high-resolution satellite remote sensing of littoral vegetation of lake sevan (armenia) as a basis...
TRANSCRIPT
LAKE SEVAN
High-resolution satellite remote sensing of littoral vegetationof Lake Sevan (Armenia) as a basis for monitoringand assessment
Jorg Heblinski • Klaus Schmieder •
Thomas Heege • Thomas Kwaku Agyemang •
Hovik Sayadyan • Lilit Vardanyan
Published online: 30 November 2010
� Springer Science+Business Media B.V. 2010
Abstract Physics-based remote sensing in littoral
environments for ecological monitoring and assess-
ment is a challenging task that depends on adequate
atmospheric conditions during data acquisition, sen-
sor capabilities and correction of signal disturbances
associated with water surface and water column.
Airborne hyper-spectral scanners offer higher poten-
tial than satellite sensors for wetland monitoring and
assessment. However, application in remote areas is
often limited by national restrictions, time and high
costs compared to satellite data. In this study, we
tested the potential of the commercial, high-resolu-
tion multi-spectral satellite QuickBird for monitoring
littoral zones of Lake Sevan (Armenia). We present a
classification procedure that uses a physics-based
image processing system (MIP) and GIS tools for
calculating spatial metrics. We focused on classifi-
cation of littoral sediment coverage over three
consecutive years (2006–2008) to document changes
in vegetation structure associated with a rise in water
levels. We describe a spectral unmixing algorithm for
basic classification and a supervised algorithm for
mapping vegetation types. Atmospheric aerosol
retrieval, lake-specific parameterisation and valida-
tion of classifications were supported by underwater
spectral measurements in the respective seasons.
Results revealed accurate classification of submersed
aquatic vegetation and sediment structures in the
littoral zone, documenting spatial vegetation dynam-
ics induced by water level fluctuations and inter-
annual variations in phytoplankton blooms. The data
prove the cost-effective applicability of satellite
remote-sensing approaches for high-resolution map-
ping in space and time of lake littoral zones playing a
Guest editor: Martin A. Stapanian / QuickBird satellite imagery
as a tool for restoration and rehabilitation of Lake Sevan,
Armenia
J. Heblinski (&)
Brauhausgasse 1, 82205 Gilching, Germany
e-mail: [email protected]
T. Heege
EOMAP GmbH & Co. KG, Friedrichshafener Str. 1,
82205 Gilching, Germany
e-mail: [email protected]
K. Schmieder � T. K. Agyemang
Institute for Landscape and Plant Ecology,
University of Hohenheim, August-v.-Hartmann-Str. 3,
70599 Stuttgart, Germany
e-mail: [email protected]
T. K. Agyemang
e-mail: [email protected]
H. Sayadyan
Department of Forestry and Agro-Ecology, State Agrarian
University, 74 Teryan, 375079 Yerevan, Armenia
e-mail: [email protected]
L. Vardanyan
Vanevan University, #19, Gulbenkyan 29A,
Yerevan, Armenia
e-mail: [email protected]
123
Hydrobiologia (2011) 661:97–111
DOI 10.1007/s10750-010-0466-6
major role in lake ecosystem functioning. Such
approaches could be used for monitoring wetlands
anywhere in the world.
Keywords Remote sensing � Littoral vegetation �Water level fluctuations � Spectral unmixing �Inversion � QuickBird
Introduction
Submerged macrophytes perform important ecologi-
cal functions: offering refuge to zooplankton, zoo-
benthos and fish populations; self-purification of the
shallow water zone; and reduction of small-scale
turbulence. Various studies underlined the importance
of structural aspects such as species diversity and
spatial complexity (patchiness) of vegetation struc-
tures for fish populations (Wilcox, 1992; Weaver
et al., 1997; Petr, 2000). Submerged macrophytes also
protect juvenile fish against predators whose catching
efficiency decreases with increasing habitat structural
complexity (Diehl, 1993). Since they can affect fish
populations, submerged macrophytes are not only
ecologically, but also economically important. Chick
& McIvor (1994) report that from the fish’s perspec-
tive, the littoral zone can be perceived as a mosaic of
underwater plants. However, water level fluctuations
strongly influence wetland vegetation patterns,
regardless of seasonal fluctuations (Strang & Dienst,
2004), annual fluctuations (Strang & Dienst, 2004;
Wilcox & Xie, 2007; Wilcox & Nichols, 2008) or
extreme events (Schmieder et al., 2002; Ostendorp
et al., 2003; Dienst et al., 2004).
The water levels of Lake Sevan (Armenia) have
changed considerably over the last few decades.
Anthropogenic activities caused a drop in the lake
level by nearly 20 m between 1940 and 1960. This
decrease in water level was accompanied by serious
ecological consequences (Babayan et al., 2006). In
recent years, the water level rose by nearly 2 m
(Karibyan, 2007), which again may have caused
changes in littoral vegetation structures, affecting
their ecological functions. This increased the need to
document recent changes and predict future changes
of vegetation structures due to any further rise in
water levels.
Large scale mapping and characterisation of
submersed vegetation patterns using remote sensing
and GIS techniques offer many opportunities to
develop ecological assessment tools, such as struc-
tural diversity of littoral vegetation and habitat
suitability models (Woithon & Schmieder, 2004;
Schmieder et al., 2010). These can be used for
monitoring changes and supporting management
decision applications.
Remote sensing image classification is a challeng-
ing task because images often contain information
redundancy, noise, or lack of spectral separation (Lu
& Weng, 2007). In land coverage classification,
remote sensing is accepted as a common technique.
Classification of water constituents is more ambi-
tious, but mapping of macrophytes, sea grasses,
sediments, coral reefs and other ground features is
possible if the water column is sufficiently transpar-
ent, and a sufficient amount of light reaches the
ground and is reflected back out of the water (Dekker
et al., 2001). Govender et al. (2007) reviewed
applications of hyper- and multi-spectral remote
sensing in vegetation and water resources studies.
Becker et al. (2005, 2007) derived optimal band
combinations from hyper-spectral scanner data and
optimal spatial resolutions for wetland vegetation
classification. They recommended an optimal spatial
resolution of 2 m and a minimum of 7 m, with
strategically located bands in the VIS–NIR wave-
length region. During the last decade, physics-based,
automatic classification methods were further devel-
oped by Heege et al. (2004) using spectrally and
spatially high-resolution images and physics-based
classification procedures. Pinnel (2007) analysed the
spectral signatures of various submersed species in
different lakes of the Alpine forelands, which formed
the basis for improved classification algorithms and
assessment tools (Schmieder et al., 2010).
The use of high-resolution images will hopefully
fulfil the needs of environmental monitoring tasks in
areas of high structural diversity. However, in remote
areas, availability of airborne hyper-spectral scanner
data is limited, and flight campaigns from foreign
countries and logistic efforts are restricted and costly.
Sawaya et al. (2003) obtained good results (80%
accuracy) for aquatic vegetation surveys from high-
resolution QuickBird and IKONOS satellite imagery.
Thus, despite spectral limitations, we tested multi-
spectral QuickBird satellite images, highly resolved
in space and time, for their suitability to monitor
littoral vegetation of Lake Sevan (Armenia)
98 Hydrobiologia (2011) 661:97–111
123
associated with water level fluctuations and other
environmental changes.
Materials and methods
Study site
Lake Sevan is the largest freshwater lake in the
Transcaucasus region (Fig. 1). The alpine catchment
of Lake Sevan results in characteristic annual water
level fluctuations with low water during winter and
flood peak in July due to snow melt (Fig. 2). Extensive
use of water for irrigation and hydroelectric power in
the twentieth century caused lowering of the lake level
by nearly 20 m (Chilingaryan et al., 2002; Jenderedjian
et al., 2005; Babayan et al., 2006). This not only
reduced the total lake surface area by 12%, the average
depth by 34.2% and the lake volume by 42.2%, but
particularly affected the littoral zone, seriously affect-
ing its ecological functions. The recently adopted
Sevan rehabilitation programme envisages rising lake
levels, which increased by nearly 2 m between 2003
and 2007 (Karibyan, 2007). Over the 3 years investi-
gated (2006–2008), the water level rose by 0.62 m
(Karibyan, 2007; MES-Armenia, 2008, Fig. 2).
In addition, extended inputs of nutrients from
municipal point sources and non-point sources due to
agriculture and soil erosion from over exploited
pastures in the catchment area increased the trophic
state of Lake Sevan over the last few decades, making
the lake highly eutrophic. Blooms of blue–green
bacteria (Cyanobacteria) during the summer season
are regular, and limit the development of submersed
macrophyte vegetation (Babayan et al., 2006).
The site chosen for this study (Gavaraget) is situated
in the Minor Sevan, the northern part of the lake at
the western shore (Fig. 1). It is the rearmost part of a
wide bay with a more or less unstructured shoreline
not directly influenced by the discharging river
Gavaraget. The littoral zone in Gavaraget gently
declines towards the open water region, with sub-
mersed vegetation of different coverage (Fig. 3A–C)
and extended reeds and bush vegetation (Fig. 3D–F).
Fig. 1 Lake Sevan,
overview (Landsat ETM?
20 August 2005) with
region of interest,
Gavaraget bay (outlined bya white box) shown in the
inset map down left(QuickBird of 16 August
2007). Lake Sevan, darkblue area in main image, is
located in Transcaucasus
region of Armenia (insetmap, top right). Hash markson the margins represent
UTM coordinates Zone
38N, WGS 84. (Color figure
online)
Hydrobiologia (2011) 661:97–111 99
123
Fig. 2 Water levels of
Lake Sevan, daily data
plotted until 19 November
2008 (Karibyan, 2007;
MES-Armenia, 2008),
period of ground truthing
(lilac bands) and days of
satellite data acquisition,
red dot with the data in red.
Months of the year as a
single letter are along the
x-axis, sampling was in
September 2006, then in
July for 2007 and 2008. The
plots demonstrate the
characteristic low levels in
winter and high levels in
summer due to snowmelt,
but the overall picture is of
increasing levels each year.
(Color figure online)
Fig. 3 Submersed
vegetation coverage classes:
A sparse, B moderate and
C dense. D Overview and
E, F emersed vegetation of
Gavaraget bay. Photographs
taken during ground truth
measurements on 23 July
2007. (Color figure online)
100 Hydrobiologia (2011) 661:97–111
123
Remote sensing data
The QuickBird satellite is an on demand working
multi-spectral sensor, and collected the highest multi-
spectral resolution imagery commercially available—
during the term of the project. The spatial resolution is
up to 61 cm Ground Sample Distance (GSD) at nadir
in panchromatic and 2.44 m GSD at nadir in multi-
spectral mode. QuickBird image bands lie at central
wavelengths of 480 nm (Blue), 560 nm (Green),
660 nm (Red), 780 nm (NIR) and 672.5 nm (PAN)
(Digital Globe, 2009 (1)).
For this study we used QuickBird images of the
Gavaraget region over three consecutive years. The
images were collected on 18 August 2006, 16 August
2007 and 23 August 2008. The spatial resolution of
the images represented 0.70/2.8 m in ‘ortho-ready
product level’ (Level 2A) with nearest neighbour
pixel re-sampling and off-nadir viewing angles of
5.2� (2008), 24.33� (2007) and 18.81� (2006). The
satellite data provider DigitalGlobe assures a geo-
location accuracy of 23 m CE90%, excluding any
topographic displacement and off-nadir viewing
(Digital Globe, 2009 (2)). Weather conditions during
remote sensing observations were good, with no
clouds and little sun glint on the water surface. All
the images represent littoral vegetation biomass at
peak state in August. The water transparency
(turbidity, waves) was excellent in 2007, but some
areas had re-suspended sediments in 2006. In
contrast, the 2008 water transparency was extremely
reduced by a blue–green bacteria bloom, with Secchi
depths around 1.5 m only measured during a field
campaign in July 2008.
Ground truth data
During the field campaigns in September 2006, July/
August 2007 and July 2008, we used the ’Radiation
Measurement Sensor with Enhanced Spectral Reso-
lution’ (RAMSES, TriOS Optical Sensors GmbH,
Oldenburg) for ground truth measurements of spec-
tral signatures of littoral vegetation and sediment
patches in shallow water areas, and for spectral
measurements in the pelagial area. The RAMSES
sensor represents a combination of radiance and
irradiance sensors for optical measurements in water.
The system applied in this study consisted of three
sensors, one measuring the up-welling radiance (Lu),
one measuring the up-welling irradiance (Eu) and one
measuring the down-welling irradiance (Ed).
For measurements in the water column, we placed
the sensor at a distance of 2 m from the boat. The
sensors were orientated with the best possible
object–sensor–sun geometry beyond shading influ-
ences. We measured spectral reflectance curves of 30
homogeneous patches of macrophyte or sediment in
2008 to verify classification results. We recorded the
spectra from station equivalent measurements
directly above (R(0?)) and below (R(0-)) the water’s
surface and directly above the top of target (R(b)) in
shallow water areas (Table 1). During a vessel trip in
pelagial areas, we measured reflectance directly
below (R(0-)) the water’s surface and in 2 m water
depth (R(2-)) to determine parameters of atmosphere
and water body as an input for the processing
algorithm (Table 2).
In addition, a dark current measurement was
collected to correct the sensor’s internal noise signal.
The radiometric instruments were linked to a GPS
(Navilock, NL-302U, Berlin Germany) and a Tough-
book CF-30 (Panasonic, Osaka, Japan) to record the
measurements and location of the respective site and
corresponding vegetation patches (see Agyemang
et al., this issue). By diving at measurement points,
we sampled submerged plants and checked the
percentage of vegetation coverage of the patch
measured by visual interpretation. We distinguished
between sparse (\30%), moderate (30–70%) and
dense coverage ([70%) (Fig. 3A–C). The results of
measurements and additional observations (station,
GPS-location, date, time, substrate, vegetation type,
sample, type of spectral measurement, sensor depth,
Secchi depth, water depth, clouds, waves, winds)
were recorded directly in the field (Table 1).
To test the plausibility of the classification results,
the modelled ground reflectance from QuickBird was
compared with ground reflectance measured with
RAMSES during field campaigns. The respective
field measurements of water depth (plumb; Table 1)
were used to check the plausibility of the generated
bathymetric map. The use of GPS during field
activities guaranteed the identification of measure-
ment point location in geo-referenced image data.
Hydrobiologia (2011) 661:97–111 101
123
Image processing
The pre-processing of the raw data was performed
using ENVI (ITT VIS, Boulder, CO, USA) image
analysis software. The pre-processing involved radio-
metric calibration and orthorectification. The images
were radiometrically calibrated and orthorectified
using the calibration coefficients and orthorectifica-
tion parameters delivered by the provider. The
images were rectified to the mean elevation of the
water level during a field trip in July 2007, i.e.
1899 m above the Baltic Sea level (Karibyan, 2007,
pers. comm.). All the images were geo-referenced to
the UTM system, Zone 38 N, WGS-84.
To improve image quality for classification, the
images were corrected for atmosphere, air–water
interface and water column effects using the Modular
Inversion and Processing System (MIP) (EOMAP
GmbH & Co. KG, Gilching, Germany) (Kisselev
et al., 1995; Heege & Fischer, 2004; Miksa et al.,
2005). The physics-based system is designed to
evaluate hyper- and multi-spectral remote sensing
data. Its systematic structure and the independent
treatment of physically independent properties allow
efficient development and implementation of itera-
tively coupled algorithms. Algorithms are defined
independently from sensor and recording conditions.
To execute atmospheric correction with MIP, we
separated terrestrial and water-covered areas. Accord-
ing to the focus of interest, we used different databases
and masked the neglected area. The retrieval of water
constituents for optically deep and shallow natural
waters, as well as the retrieval of littoral ground
characteristics in shallow water areas of lakes and sea,
are major applications of MIP (Heege et al., 2003;
Miksa et al., 2005). Such retrieval is performed in
combination with retrieving atmospheric parameters
and, in specific cases, sun glitter correction algorithms
Table 1 Ground truth measurements in shallow water area Gavaraget (abridged)
Date Time
(local)
Location (UTM
38N, WGS-84)
Sample Measurement
type
Water
depth (m)
Secchi
depth (m)
Wind
force
Wave
height (m)
Cloud
coverage
X (m) Y (m)
20 July
2008
12:53 509310 4475460 P. pectinatus R(0-)/R(b) 1.7 Bottom
visible
Gentle
breeze
0.5–1.0 1/8 (only at
horizon)
25 July
2008
11:45 509275 4475469 P. pectinatus,
Chara sp.
R(0-)/R(b) 2.0 Bottom
visible
Light
breeze
0.2–0.5 0/8
20 July
2008
13:23 509226 4475556 Z. palustris,
Chara sp.
R(0-)/R(b) 1.8 Bottom
visible
Gentle
breeze
0.5–1.0 1/8 (only at
horizon)
20 July
2008
13:38 509153 4475488 Mud Area R(0-)/R(0?) 1.0 0.4 Gentle
breeze
0.5–1.0 1/8 (only at
horizon)
22 July
2008
15:47 509672 4474932 M. spicatum R(0-)/R(0?) 1.0 0.8 Gentle
breeze
0.5–1.0 1/8 (only at
horizon)
22 July
2008
15:31 509672 4474932 White Sand R(0-) 0.3 Bottom
visible
Light
breeze
0.2–0.5 1/8 (only at
horizon)
Table 2 Ground truth measurements in pelagial area used for retrieval of SIOP
Station Time
(local)
Location (UTM 38N, WGS-84) Measurement
type
Secchi
depth (m)
Wind force Wave
height (m)
Cloud coverage
X (m) Y (m)
1 11:15 502497 4488737 R(0-)/R(2-) 4.8 Calm 0 0/8
2 11:45 499266 4489201 R(0-)/R(2-) 5.0 Calm 0 0/8
3 12:30 497607 4485543 R(0-)/R(2-) 4.2 Light breeze 0.2–0.5 0/8
4 13:10 503002 4485081 R(0-)/R(2-) 3.5 Light breeze 0.2–0.5 1/8 (only at horizon)
5 14:50 513678 4475157 R(0-)/R(b) Bottom visible,
1.8 m depth
Light breeze 0.2–0.5 1/8 (only at horizon)
6 15:20 510717 4478459 R(0-)/R(2-) 4.5 Light breeze 0.2–0.5 1/8 (only at horizon)
102 Hydrobiologia (2011) 661:97–111
123
(Heege et al., 2003). The MIP software is an efficient
processing system for inversion of remote sensing data
from natural waters (Heege et al., 2003; Heege &
Fischer, 2004). The overall workflow of data process-
ing in MIP follows the structure presented in Fig. 4. An
azimuthally resolved radiative transfer model for a
multilayer atmosphere–ocean system with a flat water
surface is used for aerosol retrieval, sun glitter and
atmospheric correction modules. The radiative transfer
modules and database system in MIP was implemented
by Kisselev et al. (1995) and is based on the Finite
Element Method (FEM) (Bugarelli et al., 1999).
Radiative transfer calculations are stored in extensive
libraries independently from specific sensor and
recording conditions (system initialisation, necessary
only once). The reduction to concrete recording
conditions and adjusting algorithms to the sensor
specifications is supported automatically in MIP. The
libraries and analytical parameterisations of the under-
water light field form the search space for the retrieval
algorithm.
Water constituent concentrations are needed as
input for the water column correction process. The
water constituent concentrations were calculated
twice. The first approximation was carried out
according to Heege & Fischer (2004) over pelagial
areas. The second adjustment, over moderate deep
benthic substrates, was performed in the frame of the
littoral ground coverage and bathymetry retrieval.
The retrieval of littoral ground reflectance was
performed in shallow water areas. As input, we used
the retrieved water constituent concentrations and the
ground reflectance images. The transformation of
ground reflectance to the ground albedo was based on
the equations published by Albert & Mobley (2003).
The unknown input value of depth was calculated
iteratively in combination with the spectral unmixing
of the respective ground reflectance (Heege et al.,
2006). The unmixing produces littoral ground cover-
age of two to three main class components and the
residual error between the modelled ground reflec-
tance and the calculated reflectance. The final depth,
ground reflectance and ground coverage are achieved
at the minimum value of the residual error. The
potential mapping area is limited due to water
transparency and remote sensor sensitivity. Image
areas without reflection components from the lake’s
ground—the so-called optically deep-water areas—
cannot be mapped with this method. Again, we used
RAMSES measurements of the littoral ground reflec-
tance to validate and consolidate the configuration.
The created underwater digital surface model has a
vertical resolution of 0.1 m.
Based on the ground reflectance image, a spectra-
normalised image, we performed a detailed classifi-
cation of littoral ground coverage. Normalised spectra
are obligatory to summarise spectrally similar pixel in
one class. The final step of the thematic processing
classifies the ground reflectance due to the spectral
signature of different sediment types and macrophyte
Fig. 4 Flow chart of the MIP: A to calculate subsurface
reflectance, B for retrieval of littoral ground coverage and
bathymetric map and C for supervised classification for
underwater cases
Hydrobiologia (2011) 661:97–111 103
123
species using a Fuzzy Logic method including
assignment of individual probability functions for
each defined littoral ground coverage class. Spectra for
the littoral ground classification were derived by
extracting the spectral characteristics from different
areas all over the satellite scene. This was performed
by analysing the statistical variation of each class and
by inspecting the spectral overlaps between the
classes. According to the class-specific spectral fea-
tures, configuration settings for the Fuzzy Logic
discrimination of classes were established. The out-
puts of the littoral ground coverage classification are
class probabilities (in percent) for each coverage class
per pixel.
Calculation of spatial metrics
To summarise spatial changes and assess spatial
diversity, landscape metrics were calculated. This
approach attempts to record the spatial distribution of
landscape elements (patches) and their configuration to
generate explicit, spatially orientated parameters
(Blaschke, 2000). The underlying objective is to record
the structure of a landscape or an ecosystem on the
basis of area, shape, edge, diversity and topology
parameters.
Based on histogram analyses of classification
images, we calculated the surface area covered by
each class. The calculated area metrics are based on a
constant mask, which includes the largest extent of
water on the landside (2008) and the largest extent of
optically deepwater on the lakeside (2008). In addition
to these summarising area metrics, diversity metrics
(Shannon diversity index; and Interspersion and
Juxtaposition Index: IJI) were calculated to document
changes in class diversity and spatial diversity in
vegetation structures over the investigated years using
Fragstats Version 3.3 (McGarigal et al., 2002).
Results
Image processing: Sevan-specific inherent optical
properties
Using the underwater reflectance spectra from RAM-
SES measurements in pelagial areas (Fig. 5A), we
inverted water constituent concentrations and opti-
mised the specific absorption and scattering spectra of
the optical active water components (total suspended
matter, coloured dissolved organic material, phyto-
plankton/chlorophyll). These so-called Lake Sevan-
Specific Inherent Optical Properties (SIOP) (Fig. 5B)
were then applied to processing all the satellite scenes
of Lake Sevan and remarkably improved the
subsequent results.
Classifying multi-spectral satellite data
The basic classification of multi-spectral satellite data
of the three investigated years revealed two main
ground coverage classes: submerged vegetation and
sediments (Fig. 6A, example for 2007). Mixed pixels
contain both vegetation and sediments. Littoral
ground is not completely covered by vegetation in
the respective pixel. The colour gradient represents
the respective proportion of each class, with class
proportions in each pixel adding up to 100%.
The bathymetric map as a direct MIP output
represented a digital surface model of the lake
bottom. Using the example for 2007, with best
conditions compared to 2006 and 2008, the map was
created down to 3 m water column between water
surface and top of target (1896 m above Baltic Sea
level; Fig. 6B). In cases of sediment coverage, the
lake bottom elevation was determined by subtraction
of water level elevation and calculated water column.
Scattered water areas in the land part of the images
are caused by seepage water in small depressions or,
vice versa, dense emergent vegetation coverage
totally hiding the water surface.
The supervised classification procedure revealed
three types of littoral sediments, two types of muddy
areas, and up to six types of submersed vegetation
(Fig. 7). The littoral sediment types represented three
different light-pure sediments without vegetation
cover. Type 1 contained white sand, Type 2 con-
tained buff sand and Type 3 contained dark sand. The
two types of muddy were distinguished by differ-
ences in water clarity.
The six vegetation types (Fig. 7) represented the
littoral ground coverage by submersed macrophytes
(emersed macrophytes were classified separately, and
the data are not presented here). Type 1 was
characterised by sparse coverage (\30%) of low-
growing macrophytes (Chara sp. and Zannichellia
palustris) or sparse coverage (\30%) of high-grow-
ing macrophytes (Potamogeton pectinatus) mixed
104 Hydrobiologia (2011) 661:97–111
123
with buff sand ([70%). Type 2 had moderate
coverage (30–70%) of low-growing macrophytes
(Chara sp. and Z. palustris) mixed with buff sand.
Type 3 had dense coverage ([70%) of low-growing
Chara sp. and Z. palustris partly mixed with sparse
coverage (\30%) of high-growing species (mainly
P. pectinatus). Type 4 was characterised by dense
coverage ([70%) of high-growing plants (mainly
P. pectinatus). Type 5 had moderate coverage
(30–70%) of high-growing species (mainly P. pec-
tinatus, Myriophyllum spicatum and Ceratophyllum
demersum) sometimes mixed with low-growing mac-
rophytes (mainly Chara sp. and Z. palustris) or with
dark sand. Finally, Type 6 was characterised by
moderate coverage (30–70%) of low-growing mac-
rophytes (Chara sp. and Z. palustris), mixed with
dark sand, spiked with stumps of dead reed patches
and/or trees, and low water clarity.
Due to the water level rise of 0.62 m over the three
study years (Fig. 2), the class allocation changed (i.e.
some classes supervened or disappeared). However,
class definition remained constant, so that we could
calculate area metrics and monitor changes of
specific classes over the years. With the collected
results of the 3-year time series, we could document
the spatial distribution of the submerged vegetation
Fig. 5 A Calculated
subsurface irradiance
reflectance R- from
RAMSES measurements at
six pelagial stations (Stat.)
in Minor Sevan. B Lake
Sevan Specific Inherent
Optical Properties (SIOP):
Absorption a* of
phytoplankton (chl) and
Gelbstoff (y), scattering b*
of suspended matter (sm).
(Color figure online)
Hydrobiologia (2011) 661:97–111 105
123
and its spatial changes. In general, due to increasing
water levels, the landside increase of the inundated
area in the gently sloped littoral zone of the ROI
Gavaraget was clearly visible (Fig. 7). Further, a
lakeside decrease of the littoral zone due to reduced
water transparency (optically deep water masked out)
from 2007 to 2008 was apparent, which may have
affected the area metrics. Thus, the calculated area
metrics (Fig. 9) were based on a constant mask. The
water-covered area was largest in 2008 compared to
2007 and 2006, and land area was the opposite
(Fig. 8). The littoral area covered by submersed
Fig. 6 Ground coverage
map of Gavaraget region,
16 August 2007, derived
from QuickBird image
processing. A Derivation of
submersed vegetation with
colour gradient from blue(sediment) to green(vegetation) signifying per
cent ratio of these
components. B Bathymetric
map, colour gradient from
red (shallow areas, about
0.1 m water depth) to green(water depth about 3 m).
Optical deep water is
indicated in black and
terrestrial vegetation in
grey. Contour lines were
digitised from topographic
maps (1:50,000 scale) and
represent elevation above
Baltic Sea level. (Color
figure online)
106 Hydrobiologia (2011) 661:97–111
123
vegetation was largest in 2007, whereas the non-
vegetated littoral area was greatest in 2008. Cover
vegetation types 4 and 5 were the highest in 2007
(Fig. 9), whereas cover of type 3 patches were the
highest in 2008. Large patches vegetation type 1
occurred only in 2006. Patches of vegetation type 6
could only be detected in 2007.
The scores for the Shannon diversity index and IJI
index were the highest in 2007 (1.47 and 83.99,
respectively), compared to 2008 (1.24 and 83.31) and
2006 (0.96 and 81.02). These results indicated higher
class diversity and higher spatial diversity in 2007
compared to the other years. However, variation in
spatial diversity of submersed vegetation structures in
the investigated area was not high over the study
years.
The modelled ground reflectance curves from
QuickBird mirrored the ground reflectance curves
measured by RAMSES in the field (Fig. 10). These
results confirmed the high quality and reliability of
Fig. 7 Detailed littoral ground coverage, Gavaraget region, A 18 August 2006, B 16 August 2007, C 23 August 2008. What each
colour represents is indicated in the key on the right. (Color figure online)
Hydrobiologia (2011) 661:97–111 107
123
the classifications (compare Agyemang et al., this
issue).
Discussion
Remote sensing methods for determining macrophyte
cover in littoral zones of lakes are effective tools for
high-resolution mapping of large areas in space and
time, and thus useful for monitoring and ecological
assessment (Becker et al., 2005, 2007; Schmieder
et al., 2010). Although most applications in lake
littoral zones are based on airborne hyper-spectral
scanner data (Woithon & Schmieder 2004; Becker
et al., 2005, 2007; Schmieder et al., 2010), less cost
and effort are required to test commercially available,
operational satellite data with lower spectral resolu-
tion. We submit that the latter technique offers a
practical and cost-effective alternative for monitoring
and ecological assessment in lake littoral zones
(Sawaya et al., 2003), particularly in remote areas.
Based on multi-spectral data from the QuickBird
satellite sensor, a first classification step allowed us to
create high spatial resolution maps of submersed
vegetation cover and uncovered sediments in the
littoral zone of Lake Sevan. Without this step, Wolter
et al. (2005) only separated two submersed vegetation
density classes, and Sawaya et al. (2003) separated
four submersed vegetation classes differing in density
and growth type. In contrast, we could obtain a higher
resolution of six classes differing in density and
growth type, as well as monitor changes in submersed
vegetation structures over the three investigated
years.
However, vegetation patches need to be large and
adequately dense for successful identification. The
minimum dimension of the patches should at least be
twice the length of a pixel diagonal (Haberacker,
1991). For QuickBird images, with maximal
Fig. 8 Spatial changes in Land-Water-Coverage of ROI
Gavaraget during the investigation period 2006–2008. The ratios
of land (yellow) to water (blue) plotted as bar graphs, clearly show
an increase in water over the 3 years. (Color figure online)
Fig. 9 Spatial changes in
submersed vegetation
classes of ROI Gavaraget
during the investigation
period 2006–2008. Bargraph comparing each year
with respect to different
vegetation types as
described along the x-axis.
(Color figure online)
108 Hydrobiologia (2011) 661:97–111
123
available resolution of 2.4 m, vegetation patches
should be at least 6.8 m on a side. The patch density
and its homogeneity affect the reliability of the class
assignment of a patch. Mixed pixels are classified as a
separate class or have less probability of assignment.
The non-vegetated littoral area was the greatest in
2008 due to a blue–green bacteria bloom. This event
seriously affected the potential of submersed vegeta-
tion classification of deeper littoral areas, leading to a
higher proportion of optical deep water. On the other
hand, the bloom also affected the potential of
submersed macrophytes to form dense high growing
patches. Collectively, this suggests that shading by
phytoplankton has more effect on the high growing
P. pectinatus stands, whereas charophyte-dominated
vegetation seems able to grow well under limited
light conditions. In contrast, P. pectinatus has been
shown to be much more tolerant of high trophic
conditions than charophyte species (e.g. Schneider,
2007). We are unable to resolve this issue which is
suggested for future research.
Our results demonstrate the limitations of remote
sensing in mapping submersed vegetation in phyto-
plankton-dominated lakes. We can only speculate
about the reasons why the regular bloom of blue–
green bacteria did not occur during the summer of
2007. Exceptional weather conditions and increased
total lake volume (approximately 2500 million m3
over the last 6 years) might have inhibited the
development of blue–green bacteria. However,
during clear water conditions in 2007, submersed
macrophytes colonised larger littoral areas, forming
denser patches of increased biomass and higher
spatial diversity, and implying a higher capacity to
perform their ecological functions in the lake
ecosystem.
The low spectral resolution of QuickBird limits the
effectiveness of the classification algorithms and,
consequently, the quality of results and their use for
further applications. Due to the high bandwidth of the
sensor, spectral characteristics of objects can only be
detected at a general level (Sawaya et al., 2003).
However, for the proposed applications, the remote
sensing data quality was acceptable (compare
Agyemang et al., this issue). Sawaya et al. (2003)
also obtained reliable results in mapping littoral
vegetation from high-resolution IKONOS and Quick-
Bird satellites. Thus, particularly for remote areas,
such high-resolution satellites provide an alternative
to airborne hyper-spectral scanner data, especially
when new generations of such satellites offer the
minimum spectral capabilities defined by Becker et al.
(2005, 2007). However, so far only airborne hyper-
spectral scanner data in combination with in situ
measurements of spectral signatures allow higher
thematic resolution up to single species patch detec-
tion (Pinnel, 2007; Schmieder et al., 2010).
In addition to the usual limitations in aquatic
environments (cloud cover, sun glitter, sediment
resuspension, etc.), the retrieval of results in our
study was affected by water level changes (0.62 m
rise between 2006 and 2008) and subsequent varia-
tions in macrophyte spectral characteristics over the
years. In particular, we detected a landside extension
of muddy areas with moderate vegetation cover,
indicating a loss of former terrestrial vegetation.
Rapid water level rise in future years will have a
strong influence on the littoral vegetation. Slowly and
continuously increasing water levels will increase the
chances of macrophytes adapting to changing condi-
tions, thus maintaining their ecological functions
(Wilcox, 1992; Weaver et al., 1997; Petr, 2000).
Conclusions
A continuous environmental monitoring plan for
Lake Sevan using remote sensing techniques is
needed to document and visualise changes in a
Fig. 10 Comparison of reflectances R measured with RAMSES
(continuous lines) and calculated with MIP from QuickBird data
(dotted lines) for different ground coverage (P. pectinatus,
Chara sp., Z. palustris, mud area and white sand as colour coded
in the key top right) in Gavaraget region. (Color figure online)
Hydrobiologia (2011) 661:97–111 109
123
contemporary and cost-effective manner. The
described workflow, using the MIP, demonstrates
the feasibility of retrieving high quality results for
supporting decisions made by lake management
authorities. Although we used a pixel-based classifi-
cation approach here, MIP is open for advanced
object- and knowledge-based classification tech-
niques as described elsewhere (Im et al., 2007;
Lu & Weng, 2007).
An operational, lake-wide application of the
developed algorithms is practicable with little addi-
tional expense. Considering a project-focused acqui-
sition schedule with the provider and stable good
weather conditions, lake-wide coverage of Lake
Sevan could be achieved within five QuickBird
overpasses during 1 month. Once the general optical
properties of a water body (SIOP) are known, data
processing for the whole lake using MIP is not more
work-intensive than for a single scene, due to high
automation of the parameterised processing system.
The use of standardised analytical ground truth
methods can boost the retrieval process. Thus,
although this study covered only part of one lake in
Armenia, it has general applications to wetland
monitoring tasks around the world.
Acknowledgements We thank the SEMIS team and our
other Armenian partners for supporting the project activities.
A special thanks goes to the EOMAP company for support in
processing and parameterisation tasks. Finally, we gratefully
thank the VW Foundation for financial support of this study.
References
Agyemang, T. K., J. Heblinski, K. Schmieder, H. Sayadyan &
L. Vardanyan, 2010. Accuracy assessment of supervised
classification of submersed macrophytes: the case of the
Gavaraget region of Lake Sevan, Armenia. Hydrobiolo-
gia. doi:10.1007/s10750-010-0465-7.
Albert, A. & C. D. Mobley, 2003. An analytical model for
subsurface irradiance and remote sensing reflectance in
deep and shallow case-2 waters. Optics Express 11:
2873–2890 [available on internet at http://opticsexpress./
org/abstract.cfm?URI=OPEX-11-22-2873].
Babayan, A., S. Hakobyan, K. Jenderedjian, S. Muradyan &
M. Voscanov, 2006. Lake Sevan—Experience and Lessons
Learned Brief: 347–362 [available on internet at http://
www.ilec.or.jp/eg/lbmi/pdf/21_Lake_Sevan_27February
2006.pdf].
Becker, B. L., D. P. Lusch & J. Qi, 2005. Identifying optimal
spectral bands from in situ measurements of Great Lakes
coastal wetlands using second-derivative analysis. Remote
Sensing of Environment 97: 238–248.
Becker, B. L., D. P. Lusch & J. Qi, 2007. A classification-based
assessment of the optimal spectral and spatial resolutions
for Great Lakes coastal wetlands imagery. Remote Sens-
ing of Environment 108: 111–120.
Blaschke, T., 2000. Landscape metrics: Konzepte eines jungen
Ansatzes der Landschaftsokologie im Naturschutz. Archiv
fur Naturschutz & Landschaftsforschung 9: 267–299.
Bugarelli, B., V. Kisselev & L. Roberti, 1999. Radiative transfer
in the atmosphere ocean system: the finite-element method.
Applied Optics 38: 1530–1542.
Chick, J. H. & C. C. McIvor, 1994. Patterns in the abundance
and composition of fishes among beds of different mac-
rophytes: viewing a littoral zone as a landscape. Canadian
Journal of Fisheries and Aquatic Sciences 51: 2873–2882.
Chilingaryan, A. L., B. P. Mnatsakanyan, K. A. Aghababyan &
H. V. Toqmagyan, 2002. Hydrology of Rivers and Lakes
of Armenia. Yerevan, Armenia.
Dekker, A., V. Brando, J. Anstee, N. Pinnel, T. Kutser, H.
Hoogenboom, R. Pasterkamp, S. Peters, R. Vos, C. Olbert
& T. Malthus, 2001. Applications of imaging spectrom-
etry in inland, estuarine, coastal and ocean waters. In van
der Meer, F. D. & S. M. de Jong (eds), Imaging Spec-
trometry: Basic Principles and Prospective Applications,
Volume IV of Remote Sensing and Digital Image Pro-
cessing. Kluwer Academic Publishers, Dordrecht, The
Netherlands.
Diehl, S., 1993. Effects of habitat structure on resource avail-
ability, diet and growth of benthivorous perch, Percafluviatilis. OIKOS 67: 403–414.
Dienst, M., K. Schmieder & W. Ostendorp, 2004. Effects of
water level variations on the dynamics of the reed belts of
Lake Constance. Limnologica 34: 29–36.
Digital Globe, 2009 (1). QuickBird Product Description [avail-
able on internet at http://www.digitalglobe.com/index.php/
85/QuickBird, accessed 24 February 2009].
Digital Globe, 2009 (2). QuickBird Imagery Products, Prod-
uct Guide, Revision 5.0 [available on internet at http://
www.digitalglobe.com/index.php/6/DigitalGlobe?Products
, accessed 03 March 2009].
Govender, M., K. Chetty & H. Bulcock, 2007. A review of
hyperspectral remote sensing and its application in vege-
tation and water resources studies. Water SA 33: 145–151.
Haberacker, P., 1991. Digitale Bildverarbeitung: Grundlagen
und Anwendungen. Carl Hanser Verlag, Munchen, Wien.
Heege, T., A. Bogner & N. Pinnel, 2003. Mapping of sub-
merged aquatic vegetation with a physically based process
chain. Proceedings of Remote Sensing, SPIE—The
International Society for Optical Engineering, Vol. 5233.
CD-ROM Proceedings.
Heege, T., C. Hase, A. Bogner & N. Pinnel, 2004. Physikalisch
basierte Prozessierung multispektraler Fernerkundungsda-
ten von Binnengewassern. Laufener Seminarbeitrage, Ba-
yer. Akad. f. Naturschutz u. Landschaftspflege 03: 67–71.
Heege, T., P. Hausknecht & H. Kobryn, 2006. Hyperspectral
seafloor mapping and direct bathymetry calculation using
HyMap data from the Ningaloo reef and Rottnest Island
areas in Western Australia, CD-ROM. 13th ARSP Con-
ference, 20–24 November 2006, Canberra, Australia: 1–7.
110 Hydrobiologia (2011) 661:97–111
123
Heege, T. & J. Fischer, 2004. Mapping of water constituents in
Lake Constance using multispectral airborne scanner data
and a physically based processing scheme. Canadian
Journal Remote Sensing 30: 77–86.
Im, J., J. R. Jensen & J. A. Tullis, 2007. Object-based change
detection using correlation image analysis and image
segmentation. International Journal of Remote Sensing
29(2): 399–423.
Jenderedjian, K., A. Babayan, S. Hakobyan, S. Muradyan &
M. Voskanov, 2005. Managing Lakes and their Basins for
Sustainable Use: A Report for Lake Basin Managers and
Shareholders. ILEC Foundation, Kusatsu, Japan.
Karibyan, M., 2007. Head of Meteorological Station, Sevan,
Hydro-Meteorological Agency, Ministry of Nature Pro-
tection, Armenia. Personal Communication 4 August
2007.
Kisselev, V. B., L. Roberti & G. Perona, 1995. Finite-element
algorithm for radiative transfer in vertically inhomoge-
neous media: numerical scheme and applications. Applied
Optics 34: 8460–8471.
Lu, D. & Q. Weng, 2007. A survey of image classification
methods and techniques for improving classification per-
formance. International Journal of Remote Sensing 28(5):
823–870.
McGarigal, K., S. A. Cushman, M. C. Neel & E. Ene, 2002.
FRAGSTATS: Spatial Pattern Analysis Program for Cate-
gorical Maps [available on internet at http://www.umass.
edu/landeco/research/fragstats/fragstats.html, accessed March
2010].
MES-Armenia (Ministy of Emergency Situations), 2008.
Hydro-Meteorological Agency. Communication Novem-
ber 2008.
Miksa, S., T. Heege, V. Kisselev & P. Gege, 2005. Mapping
water constituents in Lake Constance using Chris/Proba.
Proceedings of the 3rd ESA Chris/Proba Workshop,
Volume ESA SP-593, June 2005, Frascati, Italy. ESRIN.
Ostendorp, W., M. Dienst & K. Schmieder, 2003. Disturbance
and rehabilitation of lakeside Phragmites reeds following
an extreme flood in Lake Constance (Germany). Hydro-
biologia 506–509: 687–695.
Petr, T., 2000. Interactions between fish and aquatic macro-
phytes in inland waters. A review. FAO Fisheries Tech-
nical Paper 396. FAO, Rome: 185 pp.
Pinnel, N., 2007. A Method for Mapping Submerged Macro-
phytes in Lakes using Hyperspectral Remote Sensing:
164 pp [available on internet at http://mediatum2.ub.
tum.de/node?id=604557].
Sawaya, K. E., L. G. Olmanson, N. J. Heinert, P. L. Brezonik &
M. E. Bauer, 2003. Extending satellite remote sensing to
local scales: land and water resource monitoring using
high-resolution imagery. Remote Sensing of Environment
88: 144–156.
Schmieder, K., M. Dienst & W. Ostendorp, 2002. Effects of the
extreme flood in 1999 on the spatial dynamics and stand
structure of the reed belts in Lake Constance. Limnologica
32: 131–146.
Schmieder, K., A. Woithon, T. Heege & N. Pinnel, 2010.
Remote sensing techniques and GIS modeling approaches
for monitoring and assessment of littoral vegetation at
Lake Constance, Germany. Verhandlungen International
Verein Limnologie 30(10): 1–3.
Schneider, S., 2007. Macrophyte trophic indicator values from
a European perspective. Limnologica 37(4): 281–289.
Strang, I. & M. Dienst, 2004. Effects of water level at Lake
Constance on the Deschampsietum rhenanae from 1989 to
2003. Limnologica 34: 22–28.
Weaver, M. J., J. J. Magnuson & M. K. Clayton, 1997. Distri-
bution of littoral fishes in structurally complex macro-
phytes. Canadian Journal of Fisheries and Aquatic Sciences
54: 2277–2289.
Wilcox, D. A., 1992. Implications for faunal habitat related to
altered macrophyte structure in regulated lakes in North-
ern Minnesota. Wetlands 12: 192–203.
Wilcox, D. A. & S. J. Nichols, 2008. The effects of water-level
fluctuations on vegetation in a Lake Huron wetland.
Wetlands 28(2): 487–501.
Wilcox, D. A. & Y. Xie, 2007. Predicting wetland plant
community responses to proposed water-level-regulation
plans for Lake Ontario: GIS-based modeling. Journal of
Great Lakes Research 33(4): 751–773.
Woithon, A. & K. Schmieder, 2004. Bruthabitatmodellierung
fur den Drosselrohrsanger (Acrocephalus arundinaceusL.) als Bestandteil eines integrativen Managementsystems
fur Seeufer. Limnologica 34: 132–139.
Wolter, P. T., C. A. Johnston & G. J. Niemi, 2005. Mapping
submerged aquatic vegetation in the US Great Lakes
using QuickBird satellite data. International Journal of
Remote Sensing 26(23): 5255–5274.
Hydrobiologia (2011) 661:97–111 111
123