Remote Sensing of Environment 97 (2005) 415 – 433 www.elsevier.com/locate/rse
Retrospective seagrass change detection in a shallow coastal tidal Australian lake Arnold G. Dekker, Vittorio Ernesto Brando *, Janet M. Anstee CSIRO Land & Water, Environmental Remote Sensing Group, GPO BOX 1666, Canberra, ACT 2601, Australia Received 22 January 2004; received in revised form 9 February 2005; accepted 12 February 2005
Abstract Satellite imagery was used to detect the change in seagrass and macroalgal communities of a shallow coastal lake over a period of 14 years. The lake benthic material was classified into sets of spectral classes representing the patterns and texture of the ecosystem, and then linked to environmentally relevant labels through a radiative transfer model. The classification results for 2002 achieved an accuracy of 76% for the least understood areas; other areas were significantly better, but not quantified. Classification results of 1988, 1991, and 1995 were consistent with past surveys and maps. Based on the change detection from 1988 to 2002 Posidonia, Ruppia and Halophila change slightly in the 14 year period from 1988 to 2002. However, Zostera has undergone significant change and adaptation. Early in the time series (between 1988 and 1991) a reduction in Zostera beds was evident, especially in the middle and south of the lake with some areas not returning by 2002. Epiphytic growth over Zostera could be a confounding factor here, but the Landsat sensors do not have sufficient spectral resolution to detect these subtleties. Hyperspectral remote sensing could resolve this issue more clearly. D 2005 Elsevier Inc. All rights reserved. Keywords: Satellite imagery; Change detection; Submerged vegetation; Landsat
1. Introduction The need to map and monitor the meadows of seagrass and associated macroalgae, microphytobenthos and physical substrata, over a range of spatial and temporal scales, is of prime importance in assessing the status of coastal systems. The first step is to provide baseline maps that document the current extent, diversity and condition of the seagrasses. The next step is to establish monitoring programs designed to detect disturbance at an early stage, and to distinguish such disturbance from natural variation in the meadows (Borum et al., 2004; Kendrick et al., 2000; Kirkman, 1996; Lee Long et al., 1996). Historically, seagrass distribution mapping and assessment has been undertaken through the use of aerial photography interpretation and direct field mapping—a
* Corresponding author. Tel.: +61 2 62465716; fax: +61 2 62465815. E-mail address:
[email protected] (V.E. Brando). 0034-4257/$ - see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2005.02.017
labour intensive and somewhat subjective assessment methodology. The use of advanced satellite and/or airborne remote sensing technology provides an opportunity to undertake more cost effective and objective monitoring (Dekker et al., in press; Duarte et al., 2004). Remote sensing techniques provide the tool of choice over more traditional mapping methods when the assessment of the entire meadow is the focus of the monitoring programme (Duarte et al., 2004; Krause-Jensen et al., 2004; Short & Coles, 2001), as substrate cover variability assessment requires a spatially comprehensive mapping system. Field methods, from diver survey to underwater videography (Norris & Wyllie-Echeverria, 1997; Norris et al., 1997) and acoustic (e.g. Lee Long et al., 1998) transect methods, produce mapping errors due to the need for spatial interpolation between the in situ data points or transects. The increased adoption of remote sensing techniques is due to the combined development of sensor sophistication, in situ optical instrumentation, underwater light climate modelling tools and inversion methods
416
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
(which invert a map of benthic vegetation from the remotely sensed signal). Landsat MSS broadband satellite imagery became available in 1973 but prior to that date seagrasses could only be synoptically mapped by using aerial photographs. Seagrass mapping was attempted using Landsat MSS data (e.g. Ackleson & Klemas, 1987; Claasen et al., 1984) but the maps produced were of limited value since a 79 m ground resolution is too large for the typical size and patchy nature of most seagrass meadows, and may not even be relevant to the scale of many estuaries (Cracknell, 1999). The accuracy that can typically be expected when mapping broad benthic habitat classes using Landsat MSS is in the range of only 30 – 60% according to Mumby et al. (1997). Landsat-5 Thematic Mapper was the first satellite system that supplied broad spectral band environmental data with a spatial resolution of 30 m from 1984 onwards: soon augmented by SPOT imagery with 20 m pixels from 1986 onwards. Satellite remote sensing is the most cost effective method for mapping and monitoring seagrasses in large and remote regions (Ferguson & Korfmacher, 1997; Mumby et al., 1999; Ward et al., 1997). In such situations it can provide coarse scale seagrass maps with an accuracy of 75– 85% (e.g. Mumby & Edwards, 2002; Mumby et al., 1997), particularly when large and/or continuous meadow forming
species such as Zostera marina dominate (Ward et al., 1997) including estimation of the biomass of submerged vegetation (e.g. Zhang, 1998). The spectral resolution of Landsat TM is somewhat better for the discrimination of benthic vegetation than that of the SPOT multispectral sensor (XS), while the radiometric resolution (i.e. number of brightness levels that can be resolved) of both the TM and XS sensors are well suited for this task (Lubin et al., 2001). The Landsat TM and SPOT XS sensors have been successfully applied, within these limitations, to benthic vegetation mapping in a variety of situations (e.g. Ackleson & Klemas, 1987; Ben Moussa et al., 1989; Chauvaud et al., 2001; Ferguson & Korfmacher, 1997; Liceaga-Correa & Euan-Avila, 2002; Macleod & Congalton, 1998; Robblee et al., 1991; Zainal et al., 1993). Multi-date satellite remote sensing is geometrically highly repeatable and a cost effective method for detecting large changes in seagrass distribution or extent over time (Macleod & Congalton, 1998; Robblee et al., 1991; Ward et al., 1997; Zainal et al., 1993). Wallis Lake is a shallow estuarine lake located about 360 km north of Sydney on the central coast of New South Wales, Australia, with a water surface areas of 94 km2 (Fig. 1). The Wallis Lake system consists of lakes and rivers with interconnecting channels. The lake is a significant environmental resource and is also used for recreational activities
Fig. 1. Wallis Lake with the field site positions of the August 2001 fieldwork. The shaded area represents a focus area for future change detection and the dashed box is where this study research was undertaken.
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
and aquaculture. The tidal range of Wallis Lake is reduced by the frictional effort of the entrance channel and the reported range within the lake is about 5% of the ocean tide (NSW Public Works, 1990). Seagrass (and macroalgae) is an important component of the Wallis Lake ecosystem. The seagrass beds help to stabilise sediments, cycle nutrients, provide valuable ‘‘nursery habitat’’ and are a significant source of detrital material for estuarine food webs. In Wallis Lake four seagrass species have been identified (Laegdsgaard, 2001; West et al., 1985). Large stands of Posidonia australis, a seagrass species of known ecologically sensitivity (Shepherd et al., 1989), together with Ruppia megacarpa, Halophila ovalis and Zostera capricorni. Unvegetated substrates vary from mud via sand to coarse sand with shell fragments. Vegetated substrates vary from very lightly vegetated with Halophila via densely vegetated areas of mixtures of Zostera, Posidonia and various macroalgae to monospecies fields of Zostera or Posidonia at various levels of density. The ecological importance of these seagrasses, together with their sensitivity to water quality parameters have led to their use as biological indicators of estuary health for the Wallis Lake Catchment Management Plan (WLCMP, 2001). The number of submerged vegetation taxa occurring in the lake creates a complex vegetated substrate environment. Thus, Wallis Lake was a challenging target for developing a method for change detection of seagrasses and macroalgae by satellite remote sensing, but due to its complexity, a representative target for other coastal lakes. This study demonstrates that more species discrimination is possible with Landsat 5 TM and 7 Enhanced Thematic Mapper. The potential use of archived Landsat data for detection of Wallis Lake benthic cover was investigated, as it is the only source of archival data going back to 1984 at a sufficient spatial resolution. High spectral resolution techniques such as atmospheric correction, field spectroradiometry, bio-optical modeling and classification were applied to the lower spectral resolution Landsat data. The methodology is first developed on a recent image (2002) with associated fieldwork and measurements. Subsequently the images from 1995, 1991 and 1988 are processed and classified and compared to the 2002 image. This leads to a change detection assessment of seagrasses and macroalgae in Wallis Lake across the four satellite images from 1988 to 2002.
2. Methodology 2.1. Water and substrate spectral characterization The remote sensing of seagrasses and related environments is based on the principle that a remote sensor can Fsee_ the substrate and the vegetation growing on or in that
417
substrate. The desirable layer of information about the seagrasses is covered by a water column that attenuates the light reaching, and being reflected from, the benthos. While the remote sensing of terrestrial plants makes significant use of the red edge (i.e. the steep slope between strong red wavelength absorption and strong near-infrared reflectance characteristic of the spectral signatures of healthy plant leaves), aquatic plants cannot be recognized by this feature since wavelengths beyond 680 nm are significantly attenuated by pure water (Kirk, 1994). In coastal waters, spectral scattering and absorption by phytoplankton, suspended organic and inorganic matter and dissolved organic substances may further restrict the light passing to and being reflected up from the benthos (Dekker et al., 2001). The spectral discrimination between aquatic plant species must therefore concentrate on their pigment related spectral features within the visible wavelengths, where light penetrates the water column and can be reflected back to the sensor (Fyfe, 2003). Thus a fieldwork was aimed at the spectral characterization of the Wallis Lake aquatic system, focusing on estimating the optical properties of the water column and on the optical properties of the substrate vegetation. First we discuss the spectral characterization of the seagrass and macroalgae species and then the spectral characterization of the Wallis Lake waters. A field campaign was conducted on 22 –24 August 2001 to measure and sample the lake’s optical properties at seven locations within Wallis Lake and its tributaries (see Table 1 for a summary of the measurements and Fig. 1 for locations). In situ measurements included (i) profiles of downwelling irradiance and upwelling radiance within the water column, using the RAMSES spectroradiometers and (ii) profiling measurements of spectral backscattering using the HydroScat-6. In situ spectral reflectance measurements of the benthic material were also collected using the RAMSES spectroradiometer system. In order to understand the absorption properties of the water, in situ samples for spectrophotometric measurement of the phytoplankton absorption, the tripton absorption and the CDOM (coloured dissolved organic matter) absorption, were kept cool and dark until analysis in the laboratory. Survey points (using a global positioning system) of target materials and sampling points were recorded. Additional GPS survey points were collected for the purpose of image geometric accuracy testing and resampling. The RAMSES submersible system deployed in Wallis Lake consists of two cosine collector sensors measuring the downwelling irradiance (E d) and one radiance collector measuring the upwelling radiance (L u), and a depth sensor. One downwelling irradiance sensor and the upwelling radiance sensor (mounted in a cage) were lowered on the ‘‘sunny’’ side of the boat to minimize the shading effects, while the second downwelling irradiance sensor was mounted on the mast of the boat to monitor the downwelling irradiance in the air (E d air).
418
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
Table 1 Summary of the Wallis Lake system fieldwork
WL22-1 Wl22-3 WL23-5 WL23-8 WL23-10 WL24-1 WL24-2
Site name
Bottom Secchi Cloud Chl a TSS TR a CDOM depth (m) depth (m) cover (%) (Ag L 1) (mg L 1) (mg L 1) Slope S CDOM
a CDOM (440)
* a TR Slope S TR
a *TR (550)
b bp
Wallamba River Pipers Ck Channel Pacific Palm Yahoo Island Coolongolook R Inlet Steps
2.2 1.9 1.4 1.3 2.3 3.0 1.7
0.808 0.485 0.619 0.365 0.670 0.239 0.344
0.0124 0.0117 0.0110 0.0127 0.0124 0.0116 0.0129
0.0060 0.0027 0.0007 0.0021 0.0051 0.0035 0.0026
0.065 0.69 0.044 0.77
1.4 1.6 * 1.2 1.2 2.0 *
10 25 10 50 50 0 0
0.64 0.83 0.34 0.99 0.78 0.80 0.72
15.47 16.69 13.00 20.53 17.22 12.60 11.55
15.42 16.63 12.98 20.46 17.17 12.54 11.50
0.0179 0.0181 0.0189 0.0185 0.0186 0.0181 0.0186
Gamma
0.106 1.06 0.098 0.89 0.032 0.72
* Bottom visibility.
2.2. Spectral characterization of the seagrasses and macroalgae Spectral reflectance measurements were collected of several macroalgae and seagrasses, taxonomically identified by Dr. Pia Laegdsgaard (from the Coastal Ecology Group, Centre for Natural Resources, Department of Land and Water Conservation), see Table 2. The irradiance reflectance R was estimated using the two spectroradiometers mounted in the cage measuring upwelling radiance (L u) from the target (that is, the seagrasses and macroalgae samples) and measuring downwelling irradiance. The seagrass and macroalgae samples (several layers of the plant material) were positioned on a black neoprene mat in order to avoid background signal contamination. The spectroradiometer set up was moved around the sample at least 5 times to get an average signal of reflectance of the sample. Fig. 2 contains a selection of averaged benthic vegetation irradiance reflectance (E u / E d) spectra of eight Wallis Lake species measured on 22 –24 August 2001. The vegetation spectra are spectrally distinct, mainly due to a varying pigment composition as exemplified by the local spectral troughs in reflectance. This adds to and confirms research by Fyfe (2003) who carried out a systematic examination of
the spectral variability of three seagrass species (Zostera capricorni, Posidonia australis and Halophila ovalis) over space and time, comparing their reflectance across different seasons, estuaries and habitats as well as under conditions of epiphytic algal growth. The seagrass species were found to be spectrally distinct despite intraspecific variability in their signatures and irrespective of the level of algal fouling on leaves, although epiphyte growth did reduce discrimination between species (Fyfe, 2003). This spectral separability of seagrasses and macroalgal species implies that hyperspectral remote sensing of Wallis Lake (either from aircraft or from satellites) will enable effective discrimination. The mapping of each of these individual species relies upon the condition that the water column is sufficiently transparent to obtain the significant discriminatory part of the spectrum of the substrate species. 2.3. Optical spectral water properties characterization Now it has been established that many of the target species are spectrally separable it is necessary to ascertain that sufficient light reaches the substrate vegetation and returns to the surface for detection by an aircraft or satellite.
Table 2 Macroalgae and seagrasses spectrally characterised in Wallis Lake Taxa
Common name
Type
Occurrence
Gracilaria sp.
Gracilaria
Red epiphytic macroalgae
Cystoseria trinodis
Cockleweed
Brown macroalgae
Sargassum sp. Chara sp.
Sargassum Stonewort
Brown macroalgae Green macroalgae
Chaetomorpha sp. Posidonia australis Zostera capricornia
Chaetomorpha Strapweed Eelgrass
Filamentous green algae Seagrass Seagrass
Halophila ovalis
Paddleweed
Seagrass
Ruppia megacarpa
Sea Tassel
Seagrass
Usually found in intertidal to subtidal regions. Commonly associated with Zostera. Usually found in intertidal to subtidal regions on sheltered and semi exposed shores. Ubiquitous. Usually prefers slow moving to stationary water. It can be found in range of water depths from 0 – 4 m. Occurs in sheltered locations. Found in shallow subtidal areas in clear waters. Found in sheltered areas with sand or mud substrate. It can be found in range of water depths from 0 – 7 m. Found between other seagrass in sandy shallow areas, adapted to low light levels. Found in a range of habitats exhibiting various degrees of turbidity and salinity. Often occurs mixed with Zostera.
The occurrence was summarized from Laegdsgaard (2001).
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
419
Fig. 2. Averaged benthic substrate reflectances measured on the 22nd – 24th August 2001 using the RAMSES field spectroradiometer. The wavelength range between 480 and 680 nm allows maximal penetration of light into the water column and on to the substrate. The higher reflectance towards the NIR (700 nm) is less relevant for remote sensing as the pure water absorption becomes significantly higher towards these longer wavelengths.
According to Dekker et al. (2001) and Maritorena et al. (1994), the reflectance just below the surface of a homogeneous water body with a reflecting substrate R(0, H) can be described as: Rð0 ; H Þ ¼ RV þ expð Kd HÞ½ Aexpð jB H Þ RV expð jC H Þ
ð1Þ
where R V represents the subsurface irradiance reflectance R(0) of a hypothetical optical deep water column, A is the reflectance of the substrate, j B and j C are the spectral diffuse vertical attenuation coefficients for the two upwelling light streams: one from the bottom and one from the water column. If one is not able to separate the two upwelling light streams, assuming that j B = j C = j, Eq. (1) simplifies to: Rð0 ; H Þ ¼ RV þ ð A RV Þexp½ ðKd þ jÞH :
ð2Þ
That is identical to formulation of Philpot and Vodacek (1989). Thus an important property of waters to be estimated in the field is the spectral diffuse vertical attenuation coefficient for downwelling irradiance K d(k). At any depth the downwelling irradiance at that depth may be described by: Ed ðk; zÞ ¼ Ed ðk; 0Þexp ð Kd ðkÞzÞ
ð3Þ
where E d(k, z) is the downwelling irradiance at any depth z and E d(k, 0) is the downwelling irradiance at the water surface and K d(k) was estimated with a regression of the log transformed downwelling irradiance versus depth for all the sampling stations. The K d(k) values estimated over the sampling sites are presented in Fig. 3. It is apparent from the
figure that the light penetration through the water column varies in this lake system from the more turbid river stations, showing spectrally higher levels of K d(k), to the coastal inlets where ocean water flows into the lake (the lowest attenuation spectrum). With this K d(k) measurement and knowing the substrate reflectance it is now possible to estimate to what depth Landsat can discriminate substrate species. The substrate will be detectable if the second term of Eqs. (1) or (2) is greater than the detecting threshold in reflectance terms for the Landsat TM sensor (see further in text). Fig. 4 illustrates these results for the measured substrate reflectances of Posidonia and Zostera and the measured K d(k)s at 550 nm, assuming a detecting threshold of 0.5% of R(0) and a R V of 7.5%. In the clearest waters Zostera will be detectable up to a depth of 2.4 m, while in more turbid waters it will be detectable up to a depth of 0.7 m. As a field sampling program is inevitably limited in duration and scope, a powerful tool for obtaining more knowledge about reflectance spectra above a shallow coastal lake system (or indeed any optically shallow water body) is the use of analytical or radiative transfer based aquatic optical model simulation tools. HYDROLIGHT (Mobley & Sundman, 2000a,b), developed by Mobley (1994), is a radiative transfer model that computes radiance distribution and derived quantities for natural water bodies. HYDROLIGHT can be used as a tool to simulate the effects of bathymetry, benthic substrate and depth, turbidity and wind speed on K d, the upwelling radiance and irradiance as well as on reflectance. HYDROLIGHT can accurately simulate water body radiance or reflectance, provided it is given correct input parameterisation files. Input of the model consists of the absorbing and scattering properties of the
420
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
Fig. 3. Spectral vertical attenuation coefficient for the downwelling irradiance K d estimated at the sampling sites.
water body, the sea surface and of the bottom of the water column, and the sun and sky radiance incident on the sea surface. Outputs can be various irradiances, K-functions and reflectances (Mobley & Sundman, 2000a,b). In this study the HYDROLIGHT model was parameterized with as much as possible relevant information for the in situ sites. The water constituents used in this work are chlorophyll, coloured dissolved organic matter, and tripton, where tripton is the ‘‘lifeless’’ component of the suspended matter. The inherent optical properties (IOP) are the properties of the medium itself regardless of the ambient light field; the IOPs are measured by active (i.e. having their own light source) optical instruments. The specific inherent optical properties (SIOP) are the inherent optical properties normalized to their concentration or in the case of coloured dissolved organic matter (CDOM) normalised to absorption at 440 nm.
The model used for the HYDROLIGHT simulations is the ‘‘ABCASE2’’ (a standard HYDROLIGHT model). This is a generic four-component IOP model that allows the user to define the component optical properties: the concentrations and IOPs of the four components (pure water, chlorophyll, CDOM and tripton). The total absorption, a TOT in this respect is represented by: * aTOT ¼ aWATER þ CCHL I a*PHY þ CCDOM I aCDOM þ C I a* TR
TR
ð4Þ
where a WATER=absorption due to water; C CHL=concentration of chlorophyll; a*PHY =specific absorption of phytoplankton (absorption of pigments normalized on chlorophyll concentration);
Fig. 4. Maximum depths at which Posidonia and Zostera should be detectable from Landsat TM for each of the sampling sites. Retrieval based on the measured substrate reflectances of Posidonia and Zostera and the measured K d at 550 nm (chosen to coincide with Landsat’s band 2 - the band with maximum water penetration), assuming a detecting threshold of 0.5% and a R V of 7.5%. The value of K d at 550 nm is reported in brackets for all sites.
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
C CDOM=concentrationof CDOM; a*CDOM =specific absorption of CDOM (normalised to absorption at 440 nm); C TR=concentration of tripton; a*TR =specific absorption of tripton. The following values were used: The ‘‘pure water’’ absorption values of Pope and Fry (1997) are used as the HYDROLIGHT default. The spectral absorption of phytoplankton (a PHY), tripton (a TR) and CDOM(a CDOM) were measured using a dual beam spectrophotometer with integrating sphere according to Clementson et al. (2001). The a TR was described as a normalized spectrum with an exponential slope, S TR: ð5Þ aTR ðkÞ ¼ CTR Ia* ðk0 Þexpð STR ðk k0 ÞÞ TR
and estimated from a linear regression of the log-transformed absorption versus wavelength where k 0 was set at 550 nm, a*TR (550) and S TR are reported in Table 1. These values are similar to those reported by Bukata et al. (1995) and Roesler et al. (1989) for inland and coastal waters. The spectral absorption a CDOM was retrieved following the same method as for a TR: aCDOM ðkÞ ¼ CCDOM Ia*CDOM ðk0 Þexpð SCDOM ðk k0 ÞÞ ð6Þ where k 0 was set at 440 nm, C CDOM is the a CDOM at 440 nm * thus a CDOM (440) is 1, and the S CDOM for each site are reported in Table 1. These values are also similar to those reported by Bukata et al. (1995) and Roesler et al. (1989) for inland and coastal waters. The entire scattering is attributed to the tripton component, so the scattering of chlorophyll is assumed to be 0. This is possible as the total backscattering estimated in situ with Hydroscat-6 (with customized band settings) is attributed to the total suspended matter (that contains algal biomass). There was assumed to be no internal source, caused by bioluminescence, nor inelastic scattering caused by (Chlorophyll and CDOM) fluorescence or Raman scattering. By using the laboratory absorption data and the in situ Hydroscat backscattering data some of these effects are implicit within the parameterization. Furthermore these effects are expected to be so small that it is more important to get optical closure without these effects. A possible sophistication could be running the model with these parameterizations. Incorporating these internal sources and elastic scattering sources has to be done with care and each of the parameterizations has to be valid for the waters under study. That went beyond the scope of this research. A semi-empirical RADTRAN atmospheric model implemented in HYDROLIGHT is used to model the atmosphere at the time of the in situ measurements. The wavelength selection is defined with corresponding Landsat bands spectral sensitivity input files. A semi-empirical sky model
421
and a wind speed of 5 m s 1 define the air –water surface boundary conditions. A solar zenith angle of 44 – 70- and a 0 –50% cloud cover representing the circumstances during the field campaign defined the sky condition ranges for the 4 Landsat images. The atmospheric parameters were obtained by modelled radio-sonde data for each month over a 25 year period from 1957– 1975 for Williamtown (45 km southwest of Wallis Lake) (Maher & Lee, 1977). The direct (solar) and diffuse (sky) components of the downwelling sky irradiance are directly calculated from the RADTRAN model. A file with RAMSES derived bottom reflectance was used for the boundary spectral condition of each site. The measured concentrations, the in situ measured downwelling irradiance and the in situ reflectances were used to simulate the subsurface L u / E d. Preference was given to estimating L u / E d instead of R(0) as the conversion factor Q of subsurface L u to subsurface E u has not been adequately measured for these Australian waters. To achieve optical closure between HYDROLIGHT and the in situ measured reflectance data, we needed to run various simulations experimenting with different parameterisations. Fig. 5 presents the HYDROLIGHT simulated spectra of L u / E d compared to the subsurface L u / E d measured in situ below the surface with a Ramses submersible spectroradiometer. The optical closure for 500 – 700 nm range of the numerical radiative transfer model with the in situ measurements confirms the parameterization and estimate of the specific inherent optical properties (SIOPs). Once this stage of optical closure has been reached, several permutations of concentrations can now be simulated as long as the parameterization of the model with SIOPs is correct. These simulations enable this research to expand knowledge beyond what is measured in situ. Thus a greater range of possible spectra of varying water columns over different substrates at different depths may now be calculated, which is essential for retrospective remote sensing applications. 2.4. Satellite imagery Landsat 5TM and 7ETM satellite imagery of Wallis Lake spanning 14 years was used to monitor the change in seagrass communities (Table 3). The Landsat archives were browsed in order to select cloud-free high quality imagery. The quality of the imagery was assessed on the absence of sun glint, wind-induced waves as well as river-outflowinduced turbidity affecting the seagrass ‘‘visibility’’ through the water column. The Landsat images were converted to top of atmosphere radiance using their ‘‘standard calibration files’’. The next step involves the correction of this at sensor measured upwelling radiance from Wallis Lake for atmospheric effects. Although the physics of atmospheric correction of remote sensing data over waters is essentially the same as for terrestrial targets, there are a few practical differences that need to be addressed. For any water body it is the signal coming from within the water body that is the desired signal,
422
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
Fig. 5. HYDROLIGHT simulation over Zostera substrate compared with in situ spectra and the Landsat spectrum for WL22-3 field site. In the ideal situation the HYDROLIGHT spectra and the in situ spectra match the Landsat spectrum. The spectra match to within the NEDR(0)E given in Fig. 6.
while the surface reflected signal is a signal that is considered as noise. Water bodies, in general, reflect less light than surrounding land and thus a highly accurate atmospheric correction is required (Dekker et al., 2001). A model based solution to this problem has been applied in this study using a ‘‘coastal Waters and Ocean MODTRAN-4 Based ATmospheric correction’’ (‘‘c-WOMBAT-c’’) procedure implemented in IDL/ENVI\ (Brando and Dekker, 2003). The atmospheric parameters were obtained by modelled radiosonde data for each month (Maher & Lee, 1977). These are only seasonal averages for this region. Therefore this atmospheric parameterization was iteratively refined using pseudo invariant feature targets within the imagery (deepest ocean, native forests and beach and dune sands). When performing such an iterative atmospheric correction method using either simulated reflectances using HYDROLIGHT or pseudo-invariant features, a criterion needs to be developed to assess when the iterative atmospheric correction is sufficiently accurate. For example, Fig. 5 shows a reasonable closure between the Landsat atmospherically and air – water interface corrected data to
Table 3 Landsat time series (L7 = L7 ETM+; L5 = L5 TM) Day Month
Year
Sensor Quality
12
September 2002 L7
21
February
1995 L5
30
March
1991 L5
18
February
1988 L5
High
Notes
Little sun glint or wind waves. Reasonable Little sun glint or wind waves, sensor striping visible, particularly in the rivers. Reasonable Little sun glint or wind waves, turbid river water. High Little sun glint or wind waves.
R(0) and the in situ measured R(0). How does one assess the term reasonable? In this study we elaborate on a method (previously described in Brando and Dekker (2003) and Dekker and Peters (1993) and further elaborated in Wettle et al. (2004)) that calculates the environmental noise equivalent reflectance difference [NEDR(0)E] in an image. Once the atmospheric correction has iteratively improved to within the NEDR(0)E limit it is not sensible to continue achieving a higher level of agreement as the noise in the image will be greater than any improvement. Thus, in order to understand the detection limits of an environmental variable with a remote sensor it is necessary to know or estimate the environmental Signal to Noise Ratio (SNR) as it exists for each image. One could refer to the instrument provider’s SNR, however these are invariably determined under laboratory conditions, often involving a bright lamp and an integrating sphere. In actual remote sensing environments there are sources of noise in the image data such as atmospheric variability, the air –water interface with swell, wave and wavelet induced reflections and refraction of diffuse skylight and direct sunlight. To estimate the sensitivity of the images from the two Landsat sensors (TM5 and ETM7) the environmental noise equivalent reflectance difference (NEDR(0)E) needs to be determined. The NEDR(0)E is calculated from the atmospherically and air – water interface corrected R(0) image according to Dekker and Peters (1993): NEDRð0 ÞE ¼ rð Rð0 ÞÞ
ð7Þ
where r(R(0)) is the standard deviation in each band over the most homogeneous area (possible) of optically deep water within the image, the size of the uniform area can be determined by increasing the number of pixels step by step (3 3, 5 5, 7 7, etc.) until r(L) reaches a first asymptotic
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
limit. Care has to be taken that whilst increasing the size of the uniform area no actual water body heterogeneities are included in the sampled area [see Wettle et al. (2004) for a more comprehensive discussion on NEDR(0)E]. These calculations were done on the Landsat images using the part of the Pacific Ocean in the imagery as a homogeneous relatively dark and representative surface for a training target; the results for the 2002 and 1988 images are presented in Fig. 6. It needs to be realized that as Landsat only has 256 digital levels of radiance detection available for the full reflectance range of 0– 100%, the theoretical limit is about 256 : 1. The digital level equivalent in reflectance terms (DLER(0)) are also reported in Fig. 6 for the two dates. In terms of reflectance Landsat 5 and Landsat 7 can resolve about 0.5– 0.7% reflectance differences These results indicate that Landsat 5 is less sensitive than Landsat 7. Because of the low quantization of the Landsat data, the DLER(0) is higher than the NEDR(0)E for TM bands 2 and 3 for both dates and thus DLER(0) represents the detecting threshold for the system. These result for the actual image radiometric accuracy together with the HYDROLIGHT simulations and the in situ measurements provided the best achievable optical closure (see Fig. 5). 2.5. Benthic substrate classification Wallis Lake has unvegetated substrates varying from muddy to silty to sandy to coarse sand. Vegetated substrates vary from lightly vegetated Halophila to densely vegetated areas of Zostera, Posidonia and various algae, to monospecies fields at various levels of density. The goal of the classification method was to be as objective as possible to increase the capacity for multitemporal and multisensor comparison of results (and in future multi-site comparison)
423
in such a complex area. In order to operationalise multitemporal remote sensing it is necessary to decouple fieldwork and satellite sensor imaging. This was one of the reasons for not attempting to achieve simultaneous field truth data and satellite data collection. The aim of the classification was to separate the benthic material within the lake into sets of spectral classes that represent the patterns and texture of the ecosystem. These classes, or their attributes, are created by the classifier, the classifier has been trained with the spectral classes collected from the imagery. For this purpose Regions Of Interest (ROIs) were selected based on a pseudo true colour (RGB) combination using the first three Landsat bands centred at 485, 560 and 660 nm. The ROIs were selected over targets that were spatially and spectrally homogeneous, covering a range of water depths and coinciding with field data or past study sites. In the first instance, it is important to obtain the statistically significant maximum number of spectral classes present in the image. Several supervised classification methods were trialled, and the Maximum Likelihood Classifier achieved the highest separation between classes. The distance based methods (Minimum Distance, Spectral Angle Mapper, etc.) performed poorly, most likely due to image noise effects on the distance retrieval. Thus a Maximum Likelihood Classification was run with ENVI 3.5 image processing software using all the mean spectra as input and applying a probability threshold value of 0.05. This was done with the Wallis Lake Landsat 5TM and 7ETM data, using the three visible bands. The near infrared band was judged to be overly influenced by sensor striping over water caused by the low sensor sensitivity and the low reflectance values, often less than 1%. The next step was to group classes that were spectrally similar and to link them to an environmentally relevant label. Firstly, the GPS location and associated substrate
Fig. 6. The environmental noise equivalent R(0) difference (NEDR(0)E) and the digital level equivalent in reflectance terms (DLER(0)) for Landsat5 TM and Landsat7 ETM+ over ocean water. Note: NEDR(0) is not a percentage error, but a difference expressed in percentages.
424
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
information with recent and historical substrate maps (Laegdsgaard, 2001; West et al., 1985) were used to identify the major groupings of substrate vegetation types. Secondly, the existing CSIRO spectral library of seagrasses and macroalgae as well as the Maximum Likelihood classification scheme was used. In all of the images, the spectral variation available in the Landsat data could be described by 10 – 15 spectral classes consisting of: 3 water types (river, turbid river and lake), 6 –11 vegetated substrate cover classes and 1 sand class. Postclassification, the spectral classes were additionally labelled using the parameterized HYDROLIGHT model. In addition a year after the spectral characterization fieldwork, two validation field trips were undertaken by the Great Lakes Council (GLC) and the NSW Department of Land and Water Conservation (DLWC) in October and November 2002 shortly after the Landsat 7 ETM 2002 data acquisition. These field trips were focused on the
further identification of these classes, in particular the classes not aligned with the spectral information available. Final classes were labelled according to the criteria described above and generally fell into sensible groupings (with knowledge of the lake ecosystem) with help from our existing field knowledge and past studies. In some images some pixels occurred where most benthic spectral information was lost due to sun glint, turbidity or wind induced waves and these pixels often formed their own classes across depth and substrate boundaries. After analysis, they were either discarded from the classification, or labelled to the closest associated spectral class.
3. Results and discussion Fig. 7 presents the substrate classification of the Landsat time series from 1988 to 2002. In most of the regions up to
Fig. 7. Benthic substrate classification of the Landsat time series. (a) Landsat 7 ETM+ classified image from 12 September 2002; (b) Landsat 5 TM classified image from 21 February 1995; (c) Landsat 5 TM classified image from 30 March 1991; (d) Landsat 5 TM classified image from 18 February 1988.
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
425
Fig. 7 ( continued ).
11 vegetated substrate classes; 3 water types (river, turbid river and lake) and 1 sand class were sufficient to describe the spectral variation. The quality of the Landsat image affected the quality of the resultant classification. Regions with visible striping impacted on the classification by mislabelling classes. High quality images (such as for 12 September 2002) were much less affected. Thus for change detection and analysis four fully classified images were available. The 2002 image was used to validate the methodology and to set the baseline. The other three images were used to evaluate the change detection and then derive the lake substrate cover evolution. Fig. 7a (Landsat 7 ETM+ on 12th September 2002) is the image closest to a field validation and was therefore processed first. The ROIs coincided with as many field positions as possible, where the substrate had been identified.
The results of the two validation field trips conducted in October and November 2002 of the September 2002 Landsat based classification are presented in Table 4. The field validation was undertaken with specific intent to check the results of the classification; in particular, classes not aligned with the field spectral measurements were investigated. It is important to note that these field campaigns were designed to check regions in the validation that were uncertain. The validation accuracy for the classification was calculated as follows. For each location the field characterization based on point measurements was compared with the Landsat TM classification result of the pixel corresponding with the field GPS location. For each location a score of the match was assessed (0 = no, 0.5 = next pixel, 1 = yes). The sum of the scores was normalised over the number of points checked.
426
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
The overall assessment of accuracy ranged from completely exact to 80% correct depending on the region of the lake visited (personal communication, Gerard Tuckerman). The two field validation campaigns therefore focused on smaller areas in the Landsat images where unidentified classes occurred or where many classes occurred together (that is, the interesting or problem areas). In some areas of most complexity, the field campaign of October 2002 gave a 76% match with the classified image of that area. In other areas of most complexity the field campaign of November 2002 matched 54% with the classified image area. The difficulty in interpreting the results of this comparison is that both the field GPS and the Landsat data have unclear spatial accuracies. After warping the 1988, 1991 and 1995 images to the 2002 data using ground control points, the geometric accuracy achieved was within a RMS of 1.5 pixels, equivalent to 45 m. The GPS used on the 10 October 2004 validation field work
was different from the GPS used on the 1 November 2004. The GPS used on 1 November did not have the required accuracy to adequately pin point field information with image data or a lesser accuracy than the one used on 10 October 2002, thus the validation results should be interpreted with caution. In addition, we are comparing field data from 10 October 2002 and 1 November 2002 with Landsat data of 12 September 2002. From an accurate analysis of the field trip results versus the classification results (full comparison tables are reported in Dekker et al., 2003a) for each location, it emerged that isolated and sparse Posidonia was mapped in the classification as Ruppia, due to not having a sparse Posidonia spectral signal in our spectral library, and Ruppia maps out the closest to sparse Posidonia in the Maximum Likelihood classification. Moreover, there is a tendency for Chara to be mapped as macroalgae which indicates the spectral library needs some refinement or a higher spectral resolution sensor
Fig. 7 ( continued ).
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
427
Fig. 7 ( continued ).
should be used. The classification results could be improved by including these areas in the next field based spectral library sampling. At the boundary between the optically deep water in the middle of the lake and the shallower parts, Ruppia is mapped. This could be a confusion with a sparse Posidonia class (see above) or the resolution of this class is at the threshold of substrate visibility, where the reliability of any optical remote sensing mapping method becomes less accurate. Table 4 Classification results for the least understood areas Day
Month
Year
Number of sites
Score (%)
12 1
October November
2002 2002
20 25
76 54
The results for other areas are significantly better, but not quantified. These results represent a Fworst case_. Several other sites were briefly visited and given a score of 80 – 100% (not further quantified).
In Fig. 7b (Landsat 5 TM data on 21st February 1995), there appears to be less macroalgae than in the 2002 image. Posidonia was classified in the southern part of the lake, where it is unlikely to occur according to local ecologists, and thus this Posidonia species location was removed from the images; this misclassification is possibly due to lower sensor data quality (e.g. areas of glint, turbidity etc.). Some Posidonia has been found (see Laegdsgaard, 2001; West et al., 1985), on the mid-western edge of the lake, and this area was mapped as Posidonia in the classification. Large homogeneous beds of Zostera and consistent beds of Ruppia and the Ruppia/Halophila mix appear in the mid and southern parts of the lake. In Fig. 7c (Landsat 5 TM data from 30 March 1991), as in the 1995 image, sensor striping is visible and again Posidonia was classified in the southern part of the lake were it is unlikely to occur (possibly due to reduced sensor quality). Large homogeneous beds of Zostera appear in the mid and southern parts of the lake and the consistent beds of
428
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
Fig. 8. Changes in substrate cover from 1988 – 2001 for (a) Zostera, (b) Posidonia and (c) Ruppia/Halophila. (
=loss,
= gain and
= no change).
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
Ruppia and the Ruppia/Halophila mix. Beds of Posidonia were mapped in the north and around Wallis Island, which are consistent with the past surveys and maps (Laegdsgaard, 2001; West et al., 1985). In Fig. 7d (Landsat 5 TM data from 18 February 1988), beds of Posidonia were mapped in the north and around Wallis Island, which are consistent with the past surveys and maps for this date. Large areas of Zostera covered the mid and south regions of the lake forming homogeneous beds adjacent to Ruppia and Ruppia/Halophila beds in the lake’s mid-east section. An unknown class (labelled Funknown_) was not identified from the field spectra but is likely to be similar to the Zostera/macroalgae class in the September 2002 image or Zostera with a mud or silt substrate. To identify the spatial change in seagrass cover over the 14 year period, the three main classes of interest (Zostera, Posidonia and the combined Ruppia/Halophila class) were selected by combining all subgroups: Zostera/Macroalgae, Sparse Zostera and Sand/Zostera into just one class of Zostera; Ruppia, Halophila/Ruppia/Sand and Sparse Halophila/Ruppia are combined into the class of Ruppia and Halophila. This presented the opportunity to look more
Fig. 9. Changes in Zostera cover (a) 1988 – 1991, (b) 1991 – 1995 ( = loss, was lost in 1991 but had re-grown by 1995.
429
closely at environmentally significant changes occurring in this coastal lake. We took care to exclude all occurrences, in all satellite images of optically deep water (i.e. where there was no substrate visibility), from this change detection analysis. The result of the change detection from 1988 to 2002 for the three main classes of interest are reported in Fig. 8. Zostera seems to be in decline; a significant loss from the central west basin of the lake (Coomba Bay) is evident from Fig. 8a as well as a major reduction from the central eastern basin. The Posidonia (Fig. 8b), Ruppia and Halophila (Fig. 8c) seem to be stable with no gross changes, however there is an apparent gain and loss in the rivers and channels leading into lake. Over the 14 years period, the submerged vegetation community however has undergone significant change as the Zostera has been replaced by macroalgae species like Chara and Nitella or is covered by dense epiphytes (see Fig. 7a). Healthy seagrass keep epiphytes under control by shedding leaves (Kendrick & Lavery, 2001) and through the maintenance of host grazers that consume the algae (Zupo et al., 2001). However when there is a change in the balance of the ecosystem, such as an increase of nutrients available,
= gain and
= no change). The highlighted region shows a scar where Zostera
430
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
Table 5 Species
Gain
Loss
Area (km2)
Area (%)
No change
Area (km2)
Area (km2)
Area (%)
Area (%)
2
(a) Seagrass change from 1988 until 2002 over the entire area of visible substrates covering a total area of 33.87 km (the dashed box in Fig. 1) Zostera 1.88 11.15 9.50 56.30 6.48 38.37 Posidonia 2.23 13.19 1.63 9.63 0.52 3.06 Ruppia/Halophila/sand 3.39 20.09 1.43 8.47 2.39 14.15 Other 26.37 55.57 21.32 25.60 24.49 44.42 Species
2002 Area (km2)
1995 Area (%)
Area (km2)
1991 Area (%)
Area (km2)
1988 Area (%)
Area (km2)
Area (%)
2
(b) Seagrass coverage in 1988, 1991, 1995 and 2002, over the entire area of visible substrates covering a total area of 33.87 km (the dashed box in Fig. 1) Zostera 11.17 32.98 10.21 30.16 10.24 30.22 16.69 49.27 Posidonia 3.39 10.00 2.73 8.07 6.05 17.85 2.54 7.50 Ruppia/Halophila/sand 6.22 18.37 5.62 16.58 5.55 16.39 4.51 13.32 Other 13.09 38.65 15.31 45.19 12.03 35.54 10.13 29.91 (c) Area of seagrasses and optically deep water class coverage as defined in the Classification for the four dates within the boundary specified in by the shaded area in Fig. 1 Zostera 1.212 7.4 0.180 1.1 1.165 7.1 1.557 9.6 Posidonia 1.422 8.7 1.493 9.2 1.889 11.6 0.988 6.1 Ruppia/Halophila/sand 2.249 13.8 2.414 14.8 1.345 8.3 1.478 9.1 Other 11.409 70.1 12.205 74.9 11.893 73.0 12.269 75.2 The FOther_ class includes optically deep waters, unvegetated or non-seagrass macrophytes. The total area of water within the boundary specified by the shaded area in Fig. 1 is 16.292 km2.
epiphytes can increase, reducing the available light for photosynthesis and therefore the seagrass growth rate or even causing the seagrass to die. To illustrate the variability of change over time, a change detection map for Zostera was produced for the 1988 – 1991 and 1991– 1995 images. The 1988– 1991 change detection image (Fig. 9a) clearly indicates Zostera loss from the central eastern basin. Zostera change is not uncommon and could be the result of seasonal or environmental conditions, although both images were acquired approximately in the same season (February 1988 and March 1991). By 1995 (Fig. 9b), Zostera had regrown covering the scar feature identified in 1991 (Fig. 9a). The change in the three major seagrass classes (Zostera, Posidonia and Ruppia/Halophila) and the Fother_ class (which includes the optically deep waters, unvegetated or
non-seagrass macrophytes) from 1988 to 2002 is reported in Table 5a and illustrated in Fig. 8a,b,c. The coverage of these 4 classes for each of the 4 years is reported in Table 5b. The change in Zostera distribution may be enhanced as the classification has identified several macroalgae classes due to a heavy infestation of epiphytes. If Zostera still exists under this macroalgae class then it needs to be identified and labelled accordingly. It would need a higher spectral or spatial resolution to resolve epiphytes on seagrasses. The shift from pure Zostera colonies to the Zostera/macroalgae class (or epiphyte infested Zostera colonies) represents an ecological indication of degradation of the system. An area of 16.292 km2 at the entrance of the lake (shaded area in Fig. 1) was chosen to be a representative area for testing future multi-temporal analyses and calculating
Table 6 Sub-group
1988 Area (km2)
2002 Area (%)
Area (km2)
Area (%) a
(a) Area of Zostera as defined in the classifications for the images of 1988 and 2002 within the boundary specified by the shaded area in Fig. 1 Zostera 0.511 32.8 0.246 20.3 Sand/sparse Zostera – 0.0 0.591 48.8 Zostera/macroalgae 0.797 51.2 0.005 0.4 Sparse Zostera 0.248 16.0 0.371 30.6 Total 1.556 100 1.213 100 (b) Area of Ruppia/Halophila as defined in the classifications for the images of 1988 and 2002 within the boundary specified by the shaded area in Fig. 1 b Ruppia 0.467 31.6 0.455 20.2 Sparse Halophila/Ruppia 1.011 68.4 1.794 79.8 Total 1.478 100 2.249 100 a
The percentages are based on the total Zostera in the region for each date (see Table 5c). The percentages are based on the total Ruppia/Halophila in the region for each date (see Table 5c).
b
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
seagrass coverage. This area is least susceptible to high turbidity levels by river water or resuspension potentially masking the substrate due to tide-induced regular clear ocean water flushing. The area covered by three major seagrass classes are displayed in Table 5c as total area covered and as a percentage of the total water covered classified region of 16.292 km2 for each year. There is an overall loss of Zostera, a small increase in Posidonia and Ruppia/ Halophila. The Ruppia/Halophila class has increased which seems to be partly at the cost of Zostera. Halophila are adapted to levels of low light and Ruppia have been found to occur in a range of turbidities and salinities and have been found to commonly occur in areas of high freshwater input (Coles et al., 2003), therefore it is possible that this change from Zostera to the Ruppia/Halophila class could be due to changing conditions (i.e. increased turbidity and salinity) within the lake. The area covered by the four sub-groups of the Zostera class and the two sub groups of Ruppia/Halophila are reported in Table 6a,b as total area in km2 and as a percentage of either the total Zostera or Ruppia/Halophila in the region. While the Zostera classes have decreased and the Ruppia/Halophila classes have increased, the sparser coverage of both have increased in comparison to the higher density classes.
4. Conclusions and recommendations Remote sensing is often seen as an expensive methodology. In reality the traditional field survey method is the most expensive if all costs are taken into account. Many of the field sampling costs fall under standard budgets whereas remote sensing is often seen as an extra cost. However, labour costs are increasing, occupational health and safety regulations are being enforced, and the spatial and temporal scale at which environmental processes need to be mapped is increasing. On the other hand, the cost of remote sensing acquisition and image processing are decreasing and it is anticipated that remote sensing will soon become a viable alternative to more traditional methods. This project analysed satellite image classifications spanning 14 years over Wallis Lake, detecting significant change in seagrass and macroalgal communities. For this lake remote sensing based mapping has significant advantages over traditional techniques as it is spatially comprehensive. Especially when used to detect change, remote sensing becomes even more cost-effective, as one methodology is applied to all images. Future improvements in methodology can be retrospectively applied. The results presented here can be improved by including areas with new classes or unsampled species in the next field based spectral library sampling, thus, the remote sensing image enables one to focus on areas needing one more (perhaps final) field campaign. In future it is
431
recommended to carry out one field survey prior to or during the satellite image capture and one survey postclassification as it is only at this stage that the spatial information in the remote sensing image can be used to identify gaps in knowledge. If this process is followed valid spatially comprehensive information can be obtained by remote sensing. After this second fieldwork the necessity for ongoing fieldwork during satellite image capture is greatly reduced or not required anymore, unless new species enter the system. Considering that in situ seagrass monitoring only started in 1996 for Wallis Lake, the use of archival Landsat or SPOT data is the only way to retrospectively establish an environmental baseline in 1988. Remote sensing methodologies based on digital data and using methodologies that are well documented such as this study, have the additional advantage of being objective and repeatable. Satellite sensor data with higher spatial resolution, and increased signal to noise ratio are currently available. These improvements will result in higher quality in the derived information products due to the increased spatial accuracy. An assessment of hyperspectral data would provide higher separability and spatial and spectral variation than Landsat imagery. Although there would be a significant increase in the cost of data acquisition, hyperspectral data provide many more opportunities than multispectral imagery. Hyperspectral data have been used to successfully map rock platform vegetation (Dekker et al., 2003b), mangroves, salt flats (Phinn et al., 1999) and water quality parameters such as total suspended sediment (TSS), chlorophyll and coloured dissolved organic matter (CDOM) concentrations (Brando & Dekker, 2003; Dekker et al., 2001). The method followed here provides an excellent base for further expansion of the spectral library of optical water properties as well as of substrates. Moreover as all optical field and laboratory data were collected hyperspectrally this method can be applied to any current and future multispectral and hyperspectral sensor imagery for Wallis Lake and other similar lakes providing a spectral library database for future studies and applications. This research did suffer from the lack of an integrated methodology for validating remotely sensed images with field data in a shallow tidal coastal lake. Such a methodology should then take into account the specific two dimensional information that is contained within 1 pixel, for example, a 30 by 30 m area in the case of Landsat, because the question needs to be solved how to compare a 30 by 30 m spatial average with points or transects? Another interesting issue deriving from this research is the following: how does mapping 1% (940 m2) of the area using points and transects with an accuracy (to account for geolocation and interpretation errors), e.g. of 95%, compare to a remotely sensed image (100% mapping or 94 km2) with 80% or 75% accuracy? Indeed which type of information is environmentally more relevant? The accurate 1% field classification extrapolated to 100% areal
432
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433
coverage or the less accurate spatially comprehensive remote sensing based classification? The low radiometric sensitivity of the Landsat sensor system for lake waters and visible substrate cover causes the classified images not always to be accurate at pixel to pixel scale, but highly effective at groups of pixels scale—which is the environmentally adaptive management scale.
Acknowledgments This work was supported by the Great Lakes Council, the NSW Department of Land and Water Conservation, the Cooperative Research Centre for Coastal Zone Estuary and Waterway Management and CSIRO Land and Water. We wish to thank Gerard Tuckerman (Great Lakes Council), Dr. Pia Laegdsgaard and Graham Carter (DLWC, NSW). The spectrophotometric measurements were carried out by Lesley Clementson at CSIRO Marine Research. The help of Nicole Pinnel, Liis Sipelglass and Cesar Urrutia during the time they spent at CLW is appreciated.
References Ackleson, S. G., & Klemas, V. (1987). Remote-sensing of submerged aquatic vegetation in lower Chesapeake Bay—A comparison of Landsat MSS to TM imagery. Remote Sensing of Environment, 22(2), 235 – 248. Borum, J., Duarte, C. M., Krause-Jensen, D., & Greve, T. M. (2004). European seagrasses: An introduction to monitoring and management, The M&MS project, September 2004, Internet version. http://www. seagrasses.org Ben Moussa, H., Viollier, M., & Belsher, T. (1989). Te´le´de´tection des algues macrophytes de l’Archipel de Mole´ne (France) radiome´trie de terrain et application aux donne´es du satellite SPOT. International Journal of Remote Sensing, 10(1), 53 – 69. Brando, V. E., & Dekker, A. G. (2003). Satellite hyperspectral remote sensing for estimating estuarine and coastal water quality, IEEE Trans. IEEE Transactions on Geoscience and Remote Sensing, 41(6), 1378 – 1387. Bukata, R. P., Jerome, J. J., Kondratyev, K. Ya, & Pozdnyakov, D. V. (1995). Optical properties and remote sensing of inland and coastal waters. Boca Raton’ CRC Press. Chauvaud, S., Bouchon, C., & Manie`re, R. (2001). Cartographie des bioce´noses marines de Guadeloupe a` partir de donne´es SPOT re´cifs coralliens, phane´rogames marines, mangroves. Oceanologica Acta, 24, S3 – S16. Claasen, D., van, R., Jupp, D. L. B., Bolton, J., & Zell, L. D. (1984). An initial investigation into the mapping of seagrass and water colour with CZCS and Landsat in North Queensland, Australia. Proc. 10th Int. Symp. Machine Proc. Remotely Sensed Data, Indiana (pp. 190 – 201). Clementson, L. A., Parslow, J. S., Turnbull, A. R., McKenzie, D. C., & Rathbone, C. A. (2001). The optical properties of waters in the Australasian sector of the southern ocean. Journal of Geophysical Research, 106, 31611 – 31626. Coles, R., McKenzie, L., & Campbell, S. (2003). The seagrasses of eastern Australia. In E. P. Green, & F. T. Short (Eds.), World atlas of seagrasses, prepared by the UNEP World Conservation Monitoring Centre. Berkeley, USA’ University of California Press. Cracknell, A. P. (1999). Remote sensing techniques in estuaries and coastal zones—An update. International Journal of Remote Sensing, 19(3), 485 – 496.
Dekker, A. G., Anstee, J. M., & Brando, V. E. (2003a). Seagrass change assessment using satellite data for Wallis Lake, a consultancy report to the Great Lakes Council and Department of Land and Water Conservation. Technical report 13/03. CSIRO Land and Water, Canberra. In Internet version at: http://www.clw.csiro.au/publications/ technical2003/ Dekker, A. G., Brando, V. E., Anstee, J. M., Fyfe, S., Malthus, T. J. M., & Karpouzli, E. (in press). Remote sensing of seagrass ecosystems: Use of spaceborne and airborne sensors. In A. W. D. Larkum, C. M. Duarte, & Orth, R. J. (Eds.), Biology of seagrasses: A treatise. Amsterdam, Elsevier. Dekker, A. G., Brando, V. E., Anstee, J. M., Pinnel, N., Kutser, T., Hoogenboom, H. J., et al. (2001). Imaging spectrometry of water. Imaging spectrometry: Basic principles and prospective applications, vol. IV (pp. 307 – 359). Dordrecht’ Kluwer Academic Publishers. Dekker, A. G., Byrne, G. T, Brando, V. E., & Anstee, J. M. (2003b). Hyperspectral mapping of intertidal rock platform vegetation as a tool for adaptive management. Technical report 09/03. CSIRO Land and Water, Canberra. Internet version at: http://www.clw.csiro.au/publications/technical2003/ Dekker, A. G., & Peters, S. W. M. (1993). The use of the thematic mapper for the analysis of eutrophic lakes: A case study in The Netherlands. International Journal of Remote Sensing, 14, 799 – 822. Duarte, C. M., Alvarez, E., Grau, A., & Krause-Jensen, D. (2004). Which monitoring strategy should be choosen? In J. Borum, C. M. Duarte, D. Krause-Jensen, & T. M. Greve (Eds.), European seagrasses: An introduction to monitoring and management, The M&MS project, September 2004, Internet version at: http://www.seagrasses.org. Ferguson, R. L., & Korfmacher, K. (1997). Remote sensing and GIS analysis of seagrass meadows in North Carolina, USA. Aquatic Botany, 58, 241 – 258. Fyfe, S. K. (2003). Spatial and temporal variation in spectral reflectance: Are seagrass species spectrally distinct? Limnology and Oceanography, 48(1, part 2), 464 – 479. Kendrick, G. A., Hegge, B. J., Wyllie, A., Davidson, A., & Lord, D. A. (2000). Changes in seagrass cover on success and parmelia banks, Western Australia between 1965 and 1995. Estuarine, Coastal and Shelf Science, 50, 341 – 353. Kendrick, G. A., & Lavery, P. S. (2001). Assessing biomass, assemblage structure and productivity of algal epiphytes on seagrasses. In F. T. Short, & R. G. Coles (Eds.), Global seagrass research methods (pp. 199 – 222). Amsterdam’ Elsevier. Kirk, J. T. O. (1994). Light and photosynthesis in aquatic ecosystems (pp. 1 – 509). Cambridge, UK’ University Press. Kirkman, H. (1996). Baseline and monitoring methods for seagrass meadows. Journal of Environmental Management, 47, 191 – 201. Krause-Jensen, D., Quaresma, A. L., Cunha, A. H., & Greve, T. M. (2004). How are seagrass distribution and abundance monitored? In J. Borum, C. M. Duarte, D. Krause-Jensen, & T. M. Greve (Eds.), European seagrasses: An introduction to monitoring and management (pp. 345 – 350), The M&MS project, September 2004, Internet version at: http://www.seagrasses.org. Laegdsgaard, P., (2001). A field guide for the identification and monitoring of the seagrasses and macroalgae in Wallis Lake. Land and Water Conservation, Centre for Natural Resources, NSW Government. Lee Long, W. J., McKenzie, L. J., Rasheed, M. A., & Coles, R. G. (1996). Monitoring seagrasses in tropical ports and harbours. In J. Kuo, R. C. Phillips, D. I. Walker, & H. Kirkman (Eds.), Seagrass biology: Proceedings of an international workshop, Rottnest Island, Western Australia, 25 – 29 January 1996 (pp. 345 – 350). Lee Long, W. J., Roder, C. A., McKenzie, L. J., & Huntley, A. J. (1998). Preliminary evaluation of an acoustic technique for mapping tropical seagrass habitats, research publication no. 52. Great Barrier Reef Marine Park Authority. Liceaga-Correa, M. A., & Euan-Avila, J. I. (2002). Assessment of coral reef bathymetric mapping using visible Landsat Thematic Mapper data. International Journal of Remote Sensing, 23(1), 3 – 14.
A.G. Dekker et al. / Remote Sensing of Environment 97 (2005) 415 – 433 Lubin, D., Li, W., Dustan, P., Mazel, C., & Stamnes, K. (2001). Spectral signatures of coral reefs: Features from space. Remote Sensing of Environment, 75, 127 – 137. Macleod, R. D., & Congalton, R. G. (1998). A quantitative comparison of change-detection algorithms for monitoring eelgrass from remote sensing data. Photogrammetric Engineering and Remote Sensing, 64(3), 207 – 216. Maher, J. V., & Lee, D. M. (1977). Upper air statistics Australia: Surface to 5 mb, 1957 to 1975. Department of Science, Bureau of Meteorology, AGP, Canberra. Maritorena, S., Morel, A., & Gentili, B. (1994). Diffuse reflectance of oceanic shallow waters: Influence of water depth and bottom albedo. Limnology and Oceanography, 39(7), 1689 – 1703. Mobley, C. D. (1994). Light and water: Radiative transfer in natural waters. London’ Academic Press. Mobley, C. D., & Sundman, L. K. (2000a). HYDROLIGHT 4.1 users’ guide, WA (pp. 85). London’ Sequoia Scientific, Inc. Mobley, C. D., & Sundman, L. K. (2000b). HYDROLIGHT 4.1 technical documentation, WA (pp. 76). Sequoia Scientific, Inc. Mumby, P. J., & Edwards, A. J. (2002). Mapping marine environments with IKONOS imagery. Mumby, P. J., Green, E. P., Edwards, A. J., & Clarke, C. D. (1997). Coral reef habitat mapping: How much detail can remote sensing provide? Marine Biology, 130, 193 – 202. Mumby, P. J., Green, E. P., Edwards, A. J., & Clark, C. D. (1999). The cost-effectiveness of remote sensing for tropical coastal resources assessment and management. Journal of Environmental Management, 55, 157 – 166. Norris, J. G., & Wyllie-Echeverria, S. (1997). Estimating maximum depth distribution of seagrass using underwater videography. 4th Int. Conf. Remote Sens. Mar. Coastal Environ., Orlando, Florida, 17 – 19 March 1997, Vol. I (pp. 603 – 610). Norris, J. G., Wyllie-Echeverria, S., Mumford, T., Bailey, A., & Turner, T. (1997). Estimating basal area coverage of subtidal seagrass beds using underwater videography. Aquatic Botany, 58, 269 – 286. NSW Public Works Dept. (1990). Coastline management manual (pp. 114). Syndey’ NSW Government. Internet version at: http://www.deh.gov.au/ coasts/publications/nswmanual/index.html Philpot, W. D., & Vodacek, A. (1989). Laser-induced fluorescence: Limits to the remote detection of hydrogen ion, aluminum and dissolved organic matter. Remote Sensing of Environment, 29, 51 – 65. Phinn, S. R., Hess, L., & Finlayson, C. M. (1999). An assessment of the usefulness of remote sensing for wetland inventory and monitoring in Australia. In C. M. Finlayson, & A. G. Speirs (Eds.), Techniques for
433
enhanced wetland inventory and modelling, supervising scientist report 147 (pp. 44 – 83). Canberra’ Supervising Scientist. Pope, R. M., & Fry, E. S. (1997). Absorption spectrum (380 – 700 nm) of pure water: II. Integrating cavity measurements. Appl.Opt., Vol. 36 (pp. 8710 – 8723). Robblee, M. B., Barber, T. R., Carlson Jr., P. R., Durako, M. J., Fourqurean, J. W., Muehlstein, L. K., et al. (1991). Mass mortality of the tropical seagrass Thalassia testudinum in Florida Bay USA. Marine Ecology. Progress Series, 71, 297 – 299. Roesler, C. S., Perry, M. J., & Carder, K. L. (1989). Modeling in situ phytoplankton absorption from total absorption spectra in productive inland marine waters. Limnology and Oceanography, 34, 1510 – 1523. Shepherd, S. A., McComb, A. J., Bulthuis, D. A., Neverauskas, D. A., Steffensen, D. A., & West, R. (1989). Decline of seagrasses. In A. W. D. Larkum, A. J. McComb, & S. A. Shepherd (Eds.), Biology of seagrasses. A treatise on the biology of seagrass with special reference to the Australian region (pp. 346 – 393). Amsterdam’ Elsevier. Short, F. T., & Coles, R. G. (Eds.) (2001). Global seagrass research methods. Amsterdam’ Elsevier. Ward, D. H., Markon, C. J., & Douglas, D. C. (1997). Distribution and stability of eelgrass beds at Izembek Lagoon Alaska. Aquatic Botany, 58, 229 – 240. West, R. J., Thorogood, C., Walford, T., & Williams, R. J. (1985). An estuarine inventory for New South Wales, Australia. NSW’ Department of Agriculture. Wettle, M., Brando, V. E., & Dekker, A. G. (2004). A methodology for retrieval of environmental noise equivalent spectra applied to four Hyperion scenes of the same tropical coral reef. Remote Sensing of Environment, 93, 188 – 197. WLCMP (2001). Wallis Lake catchment management plan—volume 1—State of the catchment report Internet version at. http://www. greatlakes.nsw.gov.au/Environ/wlcmp/wIndex.htm Zainal, A. J. M., Dalby, D. H., & Robinson, I. S. (1993). Monitoring marine ecological changes on the east coast of Bahrain with Landsat TM. Photogrammetric Engineering and Remote Sensing, 59, 415 – 421. Zhang, X. (1998). On the estimation of biomass of submerged vegetation using Landsat thematic mapper (TM) imagery: A case study of the Honghu, PR China. International Jounal of Remote Sensing, 19(1), 11 – 20. Zupo, V., Nelson, W. G., & Gambi, M. C. (2001). Measuring invertebrate grazing on seagrasses and epiphytes. In F. T. Short, & R. G. Coles (Eds.), Global seagrass research methods (pp. 271 – 292). Amsterdam’ Elsevier.
