both within and outside of the 100-ft buffer was
positively correlated with each other (r=0.71), so the
properties of the 100-ft buffer were not independent of
the surrounding land cover. Impervious area and total tree
cover were inversely correlated (r=0.55), as were imper-
vious area and tree buffers (r=0.43), but these relation-
ships were reduced in agricultural areas, which had both
low tree and impervious cover. As implied by the ANOVA,
the individual independent variables in the logistic regres-
sion were each statistically significant, and stepwise inclu-
sion tests revealed that impervious cover was the primary
predictive variable, followed by tree cover within the
buffers, and then total tree cover. Residual deviance was
significantly reduced when riparian buffers were added to a
model based on impervious surfaces alone, but subsequent
addition of total watershed tree cover did not further
contribute to model significance. The results suggest that
guidelines for excellent stream health rating would be no
more than 6% impervious with at least 65% forested
buffers, and no more than 10% impervious with at least
60% buffered for a rating of good. These values would
clearly be affected by the landscape configuration of buffers
in relation to impervious areas, and by mitigation measures
like sediment retention ponds.
5. Discussion
5.1. Mapping applications
Land cover classification in the study region using
IKONOS imagery with traditional supervised and unsu-
pervised approaches produced substantially different maps
and, for example, tree cover estimates. There were several
factors that impacted the classifications, including the
timing of data acquisition (thus phenological stage),
associated viewing and atmospheric conditions, the qual-
ity of training data (in the case of supervised classifica-
tion), and insufficient discrimination resulting from
limited IKONOS spectral bands. Some of the more
pronounced classification errors were between wet agri-
cultural fields, dense forests, and shadows. For example,
spring small grains (wheat, barley, and rye) and hay
(alfalfa and grasses) had emerged in agricultural fields,
which produced a similar spectral response as deciduous
trees in partial leaf flush. In other areas, the visibility of
the forest floor through the tree canopy produced spectral
responses similar to residential areas. Shadows were also
an especially important consideration with the use of this
high spatial resolution imagery, particularly shadowing
within the forest canopy from adjacent trees. Some of
these issues can be reduced through the use of multi-
temporal imagery, but this would be logistically difficult
Table 8
Small watershed sample size and average statistics by stream health rating
category
Stream health
rating
nArea
(km
2
)
Impervious
(%)
Tree
cover (%)
Buffered
(%)
Excellent 38 272 3.6 50.6 76.8
Good 81 658 4.9 44.6 71.3
Fair 76 451 13.9 37.0 63.2
Poor 50 356 19.5 29.6 56.3
Fig. 7. Stream health rankings in relation to (a) impervious surface cover,
(b) watershed tree cover, and (c) riparian buffer tree cover, each derived
from the IKONOS image data.
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208 205

to accomplish in many areas, and cost-prohibitive for
most applications using IKONOS.
In order to reduce the impact of some of these issues,
the decision tree classification approach we adopted made
use of a large number of reference (training) data for the
study area. The resulting classification reinforced the
considerations noted above regarding the use of IKONOS
imagery for vegetation mapping. Two classification issues
were especially problematic, each dependent on the timing
of scene acquisition. In the early spring images (Panels 4
and 7) (see Table 1 and Fig. 1), the primary source of
classification error was confusion between mature decidu-
ous forests and bare ground areas such as untilled agri-
cultural or partially vegetated abandoned areas. In these
scenes, mature forests had just begun to leaf out, and
undergrowth, bare ground, and litter beneath the trees
were spectrally similar to agricultural fields with crop
residue still in place from no-till farming practices. In
the spring and early summer images (Panels 1– 3 and 5),
the primary confusion occurred between active agricultural
fields and deciduous trees. Despite these issues, we were
able to achieve accurate maps of tree and impervious
cover.
Although the above points are not unique to IKONOS
imagery, they are accentuated by the high spatial resolu-
tion relative to other imaging sensors such as Landsat
TM. In the case of the early season deciduous forest, the
4-m IKONOS resolution included data in the intercanopy
space, whereas with TM, the pixels include a mixture of
canopy and higher biomass woody stems, generally
producing higher NDVI values. In the case of the leaf-
on scenes, the high vegetation index values of actively
growing agricultural fields appeared spectrally similar to
that of the dense deciduous tree canopy. This was partly
due to greater spatial variability in IKONOS scene
elements. For example, a single tree may be contained
within one TM pixel, whereas the same tree may be
represented by 5 or 10 IKONOS pixels, resulting in
greater variation with foliage density and canopy shape
(see also Blonski, 2001; Goward, Davis, Fleming, Miller,
& Townshend, this issue). Inclusion of the ARVI image
helped reduce this effect, but cannot completely resolve
this unique aspect of high spatial resolution imagery.
Clearly, this ‘‘issue’’ could be considered an advantage
for some applications (e.g., assessments of defoliation or
tree health). There have also been suggestions that high-
resolution imagery could provide information capable of
discriminating different habitat types and complex forest
associations (e.g., oak/hickory versus beech/maple), or
permit mapping of invasive species such as Kudzu. Based
on our results with IKONOS, we believe that these
applications would be difficult to accomplish within
reasonable confidence limits.
The capability of IKONOS to point in a desired
direction adds to the flexibility in acquisition strategies,
and also has implications for the spectral properties of
image components. The relationship between sun, surface,
and sensor for our image acquisitions (Table 1) shows that
Panels 0 and 1 were acquired on the same orbit, but the
sensor azimuth angles differed. As a result, many of the
water bodies in Panel 0 show the effect of sun glint,
whereas those in Panel 1 do not. Similar, albeit less
dramatic, conditions occur in areas that are mostly vege-
tation or impervious cover types. This type of information
can be useful for analyses of bidirection reflectance prop-
erties, but complicates image classification for many appli-
cations, particularly where many IKONOS tiles or
overlapping image panels are required to completely map
a study area.
The work reported here also permitted algorithm devel-
opment for application to other IKONOS imagery in the
region, and significantly allowed for the development of
subpixel mapping techniques using Landsat imagery,
which we will report on in future publications (e.g., Smith
et al., fourthcoming).
5.2. Stream health application
Riparian buffers have been recognized as important
landscape features that provide unique habitat for many
wildlife species (Iverson, Szafoni, Baum, & Cook, 2001),
as well as filtering capabilities for removing nutrient
pollutants from agricultural runoff before they reach water-
ways (Cooper, Gilliam, Daniels, & Robarge, 1987; Cor-
rell, 1997; Lowrance et al., 1997). Traditional approaches
to mapping and monitoring riparian zone vegetation have
relied on photographic interpretation (e.g., Lonard, Judd,
& Desai, 2000), but this is not practical over large areas.
For example, the Chesapeake Bay Program has been
tasked with establishing 2000 miles of forested riparian
buffers by the year 2010, but does not have a practical
methodology to accurately assess current buffer statistics,
let alone monitor new plantings. Landsat data are simply
not of sufficient spatial resolution to adequately map
riparian buffer vegetation within the widely accepted
100-ft ( f30 m) buffer width used as a common reference
for buffer effectiveness.
Our analyses of impervious surfaces and tree cover
within the small watersheds and riparian buffer zones of
Montgomery County provide an example of the capabil-
ities conveyed by consistent mapping over large areas
using IKONOS imagery. A key advantage to the IKO-
NOS-derived maps was the fine spatial resolution that
allowed very-local-scale analysis of riparian buffers, and
statistically meaningful sample populations within the
small watersheds. The results confirm that there are clear
linkages between the land cover within a watershed and
the stream water quality. Specifically, the amount of
impervious surface area within a watershed and tree cover
within riparian buffer zones provided robust indicators of
stream health rating. This suggests that the buffers were
functioning to reduce pollutants and sediments from built
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208206

areas before they entered the waterways. Conversely,
buffers alone were insufficient for protecting waterways
in highly impervious areas or where forest cover outside
buffer zones was low, including agricultural areas that
presumably affected stream health through chemical run-
off. Additional analyses beyond the scope of this paper
are being completed to explore these relationships further,
and to assess the effectiveness of mitigation approaches in
controlled and uncontrolled paired watersheds.
The high-resolution land cover maps that can be derived
from IKONOS imagery and the associated landscape
variables that can be calculated from them (e.g., buffer
configuration) thus provide an important contribution to
these types of resource management applications. In the
case of Montgomery County, these statistics are actively
used to determine whether resources are focused on
watershed preservation (streams of excellent health), pro-
tection (good or excellent health), or restoration (fair or
poor health).
6. Conclusions
Utilizing high spatial resolution imagery can be benefi-
cial to many different resource management applications.
Imagery like IKONOS can aid in the development of a wide
range of mapping and spatial modeling applications, but a
number of issues must be considered with these relatively
new and unique data sets. The issues include (i) program-
matic considerations of the timing of acquisitions to capture
features of interest (e.g., leaf-off for impervious versus leaf-
on for tree cover), as well as timely data acquisition (ours
took 18 months to acquire); (ii) technical considerations,
such as compensation for the effects of object shadowing
resulting from improved resolution of individual scene
elements, and limited spectral resolution and range; and
(iii) economic factors, such as cost per square kilometer
relative to DOQs, and feasibility of consistent repeat
acquisitions. For example, we noted an inability of IKO-
NOS to adequately spectrally discriminate some land cover
types due to high spatial variability within scene elements
resulting from variable illumination and viewing conditions.
Spectral variability within scene objects also contributed to
reduced class type discrimination between more generalized
land cover types (e.g., confusion between deciduous forest
and some agricultural crops).
New sets of interpretation strategies need to be devel-
oped to maximize the information obtained from IKONOS,
while minimizing the problematic issues specific to high
spatial resolution imagery. Our exploration of several such
approaches relevant to resource mapping applications sug-
gests the great practical utility of IKONOS imagery, partic-
ularly for impervious surface, tree cover, and riparian
buffers, all of which are related to stream health. Our results
provided very specific guidelines for predicting stream
health ratings, which allows for targeted and adaptive
protection and restoration management decision making.
Other applications for which the maps are currently being
utilized include resource lands monitoring and acquisition
planning, analyses of habitat connectedness and bird habitat
suitability, and hydrodynamic modeling for mitigation of
stream bank erosion. IKONOS imagery can augment tradi-
tional mapping approaches and, in many cases, may provide
a cost-effective alternative.
Acknowledgements
This work was funded by NASA’s Earth Science
Applications Division (grants NAG1302010 and NAG-
1399011). We acknowledge Montgomery County M-
NCPPC (Mary Dolan and Nazir Baig) and Montgomery
DEP (Cameron Wiegand) for data sets and collaborative
assistance. IKONOS image data were acquired by Space
Imaging and provided by the NASA Scientific Data
Purchase program.
References
Birk, R.J., Stanley, T., & Snyder, G.L. (2003). U.S. Government commercial
data purchase programs for use in government-sponsored science and
applications activities. Remote Sensing of Environment,88, 3–16.
(doi:10.1016/S0034-4257(03)00227-X)
Blonski, S. (2001). IKONOS-based simulations of Landsat 7 VNIR data:
Comparison with actual, coincident images. In Proceedings of the 2001
High Spatial Resolution Commercial Imagery Workshop, Greenbelt,
MD, USA, March 19 – 21 (sponsored by NASA/NIMA/USGS Joint
Agency Commercial Imagery Evaluation Team) (CD-ROM).
Breiman, L., Freidman, J., Olshend, R., & Stone, C. (1984). Classification
and regression trees. Monterey, CA: Wadsworth.
Cooper, J. R., Gilliam, J. W., Daniels, R. D., & Robarge, W. P. (1987).
Riparian areas as filters for agricultural sediment. Soil Science Society of
America Journal,51, 416 – 420.
Correll, D. L. (1997). Buffer zones and water quality protection: General
principles. In N. E. Haycock, T. P. Burt, K. W. T. Goulding, & G. Pinay
(Eds.), Buffer zones: their processes and potential in water protection
( pp. 7– 20 ). Harpenden, UK: Quest Environmental.
Dial, G.F., Bowen, H., Gerlach, B., Grodecki, J., & Oleszczuk, R. (2003).
IKONOS satellite, sensor, imagery, and products. Remote Sensing of
Environment,88, 23 – 36. (doi:10.1016/S0034-4257(03)00229-3)
Fisher, J., & Goetz, S. J. (2001). Considerations in the use of high spatial
resolution imagery: An applications research assessment. In American
Society for Photogrammetry and Remote Sensing (ASPRS) Conference
Proceedings, St. Louis, MO (p. 8), Available from ASPRS http://
www.asprs.org and at http://www.geog.umd.edu/resac.
Friedl, M. A., & Brodley, C. E. (1997). Decision tree classification of land
cover from remotely sensed data. Remote Sensing of Environment,61(3),
399 – 409.
Friedl, M., Brodley, C., & Strahler, A. (1999). Maximizing land cover
classification accuracies produced by decision trees at continental to
global scales. IEEE Transactions on Geoscience and Remote Sensing,
37(2), 969– 981.
Goetz, S. J. (1997). Multi-sensor analysis of NDVI, surface temperature
and biophysical variables at a mixed grassland site. International Jour-
nal of Remote Sensing,18(1), 71 – 94.
Goward, S.N, Davis, P.E., Fleming, D., Miller, L., & Townshend, J.R.G.
(2003). Empirical comparison of Landsat 7 and IKONOS multispectral
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208 207

measurements for Selected Earth Observation System (EOS) Validation
Sites. Remote Sensing of Environment,88, 79 – 98. (doi:10.1016/j.rse.
2003.06.007)
Hansen, M., Dubayah, R., & DeFries, R. S. (1996). Classification trees: An
alternative to traditional land cover classifiers. International Journal of
Remote Sensing,17(5), 1075 – 1081.
Helder, D., Coan, M., Patrick, K., & Gaska, P. (2003). IKONOS geometric
characterization. Remote Sensing of Environment,88, 68 – 78.
(doi:10.1016/S0034-4257(03)00232-3)
Hosmer, D. W., & Lemeshow, S. (1989). Applied logistic regression. New
York: Wiley.
Iverson, L. R., Szafoni, D. L., Baum, S. E., & Cook, E. A. (2001). A
riparian wildlife habitat evaluation scheme developed using GIS. Envi-
ronmental Management,28(5), 639 – 654.
Jantz, C. A., Goetz, S. J., & Shelley, M. A. (2003). Using the SLEUTH
urban growth model to simulate the land use impacts of policy scenarios
in the Baltimore – Washington metropolitan region. Environment and
Planning, B (in press).
Kaufman, Y. J., & Tanre, D. (1992). Atmospherically Resistant Vegetation
Index (ARVI) for EOS-MODIS. IEEE Transactions on Geoscience and
Remote Sensing,30, 260 – 271.
Lonard, R. I., Judd, F. W., & Desai, M. D. (2000). Evaluation of color-
infrared photography for distinguishing annual changes in riparian for-
est vegetation of the lower Rio Grande in Texas. Forest Ecology and
Management,128(1/2), 75 – 82.
Lowrance, R., Altier, L. S., Newbold, J. D., Schnabel, R. R., Groffman,
P. M., Denver, J. M., et al. (1997). Water quality functions of ripar-
ian forest buffer systems in the Chesapeake Bay watershed. Environ-
mental Management,21, 687 – 712.
McCauley, S., & Goetz, S. J. (2003). Mapping residential density patterns
using multi-temporal Landsat imagery and a decision-tree classifier.
International Journal of Remote Sensing (in press).
Menard, S. (1995). Applied logistic regression analysis. Quantitative Ap-
plications in the Social Sciences, 106. Thousand Oaks, CA: Sage Pub-
lications.
M-NCPPC (2000). Legacy Open Space functional master plan, Mont-
gomery County Department of Park and Planning, Silver Spring,
MD, 80 pp. plus appendices and maps (http://www.mc-mncppc.org/
legacy_open_space).
Sawaya, K., Olmanson, L., Holden, G., Sieracki, J., Heinert, N., & Bauer,
M. (2003). Extending satellite remote sensing to local scales: Land and
water resource management using high resolution imagery. Remote
Sensing of Environment,88, 143 – 155. (doi:10.1016/j.rse.2003.04.006)
Smith, A. J., Goetz, S. J., Prince, S. D., Wright, R., Melchoir, B., Mazza-
cato, E. M., & Jantz, C. (fourthcoming). Estimation of sub-pixel imper-
vious surface area using a decision tree approach, IKONOS and Landsat
imagery. Remote Sensing of Environment.
Stoney, W. (2001). The past, present, and future of the medium- and high-
resolution satellite world. In: J.C. Baker, K.M. O’Connell, & R.A.
Williamson (Eds.), Commercial observation satellites: At the leading
edge of global transparency (pp. 652 – 661 ). Santa Monica, California:
RAND Corporation and the American Society for Photogrammetry and
Remote Sensing. (Available online at http://www.rand.org/publications/
MR/MR1229/)
Stribling, J. B., Jessup, B. K., White, J. S., Boward, D., & Hurd, M. (1998).
Development of a benthic index of biotic integrity for Maryland
streams. Annapolis: Montgomery Department of Natural Resources,
62 pp.
Tanaka, S., & Toshiro, S. (2001). A new frontier of remote sensing from
IKONOS images. International Journal of Remote Sensing,22(1),
1–5.
Van Ness, K., Brown, K., Haddaway, M., Marshall, D., & Jordahl, D.
(1997). Montgomery County water quality monitoring program: Stream
monitoring protocols. Watershed Management Division, Department of
Environmental Protection, Silver Spring, Maryland.
Varlyguin, D., Wright, R., Goetz, S. J., & Prince, S. D. (2001). Ad-
vances in land cover classification for applications research: A case
study from the mid-Atlantic RESAC. American Society for Photo-
grammetry and Remote Sensing (ASPRS) Conference Proceedings,
St. Louis, MO . (p. 7).
Weber, T., & Wolf, J. (2000). Maryland’s green infrastructure—using land-
scape assessment tools to identify a regional conservation strategy.
Environmental Monitoring and Assessment,63, 265 – 277.
S.J. Goetz et al. / Remote Sensing of Environment 88 (2003) 195–208208