ArticlePDF Available

Assessing The Use Of Multiseason Quickbird Imagery For Mapping Invasive Species In A Lake Erie Coastal Marsh


Abstract and Figures

QuickBird multispectral satellite images taken in September 2002 (peak biomass) and April 2003 (pre-growing season) were used to map emergent wetland vegetation communities, particularly invasive Phragmites australis and Typha spp., within a diked wetland at the western end of Lake Erie. An unsupervised classification was performed on a nine-layer image stack consisting of all four spectral bands from both dates plus a September Normalized Difference Vegetation Index image. The resulting eight cover classes distinguished three monodominant genera (Phragmites australis, Typha spp., Nelumbo lutea), three multigenera plant communities (wet meadow, other non persistent emergents, woody vegetation), and two unvegetated cover types (water, bare soil). Field validation at 196 data points yielded an overall classification accuracy of 62%, with producer’s accuracy for the eight individual classes ranging from 41 to 91% and user’s accuracy from 17 to 90%. Three-fourths of areas designated as Phragmites were correctly mapped, but 14% were found to be cattail (Typha) during field validation. Lotus (Nelumbo lutea) beds were accurately mapped on multiseason imagery (producer’s accuracy = 91%); these beds had not yet emerged above water in April, but were fully developed in September. Other types of non persistent vegetation were confused with managed areas in which vegetation had been cut and burned to control invasive Phragmites. Multiseason QuickBird imagery is promising for distinguishing certain wetland plant species, but should be used with caution in highly managed areas where vegetation changes may reflect human alterations rather than phenological change.
Content may be subject to copyright.
Dana M. Ghioca-Robrecht
, Carol A. Johnston
, and Mirela G. Tulbure
Department of Biology and Microbiology, South Dakota State University
Brookings, South Dakota, USA 57007
P.O. Box 4274
Roanoke, Virginia, USA 24015
Abstract: QuickBird multispectral satellite images taken in September 2002 (peak biomass) and April
2003 (pre-growing season) were used to map emergent wetland vegetation communities, particularly
invasive Phragmites australis and Typha spp., within a diked wetland at the western end of Lake Erie. An
unsupervised classification was performed on a nine-layer image stack consisting of all four spectral
bands from both dates plus a September Normalized Difference Vegetation Index image. The resulting
eight cover classes distinguished three monodominant genera (Phragmites australis,Typha spp., Nelumbo
lutea), three multigenera plant communities (wet meadow, other non persistent emergents, woody
vegetation), and two unvegetated cover types (water, bare soil). Field validation at 196 data points
yielded an overall classification accuracy of 62%, with producer’s accuracy for the eight individual classes
ranging from 41 to 91%and user’s accuracy from 17 to 90%. Three-fourths of areas designated as
Phragmites were correctly mapped, but 14%were found to be cattail (Typha) during field validation.
Lotus (Nelumbo lutea) beds were accurately mapped on multiseason imagery (producer’s accuracy 5
91%); these beds had not yet emerged above water in April, but were fully developed in September. Other
types of non persistent vegetation were confused with managed areas in which vegetation had been cut
and burned to control invasive Phragmites. Multiseason QuickBird imagery is promising for
distinguishing certain wetland plant species, but should be used with caution in highly managed areas
where vegetation changes may reflect human alterations rather than phenological change.
Key Words: Phragmites, remote sensing, Typha, wetland mapping
Satellite remote sensing has many advantages for
mapping wetlands, including frequent acquisition,
repeat coverage for monitoring changing conditions,
and low image cost in comparison to high-altitude
photography (Ozesmi and Bauer 2002). The ability
to remotely identify dominant wetland plant species
is desirable, because plant species are indicators of
wetland condition (Johnston et al. 2007a). It would
be particularly useful to identify the presence and
spread of invasive plant species that displace native
vegetation and degrade wetland habitat values
(Madden 2004). One such species is common reed
(Phragmites australis), which has been the subject of
remote sensing research in coastal brackish marshes
(Bachman et al. 2002, Artigas and Yang 2006) as well
as coastal freshwater marshes of the North American
Great Lakes (Arzandeh and Wang 2003, Wilcox et
al. 2003, Lopez et al. 2004, Pengra et al. 2007).
The identification of non persistent emergents
would also be desirable, because non persistent
coastal wetlands provide important faunal habitat
(Burton et al. 2004, Brazner et al. 2007). As defined
by the U.S. National Wetlands Inventory, non
persistent emergents are a subclass ‘‘dominated by
plants which fall to the surface of the substrate or
below the surface of the water at the end of the
growing season so that, at certain seasons of the
year, there is no obvious sign of emergent vegeta-
tion’’ (Cowardin et al. 1979). Examples of non
persistent emergents include arrow arum (Peltandra
virginica (L.) Schott), pickerelweed (Pontederia
cordata L.), and arrowheads (Sagittaria). Non
persistent wetlands were poorly mapped by the
National Wetlands Inventory due to its use of ‘‘leaf-
off’’ aerial photography that was taken too early in
the growing season to detect non persistent emer-
gents (Johnston and Meysembourg 2002). Accurate
remote identification of non persistent emergents
requires images representing both mature and
senescent vegetation periods.
Imagery from early satellite sensors such as
Landsat (30 m resolution) was inadequate for many
WETLANDS, Vol. 28, No. 4, December 2008, pp. 1028–1039
2008, The Society of Wetland Scientists
wetland mapping applications due to coarse spatial
resolution, but newer finer-resolution imagery offers
promise for more detailed classification of wetland
vegetation. Multispectral SPOT (10 m and 20 m
resolution) or IKONOS images (4 m resolution)
have been used to identify multiple classes of
emergent wetlands, with mapping accuracies of
79.5 to 85.1%(Rutchey and Vilchek 1999, Sawaya
et al. 2003, Phillips et al. 2005). Airborne hyper-
spectral imagery (5 m resolution) improved mapping
of Phragmites and Typha in Great Lakes coastal
wetlands, with estimated accuracies of 80%based on
photointerpreted aerial imagery and 91%based on
field measurement data (Lopez et al. 2004).
Multiseason remote sensing has aided discrimina-
tion of wetland types by detecting hydrological and
phenological changes characteristic of those types
(Jensen et al. 1993b, Sersland et al. 1995, Wolter et
al. 2005, Baker et al. 2006). The advantage of using
multiseason imagery is that it provides additional
classification information for distinguishing plant
species within a single growing season. This
information is especially important given that many
wetland species have overlapping spectral reflec-
tances at peak biomass (Ernst-Dottavio et al. 1981,
Spanglet et al. 1998, Schmidt and Skidmore 2003).
To the best of our knowledge, QuickBird imagery,
with resolution of less than 3 m, has not been
previously used for mapping individual wetland
species. The goal of this research was to evaluate the
use of multiseason QuickBird imagery for mapping
emergent wetland vegetation. Our primary objective
was to distinguish invasive emergent species that
occurred in monodominant stands, including Phrag-
mites australis,andTypha spp. A secondary
objective was to distinguish persistent from non
persistent emergents.
Study Site
Erie Marsh (41u459050N, 83u279180W) is located
at the west end of Lake Erie in North Maumee Bay,
Monroe County, Michigan (Figure 1). This 918 ha
area is situated 10 km north of Toledo, Ohio, and
70 km south of Detroit, Michigan. Major land
holdings include the Michigan Department of
Natural Resources’ Erie State Game Area and the
Erie Marsh Preserve, an area protected by The
Nature Conservancy (TNC). Erie Marsh represents
11%of the remaining marshland in southeastern
Michigan and is one of the largest marshes on Lake
Erie (TNC 2006). Two state-threatened vascular
species, American lotus (Nelumbo lutea Willd.) and
swamp rose-mallow (Hibiscus moscheutos L.), are
found on the preserve. The marsh also serves as a
migratory and nesting area for shorebirds and
The hydrology of Erie Marsh has been altered
over the past century by construction of dikes for
water level control that was initiated by waterfowl
hunting clubs in the early 1900s (Johnston et al.
2007b). Although the dikes alter the natural
dynamics of water level change in Erie Marsh, they
also give wetland managers greater control over the
vegetation within the dikes. The Nature Conser-
vancy’s management plan for controlling invasive
Phragmites within the Erie Marsh Preserve involves
a sequence of draining, prescribed burning, herbi-
ciding of Phragmites, and reflooding (NOAA 2006).
Satellite Imagery
We utilized two QuickBird images of the Erie
Marsh Area: an early fall image (6 September 2002),
when vegetation was at peak biomass, and an early
spring image (10 April 2003) when deciduous trees
were leafless and non persistent vegetation was
absent (Figure 2). The QuickBird images covered a
10 km 310 km area, with a pixel size of 2.8 m 3
2.8 m. The QuickBird images were orthorectified by
the vendor to 1:24,000 base maps with RMS error of
7.7 m. The projected coordinate system was UTM
Zone 17, North American Datum 1983. Each image
Figure 1. Study area and reference point locations.
includes four layers, corresponding to blue, green,
red, and near-infrared (NIR) wavelengths. The
QuickBird images were initially acquired for a
previous study of submergent aquatic vegetation
(Wolter et al. 2005).
An Area of Interest (AOI) mask was created to
exclude non-wetland areas (Figure 1). The lakeward
side of the mask was determined digitally by
computing the Normalized Difference Vegetation
Index (NDVI) for the September image and
selecting the NDVI threshold between water and
vegetation, which was used with the ERDAS
Imagine 8.7 (LEICA Geosystems, Norcross, GA,
USA) CLUMP command to generate a GIS file of
contiguous water pixels. The pixel clump represent-
ing the waters of Lake Erie and Maumee Bay was
converted into a water mask polygon. The landward
side of the mask (i.e., the upland-wetland boundary)
was digitized manually based on visual interpreta-
tion by an experienced aerial photo interpreter
(Johnston et al. 2007b), because upland areas could
not be distinguished from wetland areas by a simple
elevation threshold. This process masked out such
features as the J. R. Whiting Power Plant and
associated facilities, seasonal hunting cabins owned
by the Erie Shooting and Fishing Club, coastal
residential areas, marinas, upland forest on Indian
Island, and agricultural lands outside of the diked
areas. Crops planted to attract waterfowl within
diked portions of Erie Marsh were not masked out.
Although paved roads in areas surrounding Erie
Marsh were masked out of the image, dirt roads on
top of dikes were not. The AOI thus created
contained shallow water areas in the interior of Erie
Marsh, excavated ponds on the western edge of Erie
Marsh, the Woodtick Peninsula south of the
Whiting Power Plant fly ash pit, dikes and their
associated dirt roads, and emergent and woody
Unsupervised Classification
An eight-layer stack of all four bands from both
dates was initially used for image analysis. During
test runs, dark tree shadows appearing on the
September image were sometimes misclassified as
water, so we added the September NDVI layer as a
ninth layer to alleviate this problem (tree shadows
were not a problem on the April image due to the
lack of deciduous foliage). Because we had minimal
knowledge of the site prior to image classification,
we opted to use unsupervised classification, which is
recommended when not much is known about the
data before classification (Leica Geosystems 2003).
We performed an unsupervised classification (ISO-
Figure 2. NIR band images of highly managed portion
of Erie Marsh showing selected types of vegetation change
during September (top image) and the following April
(bottom image). ‘‘Spring’’ is the location of a diked
sulphur spring visible on both images.
1030 WETLANDS, Volume 28, No. 4, 2008
DATA) with 100 classes in ERDAS Imagine 8.7,
selecting to initialize means along the Principal Axis
and limiting the processing operation to 25 itera-
tions or until it converged to the 0.95 threshold level.
The 100 ISODATA classes were grouped into
eight classes (Table 1) using a classification scheme
developed from ground reconnaissance conducted
contemporaneously with summer image acquisition
(28 August 2002) and on quantitative vegetation
cover data collected for the Great Lakes Environ-
mental Indicators project on 20 July 2002 and 25–26
June 2003 (Johnston et al. 2007a). The classification
distinguished three monodominant genera (Phrag-
mites australis,Typha spp., and Nelumbo lutea),
three multi-genera plant communities (wet meadow,
non persistent emergents, and woody vegetation),
and two unvegetated cover types (water and soil).
The Phragmites class was initially divided into two
water depth classes, but field measurements showed
that they could not be consistently distinguished and
we collapsed the two Phragmites classes into a single
Accuracy Assessment
The initial reconnaissance data were unsuitable
for accuracy assessment because they had been used
for image analysis and were non-random. Therefore,
we conducted field accuracy assessment of the maps
on 17–19 July 2005 (27 points) and 4–16 July 2006
(169 points). A minimum of 18 data points were
collected per class, with a greater number (51 points)
being collected for the species of greatest interest and
coverage, Phragmites. We developed a stratified
Table 1. Category names and descriptions used to classify wetlands, spring and summer field appearance, and area
mapped in each class at Erie Marsh, North Maumee Bay, Michigan.
Unvegetated or cultivated
1 Water Water Water, may contain floating or
submerged aquatic vegetation
Lemna minor L.,
Stuckenia pectinata
(L.) Bo¨erner
2 Soil Beach, bare moist
soil, stubble in
cultivated lands
Beach, sparsely vegetated moist
soil, crops in cultivated lands
corn (Zea mays L.) 54.2
Emergent, nonpersistent
Water Large, circular leaves that may
float on water surface or be on
emergent petioles; growing in
shallow water
Nelumbo lutea Willd. 33.5
4 - Other nonpersistent,
Water Various plant species that emerge
out of shallow water; may be
mixed with Nelumbo
Sagittaria latifolia Willd.,
Eleocharis R. Br.
Emergent, persistent
monodominant or
dominant mix
Dense standing and
fallen plant litter
Dense reeds with alternate leaves
on very tall (2–4 m) stems,
flowering heads plumelike;
growing in shallow water or
moist soil
Phragmites australis
(Cav.) Trin. ex Steud.
Dense standing and
fallen plant litter
Long, erect leaves emerging from
base of plant, stems 1.5–3 m,
flowering heads brown &
cylindrical; growing in shallow
water or moist soil
Typha angustifolia L.,
Typha xglauca Godr.
(pro sp.)
7 - Wet meadow,
Primarily fallen
plant litter
Various plant species growing in
moist soil; canopy height 1.5 m
or shorter
Phalaris arundinacea L.,
Carex L.
8 - Trees, shrubs Woody stems
without leaves
Woody stems with leaves Populus deltoides
Bartram ex Marsh.,
Salix L.
random sampling scheme, using ERDAS Imagine
and ArcView (ESRI, Redlands, CA, USA) to
generate random points. Each random point was
defined as a group of 3 33 neighboring pixels and
belonging to the same class. In the field these
random points represented approximately 8 m 3
8 m plots. The random locations were uploaded into
a handheld GPS unit with accuracy to within 30 cm
(Geo XH 2005 Series, Trimble Navigation Limited,
Sunnyvale, CA, USA). We then used the GPS unit
in the field to navigate to the reference points (i.e.,
accuracy assessment points) (Johnston et al. in
In the field, each plot was divided in four subplots
oriented towards each cardinal direction. For each
plot we recorded the elevation (m). Maximum water
depth (cm) was measured at most plots (166 out of
196). We estimated the vegetation/bare soil/water
cover (%) in each of the four subplots and calculated
average cover for each plot. Based on the dominant
cover type, one of the eight classes was assigned to
each plot. Some plots have several layers of
dominant vegetation on vertical structure; we
assigned classes based on the tallest vegetation layer
which likely would have determined the spectral
value of a pixel. For emergent vegetation (persistent
and non persistent) we used a threshold of 30%
vegetation cover (Cowardin et al. 1979).
To assess the accuracy of the classification, we
constructed an error matrix and computed the
overall, producer’s, and user’s accuracy values. We
computed average digital numbers (DN) for green,
red, and NIR bands for both QuickBird images to
assess the spectral differences among the reference
classes. We also calculated k
ˆ5(observed accuracy -
chance accuracy)/(1-chance accuracy), as an indica-
tor of the extent to which the percentage correct
values of an error matrix are due to the ‘‘true’’
agreement versus ‘‘chance’’ agreement (Congalton
and Green 1999, Lillesand and Kiefer 2000).
Wetland Classes and Reference Data
Human activity has altered most of the vegetation
within the 918 ha study area. Erie Marsh is actively
managed for waterfowl hunting, and grain crops are
planted on plowed fields within the marsh to attract
waterfowl (Figure 2). Trails and staging areas are
mowed within the planted crops for hunter access
into the interior of the wetland. Diked areas are
periodically flooded. Control efforts for Phragmites
include mowing, herbicide spraying, burning, and
reflooding. Such types of alteration are typical of
managed Lake Erie coastal marshes, but complicate
remote sensing of vegetation.
The eight classes utilized were ecologically fairly
distinct. In addition to differences in dominant plant
species (Table 1, Figure 3), these classes had differ-
ent water depths (Table 2). Sampled open water
areas and Nelumbo beds had water depths of about
50 cm, whereas other non persistent emergents grew
in water averaging 38 cm in depth. Water depth
ranges were similar for Typha and Phragmites,but
average water depth was different: Typha grew in
water averaging 16 cm in depth, whereas Phragmites
tended to grow in drier areas (mean water depth 5
5 cm). Sites with woody vegetation had little
standing water, and the wet meadows and bare
soil/cultivated areas had no standing water when the
field work was conducted in July. As mapped,
Phragmites was the most extensive class, covering
38%of the study area, followed by open water and
Typha (Table 1).
The eight classes were also spectrally distinct. The
spectral characteristics of the eight wetland classes
were very different in September versus April due to
the overwinter change from photosynthesizing
vegetation to dead plant litter or water, and the
average digital number (DN) in the NIR band was
much greater in September than it was in April
(Figure 4). The Nelumbo and non persistent emer-
gent classes changed the most, from dense vegeta-
tion in September to open water areas with very low
NIR values in April (Figures 4 and 5). The average
NIR DN for these two classes was about five times
greater in September than in April. Stands con-
firmed in the field to be Phragmites and Typha had
distinct September NIR DN values, with the average
DN values for Phragmites being 69%greater than
those of Typha.BothPhragmites and Typha had
similar NIR DN values in April. Woody vegetation
had September NIR DN values intermediate be-
tween those of Phragmites and Typha, and lower
September DN values in the visible light bands than
other vegetation types. The wet meadow class
differed from the other vegetation classes in April
due to its high DN values in the visible light bands.
We believe that this difference is due to the lack of
subcanopy water and the reflectivity of prostrate
grass detritus from the previous growing season. The
wet meadow class also had higher NIR DN values in
April than all other classes, which may be due to
earlier green-up of vegetation in these drier areas of
the wetland. The magnitude of change in the NIR
DN values of the soil/sand/cultivated class was
surprising, nearly three times greater in September
than in April. This change is attributed to the
growth of cultivated crops on portions of this
1032 WETLANDS, Volume 28, No. 4, 2008
Figure 3. Field photos of the eight cover types mapped. A) Soil/cultivated (plowed ground in foreground, corn in
background); B) Nelumbo beds (foreground) and woody vegetation (background); C) Sagittaria latifolia, a non persistent
emergent; D) Typha spp.; E) wet meadow dominated by Phalaris arundinacea;andF)Phragmites australis.
category (Figure 2); sandy beach areas on the
Woodtick Peninsula did not exhibit a similar
increase. Predictably, water had the lowest NIR
DN values of any class on both dates, and the two
non persistent emergent classes had April DN values
similar to that of water. The September NIR DN
value of the ‘‘water’’ class was twice greater than its
April value, which we attribute to the presence of
submergents, algae, and duckweed, a tiny free-
floating plant that covered some wind-protected
water areas in the marsh during September (e.g.,
area marked ‘‘Lemna’’ on Figure 2).
Accuracy Assessment
Assignment of field vegetation to a wetland cover
class for reference data was done without knowledge
of the mapped class that might bias the assignment.
An error matrix of the image analysis yielded an
overall accuracy of 62%and a k
ˆvalue of 0.542, but
map accuracy varied substantially among the eight
classes mapped (Table 3). Both user’s and produc-
er’s accuracies were .70%for water, soil/cultivated,
and woody vegetation, but there was moderate to
severe misclassification of the five emergent wetland
classes. Inspection of the imagery after field data
collection revealed a variety of reasons for these
All but two of the points mapped as water were
found to be water in the field (90%user’s accuracy);
the two erroneous points were located on a
boundary between water and cattail. Eight points
were found in the field to be water but not mapped
as such (70%producer’s accuracy). Inspection of the
September 2002 imagery showed that five of these
were clearly vegetated as of that date, so this ‘‘error’’
actually represents a change in conditions between
the image date and the date of the reference data.
The three points mapped as Nelumbo but found to
be water were all east of Gard Island, and visual
inspection of the imagery confirmed the mapped
classification (Figure 5). However, this area of
North Maumee Bay is experiencing active erosion,
so a change in wetland configuration near Gard
Island is likely. The points classified as soil and
Typha that were found to be water were within diked
areas that had clearly been flooded since the image
date. The producer’s accuracy for water was
increased to 86%by assuming that these five points
correctly portrayed conditions as of the image date,
and overall accuracy was raised to 64%(Table 3).
Table 2. Average water depth for the eight classes, based
on field measurements made at 166 reference points at
Erie Marsh, North Maumee Bay, Michigan.
depth (cm) SE Min Max N
1–Water 48.40 12.71 20.00 150.00 10
2–Soil 0.00 0.00 0.00 0.00 18
3–Nelumbo 51.00 9.80 20.00 80.00 5
4–Other non persistents 37.50 7.27 5.00 50.00 6
5–Phragmites 4.81 1.18 0.00 45.00 70
6–Typha 16.00 2.50 0.00 40.00 26
7–Meadow 0.00 0.00 0.00 0.00 9
8–Woody 1.59 1.37 0.00 30.00 22
Figure 4. Digital numbers (mean +SE) for green, red,
and NIR bands at points confirmed to be correctly
mapped within the eight wetland classes.
1034 WETLANDS, Volume 28, No. 4, 2008
The high level of accuracy for the soil/beach/
cultivated class was surprising given the diversity of
conditions represented by this class and its propen-
sity for change (Figure 2). All but two of the points
found in the field to be bare soil/cultivated were
correctly mapped (89%producer’s accuracy), and
inspection of those points suggested that cultivation
changes may have occurred between the time of
image acquisition and field checking. Five points
mapped as bare soil/cultivated were other classes in
the field. Excluding the diked flooded soil (see
above) yielded an 81%user’s accuracy. One of the
erroneous soil/cultivated points was at the edge of a
field, and the error could be due to expansion of
field boundaries or georeferencing error.
All of the 11 Nelumbo beds visited in the field were
correctly mapped except for one mapped as non
persistent (producer’s accuracy 591%). User’s
accuracy was much lower (50%). Already noted
was the problem of Nelumbo bed loss in the vicinity
of Gard Island. Five points classified as Nelumbo
beds were found in the field to be Phragmites; all five
had unvegetated April soils that were dark due to
wetness or burn scars. Two points classified as
Nelumbo were found to be other non persistents,
which is not considered to be a serious error.
The other non persistents class was quite inaccu-
rate (17%user’s accuracy, 50%producer’s accura-
cy). Only eight points were actually found to have
this type of vegetation in the field. Areas mapped as
non persistent based on the multiseason image were
usually found to be managed vegetation, where
mowing and/or raised water levels within diked
areas of the wetland caused the April image to
appear as water (e.g., areas marked as ‘‘burned’’ on
Figure 2). Nearly half of the points mapped as non
persistent were found to be Phragmites in the field,
and another three points were found to be Typha.
This class represents a disadvantage of using multi-
season imagery to infer non persistent vegetation:
anthropogenically caused vegetation changes cannot
be distinguished from natural phenological changes
by spectral reflectance alone.
Areas mapped as Phragmites were generally so in
the field (76%user’s accuracy), but Phragmites was
much more prevalent than expected (53%producer’s
accuracy). Most of the points erroneously mapped
as Phragmites were actually Typha (7 out of 51), and
a number of points mapped as Typha were actually
Figure 5. Nelumbo lutea beds near Gard Island, showing
appearance on selected image layers and field determina-
tion of reference point classes. A. Nelumbo beds are clearly
visible on September NIR image. B. bright photosynthe-
sizing vegetation contrasts with dark water on September
NDVI image, and tree shadows are minimized. C.
Nelumbo beds are undetectable on April image, prior
to emergence.
Phragmites (8 out of 21), illustrating that the
separation between these two species was imperfect.
Most of the errors of Phragmites omission were due
to confusion between managed Phragmites and non
persistent emergents (see preceding). Eight points
found to be Phragmites were mapped as wet
meadow; most of these points were located at the
end of the Woodtick Peninsula, and may be due to
differences in substrate that were misinterpreted as
vegetation differences. Points on the Woodtick
Peninsula had sandy soils, whereas points elsewhere
in the marsh generally did not.
Typha was inconsistently mapped. Nearly half of
the points mapped as Typha were actually Phrag-
mites, and only 41%of the points found in the field
to be Typha were mapped as such. In addition to
confusion with Phragmites,errorsofomission
included all classes except for Nelumbo and water.
Wet meadow was much less common than
anticipated, occurring in the field at only nine
points. Of these, five were correctly mapped. Of
the 18 field check points mapped as wet meadow,
eight were actually Phragmites, as discussed above.
The resulting accuracies were low (28%user’s
accuracy, 56%producer’s accuracy).
Points mapped as woody vegetation were usually
correct (85%user’s accuracy). Two points errone-
ously mapped as woody vegetation were actually
Typha, which had a very similar spectral signature
(Figure 4). The other point misclassified as woody
was a Phragmites stand that fell within a tree
shadow on the September image. Producer’s accu-
racy was 77%, with errors partly due to confusion
with Typha and misclassification of a dense stand of
sandbar willow (Salix interior Rowlee) as Phrag-
Using multiseason imagery to map Erie Marsh
aided detection of some plant species but confused
interpretation of others. Multiseason imagery was
essential for identifying Nelumbo beds, which emerge
anew from the waters of coastal wetlands every year.
Nelumbo lutea is listed as ‘‘threatened’’ by the state
of Michigan, but in other states (e.g., Connecticut) it
is banned as a potentially invasive species (USDA
2008). In either case, the ability to detect Nelumbo
beds with satellite imagery is beneficial.
The use of multiseason imagery was expected to
aid identification of other non persistent emergents,
but problems arose due to confusion with areas
where anthropogenic action had artificially removed
overlying plant material. For example, vegetation
changes that occurred within a burned area of the
wetland (‘‘burned,’’ Figure 2) were misclassified as
non persistent emergents because of the change from
extensive vegetation in September 2002 (before the
burn) to shallow open water in April 2003 (after the
burn). It was clear from visual inspection of the
imagery prior to classification that these areas were
not true non persistent emergents, but they were
spectrally inseparable because of the comparable
temporal change from photosynthesizing vegetation
to water. The multiseason analysis was responding
to two different causes of change, plant phenology
and anthropogenic alteration, that had the same
ultimate result but affected very different plant
species. We easily could have improved our classi-
fication accuracy by masking out these human-
altered areas prior to digital image analysis, but
human alteration of wetland vegetation is a common
occurrence in coastal wetlands, so our work
Table 3. The error matrix and producer’s and user’s accuracies resulting from an unsupervised classification with eight
final classes and based on QuickBird high-resolution imagery of the Erie Marsh, North Maumee Bay, Michigan. Numbers
in parentheses and bold represent corrected values based on visual interpretation of the images.
Reference Classification
1 2 3 4 5 6 7 8 Raw Total User’s Accuracy
1 Water 19 1 1 21 90%
2 Soil 1 (0) 16 (17) 1 1 1 1 21 76%(81%)
3–Nelumbo 3(0) 10 (13) 2 5 20 50%(65%)
4 Other
31411 3 1 1 24 17%
5–Phragmites 139 7225176%
6–Typha 1(0) 8 11 (12) 12152%(57%)
7 Meadow 1 8 3 511828%
8 Woody 1 2 17 20 85%
Column Total 27 (22) 18 (19) 11 (14) 8 74 27 (28) 9 22 196
56%77%Overall Accuracy
1036 WETLANDS, Volume 28, No. 4, 2008
provides a more realistic analysis of potential
application problems.
The ability to reliably distinguish Phragmites from
Typha is desirable because rapid invasion by the
invasive Eurasian Phragmites genotype is an in-
creasing problem in Great Lakes coastal wetlands
(Johnston et al. 2007b, Pengra et al. 2007, Tulbure et
al. 2007). Although we achieved moderate success at
identifying Phragmites, the level of confusion
between Typha and Phragmites was greater than
we hoped for. Our overall accuracy would have
improved from 64%to 72%by merging the
Phragmites and Typha classes, but that would have
defeated our objective to differentiate major wetland
species. Other workers have also found that
Phragmites and Typha are spectrally similar, but
that airborne hyperspectral imagery has promise for
differentiating them (Lopez et al. 2004). Future
studies should also investigate the use of LiDAR to
distinguish Phragmites and Typha stands based on
height differences.
Phragmites and Typha have overlapping water
depth preferences (Table 2), so some of the reported
errors may have been due to actual vegetation
changes that occurred between the time of image
acquisition and reference data collection. Phragmites
could have invaded Typha stands or vegetation
management could have promoted Typha over this
time period. Typha is a preferred food of muskrats,
so herbivory could have influenced its abundance
(Kroll and Meeks 1985); muskrat lodges were clearly
visible to the human eye on the April image. We
know that the lag time between image and reference
data acquisition caused fictional errors in our
mapping of Nelumbo beds (Figure 5), but we have
no way to assess whether this happened between
Phragmites and Typha. Optimally, reference data
should be collected at the same time as image
In the field, vegetation interspersion and alter-
ation sometimes made it difficult to assign a
reference point to a particular class. Our require-
ment that reference points be restricted to 3 33
neighboring pixels belonging to the same class
slightly reduced interspersion, but a number of
reference points were within areas where patch sizes
were small and interspersed, or at the edge between
two different classes. Because minor georeferencing
errors in the imagery and/or the field GPS unit
might have caused a spatial mismatch between the
image point and its field reference point, choosing
reference points within a larger contiguous clump of
same-class pixels might have reduced such edge
effects. Assignment of field vegetation to a class was
done without knowledge of the mapped class that
might bias the assignment, but the field assignment
was sometimes ambiguous, and later inspection of
the field data for erroneously mapped points often
revealed that the reference point had some attributes
of the mapped class. For example, a point mapped
as Nelumbo that contained 20%Nelumbo and 20%
Sagittaria cover was field-assigned to the non
persistent emergent class and considered incorrectly
mapped. Utilization of a fuzzy matrix approach that
incorporates variability into the reference data might
have reduced the effect of such minor errors
(Congalton and Green 1999).
Collection of reference data at random points was
extremely difficult. We walked to most of the
reference points because much of the study area
was inaccessible by canoe, but walking through
Phragmites stands was difficult due to slippery clay
soils and dense, tall vegetation (Figure 3f). Access
on foot was sometimes impeded by deep ditches.
Thus, some sections of the area of study had fewer
data points than intended (e.g., Woodtick Peninsu-
la). The final reference data set contained
196 points, an average of one point per every 4.7 ha.
Our number of reference points was less than the
recommended ‘‘rule-of-thumb’’ of 50 points per
mapped class (Congalton 1991). However, this
rule-of-thumb is rarely adhered to in remote sensing
studies of freshwater emergent wetlands due to the
logistical difficulties of obtaining field data. For
example, Phillips et al. (2005) used 10, 14, and 59
reference points for their mapped classes of deep
marsh, wet meadow, and prairie grassland. Sawaya
et al. (2003) used only 5 to 27 field reference points
per mapped class (83 reference points total) when
they mapped wetlands of the .3600 ha Swan Lake,
an average of one reference point per every 43 ha.
Furthermore, their points were not randomly
generated as ours were, but were selected by taking
the classified image into the field so that they could
‘‘identify unique areas with different spectral-radio-
metric responses on the image and target them for
field identification’’ (Sawaya et al. 2003:152). Our
approach of randomly selecting points in advance of
the field work and of gathering field data without
knowledge of the mapped class imposed a much
more statistically rigorous test that should be
considered when comparing our overall accuracy
with that of other studies.
Digital analysis combined with visual interpreta-
tion of displayed satellite imagery might yield the
best results. The high resolution QuickBird imagery
can be visually interpreted much like digital aerial
photography, a capability which Johnston et al.
(2007b) used to map land use. The size, shape,
location, and context of mowed and burned areas
made them easily discernable on the imagery by a
trained air photo interpreter (co-author Johnston).
Improved results might have been obtained by
inferring the vegetation present in human-altered
areas from these contextual clues, rather than trying
to use spectral characteristics alone. Alternatively,
an object-oriented image classification system such
as eCognition that takes into consideration some of
these size, shape, and contextual clues might
improve the classification (Hurd et al. 2006).
QuickBird imagery has only four spectral bands.
Unlike Landsat TM, ETM+or SPOT imagery,
QuickBird does not have a mid-IR band, but only
NIR (760 to 900 nm). The fact that mid-IR bands
and hyperspectral imagery have been demonstrated
to provide separability between wetland types
(Jensen et al. 1993a, Hirano et al. 2003, Bachmann
et al. 2002, Lopez et al. 2004), constitutes a
disadvantage of using QuickBird for wetland
mapping. However, QuickBird’s fine spatial resolu-
tion offers advantages over course resolution
hyperspectral satellite imagery, such as Hyperion,
in instances where wetlands are configured in strips
narrower than image pixel dimensions (Pengra et al.
2007). Thus, there are trade-offs between Quick-
Bird’s higher spatial resolution and its lower spectral
Despite our somewhat low accuracy, we felt that
the QuickBird imagery itself was very appropriate
for use in wetland vegetation mapping. The fine
pixel resolution allowed us to reliably distinguish
features that were long and narrow, such as rows of
trees growing on dikes and berms. Coastal wetland
vegetation often occurs in narrow bands controlled
by water depth or may invade in linear anthropo-
genic features, such as roadside ditches (Maheu-
Giroux and de Blois 2007), so fine resolution is
crucial for detecting such zonation. Multiseason,
multispectral, high-resolution imagery such as
QuickBird could be successfully used for mapping
certain wetland plant communities such as invasive
Phragmites in unaltered wetlands, but in highly
fragmented and managed areas, multiseason imag-
ery may not reflect natural processes but rather
human alterations of the landscape.
We thank Heidi Walking for field assistance and
Tami Watson for assistance with figures. Peter
Wolter collected the 2002 field reconnaissance data
and provided helpful comments on a manuscript
draft. Special thanks go to Jeff Savino at University
of Toledo Lake Erie Center for help with the
logistics in the field. We also thank Joseph Robison
at Michigan DNR and Douglas Pearsall at East
Michigan TNC for allowing us access on their
properties. This research was supported by a grant
from the National Aeronautics and Space Admin-
istration (NAG5-11262-Sup 5) and through a
cooperative agreement with U.S. Environmental
Protection Agency’s Science to Achieve Results
Estuarine and Great Lakes program through
funding to Great Lakes Environmental Indicators,
U.S. EPA Agreement R828675-00. Although the
research described in this article has been funded in
part by the U. S. EPA, it has not been subjected to
the agency’s required peer and policy review and
therefore does not necessarily reflect the views of the
agency and no official endorsement should be
Artigas, F. J. and J. Yang. 2006. Spectral discrimination of marsh
vegetation types in the New Jersey meadowlands, USA.
Wetlands 26:271–77.
Arzandeh, S. and J. Wang. 2003. Monitoring the change of
Phragmites distribution using satellite data. Canadian Journal
of Remote Sensing 29:24–35.
Bachmann, C. M., T. F. Donato, G. M. Lamela, W. J. Rhea, M.
H. Bettenhausen, R. A. Fusina, K. R. Du Bois, J. H. Porter,
and B. R. Truitt. 2002. Automatic classification of land cover
on Smith Island, VA, using HyMAP imagery. IEEE Transac-
tions on GeoScience and Remote Sensing 40:2313–30.
Baker, C., R. Lawrence, C. Montagne, and D. Patten. 2006.
Mapping wetlands and riparian areas using Landsat ETM+
imagery and decision-tree-based models. Wetlands 26:
Brazner, J. C., N. P. Danz, G. J. Niemi, R. R. Regal, A. S.
Trebitz, R. W. Howe, J. M. Hanowski, L. B. Johnson, J. J. H.
Ciborowski, C. A. Johnston, E. D. Reavie, V. J. Brady, and G.
V. Sgro. 2007. Evaluating geographic, geomorphic and human
influences on Great Lakes wetland indicators: multi-assemblage
variance partitioning. Ecological Indicators 7:610–35.
Burton, T. M., D. G. Uzarski, and J. A. Genet. 2004. Invertebrate
habitat use in relation to fetch and plant zonation in northern
Lake Huron coastal wetlands. Aquatic Ecosystem Health &
Management 7:249–67.
Congalton, R. and K. Green. 1999. Assessing the Accuracy of
Remotely Sensed Data: Principles and Practices. Lewis Press,
Boca Raton, FL, USA.
Congalton, R. G. 1991. A review of assessing the accuracy of
classifications of remotely sensed data. Remote Sensing of
Environment 37:35–46.
Cowardin, L. M., V. Carter, F. C. Golet, and E. T. LaRoe. 1979.
Classification of wetlands and deepwater habitats of the United
States. U.S. Fish and Wildlife Service, Office of Biological
Services, Washington, DC, USA. FWS/OBS–79/31.
Ernst-Dottavio, C. L., R. M. Hoffer, and R. P. Mroczynski. 1981.
Spectral characteristics of wetland habitats. Photogrammetric
Engineering and Remote Sensing 47:223–27.
Hurd, J. D., D. L. Civco, M. S. Gilmore, S. Prisloe, and E. H.
Wilson. 2006. Tidal wetland classification from Landsat
imagery using an integrated pixel-based and object-based
classification approach. American Society for Photogrammetry
and Remote Sensing, 2006 Annual Conference, Reno, Nevada.
Jensen, J. R., D. Cowen, J. D. Althausen, S. Narumalani, and O.
Weatherbee. 1993a. An evaluation of the Coast Watch change
detection protocol in South Carolina. Photogrammetric Engi-
neering and Remote Sensing 59:1039–46.
1038 WETLANDS, Volume 28, No. 4, 2008
Jensen, J. R., S. Narumalani, O. Weatherbee, and H. E. Mackey.
1993b. Measurement of seasonal and yearly cattail and
waterlily changes using multidate SPOT panchromatic data.
Photogrammetric Engineering and Remote Sensing 59:519–25.
Johnston, C. A., T. Brown, T. Hollenhorst, P. Wolter, N. Danz,
and G. Niemi. In press. GIS in support of ecological indicator
development. Manual of Geographic Information Systems.
American Society for Photogrammetry and Remote Sensing,
Bethesda, MD, USA.
Johnston, C. A., B. L. Bedford, M. Bourdaghs, T. Brown, C. B.
Frieswyk, M. Tulbure, L. Vaccaro, and J. B. Zedler. 2007a.
Plant species indicators of physical environment in Great Lakes
coastal wetlands. Journal of Great Lakes Research 33(Special
Issue 3):106–24.
Johnston, C. A., T. Watson, and P. T. Wolter. 2007b. Sixty-three
years of land alteration in Erie Township. Journal of Great
Lakes Research 33(Special Issue 3):253–68.
Johnston, C. A. and P. Meysembourg. 2002. Comparison of the
Wisconsin and National Wetlands Inventories. Wetlands
Kroll, R. W. and R. L. Meeks. 1985. Muskrat population
recovery following habitat re-establishment near southwestern
Lake Erie. Wildlife Society Bulletin 13:483–86.
Leica Geosystems. 2003. ERDAS Field Guide, seventh edition.
Leica Geosystems GIS and Mapping, LLC, Atlanta, GA, USA.
Lillesand, T. M. and R. W. Kiefer. 2000. Remote Sensing and
Image Interpretation. John Wiley and Sons Inc., New York,
Lopez, R. D., C. M. Edmonds, A. C. Neale, T. S. Slonecker, K.
B. Jones, D. T. Heggem, J. G. Lyon, E. Jaworski, D. Garofalo,
and D. Williams. 2004. Accuracy assessments of airborne
hyperspectral data for mapping opportunistic plant species in
freshwater coastal wetlands. p. 253–267. In R. S. Lunetta and J.
G. Lyon (eds.) Remote Sensing and GIS Accuracy Assessment.
CRC Press, New York, NY, USA.
Madden, M. 2004. Remote sensing and GIS methodologies for
vegetation mapping of invasive exotics. Weed Technology
Maheu-Giroux, M. and S. de Blois. 2007. Landscape ecology of
Phragmites australis invasion in networks of linear wetlands.
Landscape Ecology 22:285–301.
NOAA (National Oceanic and Atmospheric Administration).
2006. NOAA and the Estuary Restoration Act restoration plan
Osezmi, S. L. and M. E. Bauer. 2002. Satellite remote sensing of
wetlands. Wetlands Ecology and Management 10:381–402.
Pengra, B. W., C. A. Johnston, and T. R. Loveland. 2007.
Mapping an invasive plant, Phragmites australis, in coastal
wetlands using the EO-1 Hyperion hyperspectral sensor.
Remote Sensing of Environment 108:74–81.
Phillips, R. L., O. Beeri, and E. S. DeKeyser. 2005. Remote
wetland assessment for Missouri Coteau prairie glacial basins.
Wetlands 25:335–49.
Rutchey, K. and L. Vilchek. 1999. Air photointerpretation and
satellite imagery analysis techniques for mapping cattail
coverage in a northern Everglades impoundment. Photogram-
metric Engineering and Remote Sensing 65:185–91.
Sawaya, K. E., L. G. Olmanson, N. J. Heinert, P. L. Brezonik,
and M. E. Bauer. 2003. Extending satellite remote sensing to
local scales: land and water resource monitoring using high-
resolution imagery. Remote Sensing of Environment
Schmidt, K. S. and A. K. Skidmore. 2003. Spectral discrimination
of vegetation types in a coastal wetland. Remote Sensing of
Environment 85:92–108.
Sersland, C. A., C. A. Johnston, and J. Bonde. 1995. Assessing
wetland vegetation with GPS linked color video image mosaics.
p. 53–62. In P. Mausel (ed.). Proc. Biennial Workshop on
Videography and Color Photography in Resource Assessment,
15th, Terra Haute, IN. 1–3 May 1995. American Society for
Photogrammetry and Remote Sensing, Bethesda, MD, USA.
Spanglet, H. J., S. L. Ustin, and E. Rejmankova. 1998. Spectral
reflectance characteristics of California subalpine marsh plant
communities. Wetlands 18:307–19.
TNC (The Nature Conservancy). 2006. Erie Marsh Preserve. http://
Tulbure, M. G., C. A. Johnston, and D. L. Auger. 2007. Rapid
invasion of a Great Lakes coastal wetland by non-native
Phragmites australis and Typha. Journal of Great Lakes
Research 33(Special Issue 3):269–79.
USDA (U.S. Department of Agriculture). 2008. Plants database.
Wilcox, K. L., S. A. Petrie, L. A. Maynard, and S. W. Meyer.
2003. Historical distribution and abundance of Phragmites
australis at Long Point, Lake Erie, Ontario. Journal of Great
Lakes Research 29:664–80.
Wolter, P. T., C. A. Johnston, and G. J. Niemi. 2005. Mapping
submerged aquatic vegetation in the U.S. Great Lakes using
Quickbird satellite data. International Journal of Remote
Sensing 26:5255–74.
Manuscript received 19 February 2008; accepted 28 July 2008.
... Some studies have employed commercial high-resolution multispectral images for mapping peatland species. For example, Ghioca-Robrecht et al. (2008) used multi-date QuickBird imagery for mapping Typha spp. and P. australis, but they faced misclassification errors due to the absence of shortwave-infrared (SWIR) bands. ...
... Given the fact that each peatland species' flowering and growth phases vary and that individual species have different best observation dates, the contribution of multi-date imagery can be clearly confirmed. This finding is in line with previous research using multispectral data (QuickBird), which also noticed that multiseason data gave the best results by incorporating phenological variation (Ghioca-Robrecht et al., 2008). When biomass is high, single-date imagery can show a similar spectral reflectance among the peatland species. ...
Full-text available
Peatlands contribute to a wide range of ecosystem services. They play an important role as carbon sinks in their natural state, but when they are drained, they cause carbon emissions. Rewetting drained peatlands is required to reduce carbon emissions and create new carbon sinks. However, drained peatlands are commonly used as grassland or croplands; therefore, alternative agriculture schemes are required following rewetting. Paludiculture, i.e., agriculture on wet and rewetted peatlands, is an option in these areas after rewetting to produce biomass sustainably. Monitoring of peatland management is challenging, yet needed to ensure a successful rewetting and plantation of, e.g., Phragmites australis and Typha spp., two plants which are commonly used in paludiculture. Remote sensing is an excellent tool for monitoring the vegetation composition of vast rewetted peatland regions. However, because many peatland species have similar spectral characteristics, such monitoring is ideally based on high-spatial, high-temporal hyperspectral images. Data that complies with all these requirements does not exist on a regular basis. Therefore, we assessed the potential for mapping peatland vegetation communities in the Peene and Trebel river basins of the federal state of Mecklenburg-Western Pomerania, Germany, using multi-date hyperspectral (PRISMA) data. We used regression-based unmixing to map fractions of different peatland vegetation classes. Results were analyzed with regard to the contribution of multi-date observations and, in comparison, to multispectral datasets (Landsat-8/Sentinel-2). Our results showed that different classes are best mapped at different observation dates. The multi-date hyperspectral datasets produced less Mean Absolute Error (MAE = 16.4%) than the single-date hyperspectral images (ΔMAE + 1%), with high accuracies for all classes of interest. Compared to the results obtained with multispectral data from similar acquisition dates and annual spectral-temporal metrics (STM), the results from hyperspectral data were always clearly superior (ΔMAE + 4%). Besides the superior performance during comparisons, our results also indicate that information that can be derived from the hyperspectral data with the regression-based unmixing goes clearly beyond that of discrete classification. With more hyperspectral sensors coming up and an expected higher availability of multi-data hyperspectral imagery, these data can be expected to play a bigger role in the future monitoring of peatlands.
... Conversely, systems with more stable conditions allow the establishment of large and persistent vegetation beds (Wilcox and Meeker 1991;Paillisson and Marion 2011). Lake Erie water fluctuations have rarely exceeded 1 m in the last 20 years (USGS gage 04215900), allowing for establishment of large vegetation beds (Ghioca-Robrecht et al. 2008;Miller et al. 2018). No study has observed winter habitat selection in the Great Lakes proper, however, evidence suggests that Grass Carp show considerable movement even during the winter in Lake Erie tributaries where many tagged fish reside . ...
Full-text available
Grass Carp, an herbivorous fish introduced in North America to control aquatic vegetation, has become an increasing ecological threat to aquatic ecosystems they invade. Grass Carp have the potential to cause declines and alterations in aquatic vegetation communities, which in turn can have negative impacts on native species reliant on aquatic vegetation. In the last ten years, frequent captures and observed spawning of wild fish have increased concerns of Grass Carp establishment in novel waterways. A particular concern is the potential establishment of fish in Lake Erie and its tributaries. Understanding Grass Carp habitat selection might be useful in population control by guiding harvest actions. In addition, understanding movement ecology of Grass Carp might help to predict where Grass Carp go and what cues initiate those movements. Identifying predictable movement behaviors to heavily used sites might also be useful in control efforts. In this study, I tracked Grass Carp to evaluate winter habitat selection and to determine whether using tagged (Judas) fish is effective in removing wild fish during winter. In addition, this study aimed to characterize upstream migrations into lake tributaries and determine if diploid and triploid fish exhibit similar behaviors. From 2017-2019, I tracked 86 Grass Carp tagged with acoustic transmitters in Truman Reservoir, Missouri to answer these questions. I employed Bayesian discrete choice models to determine winter habitat selection of tagged fish. Generalized linear mixed models were used to determine environmental conditions associated with upstream migrations. Tagged Grass Carp showed strong selection for shallow littoral habitats, and 75% of locations were in waters ≤ 3 m deep. Attempts to capture fish with trammel nets using the Judas method proved difficult with only 1.9 fish caught/netting attempt versus 1.2 caught/netting attempt when selecting areas with suitable Grass Carp habitat for harvest. Fish were often observed near inundated vegetation, with food material collected in 29 of the 31 guts I sampled from fish harvested during capture attempts, which is indicative of overwinter feeding. Eleven fish (6 diploid and 5 triploid) exhibited upstream migrations over the 2018 and 2019 spawning seasons on the Osage River above Truman Reservoir, with distances between 13.6 – 108.1 river km observed. Three of the ten fish in 2019 (2 diploid and 1 triploid) made two migrations and one diploid made three migrations. Upstream migrations were observed from late April to early July and were strongly associated with river temperatures between 15 – 26 ℃ when river levels were rising at a rate > 0.9 m/ 12 h. Five lake resident fish made upstream movements into tributaries during rising river events, indicative that fish residing in the lake proper respond to cues to move upstream. Winter habitat selection suggests that efforts to reduce Grass Carp populations via harvest may be difficult in large lake systems when fish are dispersed throughout the available littoral habitat. Upstream migrations were made by both diploid and triploid fish during conditions suitable for spawning, possibly explaining a motive for these large movements. These results may need to be considered with efforts to remove Grass Carp in large lake systems and for the utilization of triploid fish in observing Grass Carp movement behavior in natural systems.
... For instance, camera traps take photos automatically whenever a motion sensor is activated-computer vision can be used to classify the species that pass by, supporting a real-time, less labor-intensive species count [64,585,643]. It is also possible to use aerial imagery to estimate the size of large herds [805] or count birds [282]. In underwater ecosystems, ML has been used to identify plankton automatically from underwater cameras [232] and to infer fish populations from the structure of coral reefs [867]. ...
Full-text available
Climate change is one of the greatest challenges facing humanity, and we, as machine learning (ML) experts, may wonder how we can help. Here we describe how ML can be a powerful tool in reducing greenhouse gas emissions and helping society adapt to a changing climate. From smart grids to disaster management, we identify high impact problems where existing gaps can be filled by ML, in collaboration with other fields. Our recommendations encompass exciting research questions as well as promising business opportunities. We call on the ML community to join the global effort against climate change.
... By combining spaceborne multi-date synthetic aperture radar (SAR) images and airborne light detection and ranging (LiDAR) elevation (bare earth elevation and vegetation height) images, they conclude that the combination has great potential to monitor wetland change, sea level rise, and invasive species. In another study [31], Ghioca-Robrecht and colleagues used QuickBird multispectral satellite (DigitalGlobe, Westminster, CO, USA) images taken in September 2002 and April 2003 to map emergent wetland vegetation communities within a diked wetland at the western end of Lake Eric. They found that the Multiseason QuickBird imagery is promising for distinguishing certain wetland plant species (P. ...
Full-text available
Biological invasion has been one of the reasons that coastal wetlands gradually lose their ecological services. The current study investigates the spread of a commonly found invasive species in coastal wetlands in Northeastern US, the Phragmites australis. Within a relatively pristine wetland complex in coastal New Jersey, we collected high-resolution multispectral remote sensing images for eight years (2011–2018), in both winter and summer seasons. The land cover/land use status in this wetland complex is relatively simple, contains only five identifiable vegetation covers and water. Applying high accuracy machine learning algorithms, we are able to classify the land use/land cover in the complex and use the classified images as the basis for the grey system coupled system dynamics simulative model. The simulative model produces land use land cover change in the wetland complex for the next 25 years. Results suggest that Phragmites australis will increase in coverage in the future, despite the stable intensity of anthropogenic activities. The wetland complex could lose its essential ecological services to serve as an exchange spot for nekton species from the sea.
... This may explain the low accuracy in modeling reedbeds, noted in other studies (e.g. Ghioca-Robrecht et al., 2008), whose surface areas change according to fluctuations in water levels and human intervention, notably mowing. Fluctuations in water levels can also explain the difficulty in modeling interface habitats (Belluco et al., 2006;van Beijma et al., 2014). ...
Full-text available
In the context of climate change and sea-level rise, coastal realignment consists in reopening polders to marine waters to favor ‘nature-based’ mitigation measures. Such operations have consequences on biodiversity, which vary depending on the parameters studied and site features. In this study, a multimetric indicator aiming to evaluate and predict the potential ecological quality of sites undergoing a realignment operation was developed. This indicator is based on the combination of two tools, (i) a biological-capacity matrix to assess the importance of different habitats of a defined typology for taxonomical, patrimonial and functional parameters; (ii) habitat maps obtained by photointerpretation for past habitats, by machine learning using space-borne imagery for present habitats and by forecasting using submersion models for future habitats. The indicator is presented in the form of a radar chart, with each axis corresponding to one parameter of the biological-capacity matrix and highlighting its different values for different coastal-realignment scenarios or different time horizons.
... Most current mapping methods for invasive P. australis focus on imagery acquired from one of three platforms: satellites, crewed aircraft, and uncrewed aerial vehicles (UAVs). Satellite-based mapping of invasive P. australis has been done with a variety of multispectral imagery (e.g., Arzandeh and Wang 2003;Labda et al. 2007Labda et al. , 2008Labda et al. , 2010Ghioca-Robrecht et al. 2008;Gilmore et al. 2008;Lantz and Wang 2013;Brooks et al. 2015;Xie et al. 2015;Liu et al. 2016;Marcaccio and Chow-Fraser 2016;Rupasinghe and Chow-Fraser 2019), hyperspectral data (e.g., Pengra et al. 2007), fusion methods that incorporate optical and radar imagery (e.g., Bourgeau-Chavez et al. 2004Bourgeau-Chavez et al. 2008a, b;Brooks et al. 2015), and methods that use radar alone (e.g., Bourgeau-Chavez et al. 2009Marcaccio and Chow-Fraser 2016). ...
Full-text available
Wetland managers in North America spend a great deal of time and money trying to control invasive Phragmites australis. Accurate mapping with remote sensing imagery is key to these efforts, which are increasingly employing uncrewed aerial vehicle (UAV) imagery. We mapped P. australis on the Crow Island State Game Area using UAV-derived single-date and multi-date RGB imagery combined with a Digital Surface Model (DSM). In addition to a traditional maximum likelihood classification (MLC), we used two machine-learning (ML) classification algorithms: support vector machine (SVM) and neural network (NN). We assessed accuracy based on both the traditional global model (overall accuracy [OA], omission [OE] and commission [CE] errors for the Phragmites class, and Kappa statistic) and local, per-patch accuracy broken down across 5 density classes and 3 size classes. Our global accuracy assessment for single-date imagery found that SVM (72% OA, 10% OE, 16% CE) performed similar to MLC (70% OA, 17% OE, 8% CE), while NN (33% OA, 7% OE, 41% CE) performed worse. The use of multi-date imagery had little effect on accuracy (MLC 64% OA, 21% OE, 12% CE; SVM 71% OA, 11% OE, 17% CE) except with NN, where the additional bands led to much higher accuracy (67% OA, 7% OE, 22% CE). These results were largely mirrored in the per-patch accuracy assessment, where SVM performed slightly better than MLC and NN performed poorly due to high commission errors. Regarding patch size and density, both larger and medium sized patches, as well as denser patches, were identified relatively accurately, but smaller patches tended to be overestimated and lower-density patches exhibited high omission errors. These results show that wetland managers can achieve very acceptable mapping accuracies with simple methods that require little in the way of resources and expertise.
... Conversely, lake systems with more stable conditions allow the establishment of large and persistent vegetation beds (Wilcox and Meeker 1991;Paillisson and Marion 2011). Lake Erie water fluctuations, for example, have rarely exceeded 1 m in the past 20 years (USGS gauge 04215900; USGS 2016a), allowing for establishment of large vegetation beds (Ghioca-Robrecht et al. 2008;Miller et al. 2018). Additionally, low turbidity in Lake Erie (2.87 NTU, on average; Son and Wang 2019) compared to values measured in Truman Reservoir during our study period (5.7-187.1 NTU) likely allow for more vegetation growth in Lake Erie. ...
Full-text available
Grass Carp Ctenopharyngodon idella were introduced in North America to control aquatic vegetation in small, closed systems. However, when they escape into larger systems in which they can reproduce, they have the potential to cause significant declines and alterations in aquatic vegetation communities. These alterations can in turn affect native species that are dependent on aquatic vegetation. Increased captures and observations of spawning have elevated concerns about Grass Carp establishment in new locations, with particular concern for establishment in Lake Erie and its tributaries. Recent efforts using telemetered fish that co‐locate with wild conspecifics, sometimes in aggregations that are susceptible to harvest, have been used successfully to control invasive Common Carp Cyprinus carpio populations. If Grass Carp aggregate in winter similarly to Common Carp, they might be susceptible to similar control or harvest methods. During the winters (December–March) of 2017–2019, we tracked 86 Grass Carp tagged with acoustic transmitters in Truman Reservoir, Missouri, to evaluate winter habitat selection and to determine the effectiveness of using tagged fish in locating and removing wild fish by comparing harvest at locations of tagged fish to harvest at control sites that we believed were suitable Grass Carp habitat. Discrete‐choice models showed that Grass Carp exhibited strong selection for shallow water, as 75% of locations were in littoral habitats with depths of 3 m or less. On average, we harvested more fish at sites where tagged fish were located (3.6 fish/attempt) than at control sites (1.2 fish/attempt). Full guts in individuals that were harvested may indicate that fish were using shallow‐water habitats to feed. Our results suggested that Grass Carp did not usually form large winter aggregations, and although targeting locations with tagged fish slightly increased harvest success compared to efforts without them, efforts to reduce populations via harvest may be difficult in large systems when fish are widely dispersed.
... Unsupervised classifications are ideal for making inferences where little or no validation data are available. A number of studies have utilised unsupervised classifiers for wetland applications [253][254][255][256][257]. Gluck et al. [258] is an example of a complementary principle component analysis (PCA), which reduces the number of image bands utilised in an ISODATA classifier, where the first principle component (PC1) highlights vegetation, PC2 indicates wetness differences, and PC3 distinguishes wetlands from uplands. ...
Full-text available
The Prairie Pothole Region (PPR) of North America is an extremely important habitat for a diverse range of wetland ecosystems that provide a wealth of socio-economic value. This paper describes the ecological characteristics and importance of PPR wetlands and the use of remote sensing for mapping and monitoring applications. While there are comprehensive reviews for wetland remote sensing in recent publications, there is no comprehensive review about the use of remote sensing in the PPR. First, the PPR is described, including the wetland classification systems that have been used, the water regimes that control the surface water and water levels, and the soil and vegetation characteristics of the region. The tools and techniques that have been used in the PPR for analyses of geospatial data for wetland applications are described. Field observations for ground truth data are critical for good validation and accuracy assessment of the many products that are produced. Wetland classification approaches are reviewed, including Decision Trees, Machine Learning, and object versus pixel-based approaches. A comprehensive description of the remote sensing systems and data that have been employed by various studies in the PPR is provided. A wide range of data can be used for various applications, including passive optical data like aerial photographs or satellite-based, Earth-observation data. Both airborne and spaceborne lidar studies are described. A detailed description of Synthetic Aperture RADAR (SAR) data and research are provided. The state of the art is the use of multi-source data to achieve higher accuracies and hybrid approaches. Digital Surface Models are also being incorporated in geospatial analyses to separate forest and shrub and emergent systems based on vegetation height. Remote sensing provides a cost-effective mechanism for mapping and monitoring PPR wetlands, especially with the logistical difficulties and cost of field-based methods. The wetland characteristics of the PPR dictate the need for high resolution in both time and space, which is increasingly possible with the numerous and increasing remote sensing systems available and the trend to open-source data and tools. The fusion of multi-source remote sensing data via state-of-the-art machine learning is recommended for wetland applications in the PPR. The use of such data promotes flexibility for sensor addition, subtraction, or substitution as a function of application needs and potential cost restrictions. This is important in the PPR because of the challenges related to the highly dynamic nature of this unique region.
... SAR has been used extensively for vegetation identification and monitoring [27][28][29][30][31][32] because the backscattered energy is indicative of the geometric and dielectric properties of surface features [33]. Spaceborne sensors with higher spatial resolutions have also been studied for wetland and species mapping [34][35][36][37][38][39][40]. Laba et al. (2008) [39] reported consistently high accuracies for identifying Phragmites using QuickBird imagery and a maximum-likelihood classifier. ...
Full-text available
Invasive plant species are an increasing worldwide threat both ecologically and financially. Knowing the location of these invasive plant infestations is the first step in their control. Surveying for invasive Phragmites australis is particularly challenging due to limited accessibility in wetland environments. Unoccupied aircraft systems (UAS) are a popular choice for invasive species management due to their ability to survey challenging environments and their high spatial and temporal resolution. This study tested the utility of three-band (i.e., red, green, and blue; RGB) UAS imagery for mapping Phragmites in the St. Louis River Estuary in Minnesota, U.S.A. and Saginaw Bay in Michigan, U.S.A. Iterative object-based image analysis techniques were used to identify two classes, Phragmites and Not Phragmites. Additionally, the effectiveness of canopy height models (CHMs) created from two data types, UAS imagery and commercial satellite stereo retrievals, and the RADARSAT-2 horizontal-horizontal (HH) polarization were tested for Phragmites identification. The highest overall classification accuracy of 90% was achieved when pairing the UAS imagery with a UAS-derived CHM. Producer’s accuracy for the Phragmites class ranged from 3 to 76%, and the user’s accuracies were above 90%. The Not Phragmites class had user’s and producer’s accuracies above 88%. Inclusion of the RADARSAT-2 HH polarization caused a slight reduction in classification accuracy. Commercial satellite stereo retrievals increased commission errors due to decreased spatial resolution and vertical accuracy. The lowest classification accuracy was seen when using only the RGB UAS imagery. UAS are promising for Phragmites identification, but the imagery should be used in conjunction with a CHM.
Full-text available
Machine learning is frequently combined with imagery acquired from uncrewed aircraft systems (UASs) to detect invasive plants. Having prior knowledge of which machine learning algorithm will produce the most accurate results is difficult. This study examines the efficacy of a voting-based ensemble classifier to identify invasive Phragmites australis from three-band (red, green, blue; RGB) and five-band (red, green, blue, red edge, near-infrared; multispectral; MS) UAS imagery acquired over multiple Minnesota wetlands. A Random Forest, histogram-based gradient-boosting classification tree, and two artificial neural networks were used within the voting-based ensemble classifier. Classifications from the RGB and multispectral imagery were compared across validation sites both with and without post-processing from an object-based image analysis (OBIA) workflow (post-machine learning OBIA rule set; post-ML OBIA rule set). Results from this study suggest that a voting-based ensemble classifier can accurately identify invasive Phragmites australis from RGB and multispectral imagery. Accuracies greater than 80% were attained by the voting-based ensemble classifier for both the RGB and multispectral imagery. The highest accuracy, 91%, was achieved when using the multispectral imagery, a canopy height model, and a post-ML OBIA rule set. The study emphasizes the need for further research regarding the accurate identification of Phragmites australis at low stem densities.
Full-text available
Submergent aquatic vegetation (SAV) is a powerful indicator of environmental conditions in both marine and fresh water ecosystems. Quickbird imagery was used to map SAV at three sites across the Great Lakes. Unsupervised classifications were performed at each site using summer Quickbird sensor data. At one site, a multi-temporal classification approach was added, combining visible red difference (May–August) with August red and green visible band data. Multi-temporal SAV classification was superior to single-date results at this site. Muck bottom was not seriously confused with SAV, which was unexpected. Multi-temporal classification results showed less confusion between deep water and SAV, although spectral variability due to sub-surface sandbar structure was a source of error in both single- and multi-date classifications. Nevertheless, some of the confounding effects of water column on SAV classification appear to have been mitigated using this multi-temporal approach. Future efforts would be well served by incorporating detailed, continuous, bathymetry data in the classification process. Quickbird sensor data are very useful for classifying SAV under US Great Lakes conditions. However, regional classification efforts using these data may be impractical at this time, as high cost, rigid tasking parameters and unpredictable water conditions limit availability of suitable imagery.
Drawdown management was used near Port Clinton, Ohio, to restore suitable Ondatra zibethicus habitat in one growing season; muskrats moved into the area immediately and reproduced at high rates, numbers of muskrat houses peaking in the 3rd year. Managers of restored marshes should initially emphasise moist soil and shallow flooding to encourage perennial plants. -from Authors
Color-infrared aerial photography taken in 1991 an d 1995, and SPOT satellite imagery taken in 1991, were utilized to create cattail coverage maps for Water Conservation Area 2A (WCA2A), an impounded portion of the remnant Everglades. Cattail stands were delineated and classified using conventional air photointerpretation and digital image processing techniques, respectively. Four interacting confounding factors (i.e., water depth/color, impacts from fire, periphyton species composition, and growth morphology within a single species) are implicated as possible elements that complicated vegetation classification. Photointerpretation techniques showed an increasing trend in cattail encroachment from 421.6 hectares of monotypic cattail in 1991 to 1646.3 hectares in 1995. A 1991 SPOT classified image appears to have overestimated cattail coverage due to the interacting confounding mechanisms. Overall accuracies for 1995 air photointerpreted map and 1991 SPOT classified image were 95.2 and 83.4 percent, respectively.
Data were analyzed using digital image processing techniques to inventory the spatial distribution of cattail and waterlily beds in a freshwater reservoir located on the Savannah River Site in South Carolina. Creation of a multiple date color composite using October 1988, 1989, and 1990 SPOT panchromatic data proved to be a very effective method to visually identify the change in aquatic macrophyte distribution through time. -from Authors
The NOAA sponsored CoastWatch Change Analysis Project (C-CAP) will utilize remote sensing technology to monitor changes in coastal wetland habitats and adjacent uplands on a cycle of 1 to 5 years. Two study areas in South Carolina were selected to test various C-CAP change detection protocols using near-anniversary Landsat Thematic Mapper data obtained in 1982 and 1988. Fort Moultrie (dominated by salt and brackish marsh) and Kittredge (40 river miles inland and dominated by bottomland hardwoods and riverine aquatic beds) study areas were used to evaluate a modified C-CAP classification scheme, image classification procedures, change detection algorithm alternatives, and the impact of tidal stage on coastal change detection.
Walpole Island in southern Ontario has extensive cattail marsh. In recent years, however, an invasive species (Phragmites) is fast spreading in this area and is affecting the ecosystems and wildlife habitat. In this research, a series of optical satellite images, including Landsat thematic mapper (TM), Satellite pour l'Observation de la Terre (SPOT), and Indian Remote Sensing Satellite (IRS), were used to monitor the recent growth of Phragmites and locate the areas subject to the highest rate of loss of cattail. The post-classification comparison method was applied for change detection. The change-detection accuracy is highly dependent on the accuracy of the classification of the images involved in the process. To improve the results, classification was performed on the combination of spectral bands and textural features. Employing a texture analysis technique, a RADARSAT image was used to resolve the effect of cloud cover on the classification result of the Landsat image. Maps showing the change of Phragmites distribution were created for different years. The change maps showed areas experiencing the highest rate of change in the study area. This work shows that remote sensing can provide important information for wetland vegetation changes.
Historical background, fundamental concepts, statistical considerations and a case study emphasize the need for absolute precision in applying remotely sensed data. This book is a complete guide to assessing the accuracy of maps generated from remotely sensed data.
The author's introduction to remote sensing provides coverage of the subject irrespective of disciplines of study or the academic department in which remote sensing is taught. All the ''classical'' elements of aerial photographic interpretation and photogrammetry are described, but equal emphasis is placed on non-photographic sensing systems and the analysis of data from these systems using digital image processing procedures. This text includes coverage of image restoration, enhancement, classification, and data merging, and new sensor systems such as the Large Format Camera, solid-state linear arrays, the Shuttle Imaging radar systems, the Landsat Thematic Mapper, the SPOT satellite system, and the NOAA Advanced Very High Resolution Radiometer. Also covers imaging spectrometry and lidar systems. It contains extensive illustrations.