Towards spatial assessment of carbon sequestration in peatlands: spectroscopy based estimation of fractional cover of three plant functional types

: Peatlands accumulated large carbon stocks as peat in historical times. Currently however, many peatlands are on the verge of becoming sources with their carbon sequestration function becoming sensitive to environmental changes such as increases in temperature, decreasing water table and enhanced nitrogen deposition. Long term changes in vegetation composition are both, a consequence and indicator of future changes in carbon sequestration. Spatial continuous accurate assessment of the vegetation composition is a current challenge in keeping a close watch on peatland vegetation changes. In this study we quantified the fractional cover of three major plant functional types (Sphagnum mosses, graminoids, and shrubs) in peatlands, using field spectroscopy reflectance measurements (400–2400 nm) on 25 plots differing in plant functional type cover. The data was validated using point intercept methodology on the same plots. Our results showed that the detection of open Sphagnum versus Sphagnum covered by vascular plants (shrubs and graminoids) is feasible with an R² of 0.81. On the other hand, the partitioning of the vascular plant fraction into shrubs and graminoids revealed lower correlations of R² of 0.54 and 0.57, respectively. This study was based on a dataset where the reflectance of all main plant functional types and their pure components within the peatland was measured at local spatial scales. Spectrally measured species or plant community abundances can further be used to bridge scaling gaps up to canopy scale, ultimately allowing upscaling of the C balance of peatlands to the ecosystem level. Abstract Peatlands accumulated large carbon stocks as peat in historical times. Currently how-ever, many peatlands are on the verge of becoming sources with their carbon sequestration function becoming sensitive to environmental changes such as increases in temperature, decreasing water table and enhanced nitrogen deposition. Long term 5 changes in vegetation composition are both, a consequence and indicator of future changes in carbon sequestration. Spatial continuous accurate assessment of the vegetation composition is a current challenge in keeping a close watch on peatland vegetation changes. In this study we quantiﬁed the fractional cover of three major plant functional types ( Sphagnum mosses, graminoids, and shrubs) in peatlands, using ﬁeld 10 spectroscopy reﬂectance measurements (400–2400 nm) on 25 plots di ◆ ering in plant functional type cover. The data was validated using point intercept methodology on the same plots. Our results showed that the detection of open Sphagnum versus Sphagnum covered by vascular plants (shrubs and graminoids) is feasible with an R 2 of 0.81. On the other hand, the partitioning of the vascular plant fraction into shrubs 15 and graminoids revealed lower correlations of R 2 of 0.54 and 0.57, respectively. This study was based on a dataset where the reﬂectance of all main plant functional types and their pure components within the peatland was measured at local spatial scales. Spectrally measured species or plant community abundances can further be used to bridge scaling gaps up to canopy scale, ultimately allowing upscaling of the C balance 20 of peatlands to the ecosystem level.


Introduction
Peatlands have been widely recognized as one of the world's largest terrestrial stores of organic carbon (Botch et al., 1995;Turunen et al., 2002). The extensive peat deposits show the historical role of peatlands as sinks for atmospheric carbon (C), but at gases, such as CO 2 and CH 4 . As a substantial part of the peatlands can be found on the northern hemisphere, especially in the region where large changes in temperature, precipitation and nitrogen deposition are expected, the response of peatlands to environmental change has warranted much scientific attention (Bubier et al., 2007;Moore et al., 2002;Wiedermann et al., 2007). 5 Most studies indicate that major changes in carbon sequestration rate coincide with changes in vegetation composition (Strack et al., 2006). In general decreases in water table, increases in temperature as well as enhanced N deposition favour vascular plant species and, ultimately, have a negative impact on the bryophyte community through increased competition for light (Bubier et al., 2007;Wiedermann et al., 2007). Further-10 more, increases in nutrient concentration and temperature seem to favour graminoid species over ericoid species (Weltzin et al., 2000(Weltzin et al., , 2003. As bryophytes and vascular plants di◆er in their litter degradability (Dorrepaal et al., 2005), changes in their fractional covers, may have direct repercussions for the ecosystems longer term carbon sequestration rate. Particularly the cover and productivity of the bryophyte component, 15 dominated by the genus Sphagnum (Bubier et al., 2007), is of major importance to the carbon sequestration potential of the ecosystem. Sphagnum is the main peat former due to its recalcitrant litter and acts as an ecosystem engineer dictating the growth environment of vascular plants by regulating the surface hydrology and nutrient availability to a large extent (Belyea and Baird, 2006). On the whole, the fractional covers 20 of bryophytes, graminoids and ericoid shrubs give a fair indication of the carbon sequestration potential, and changes in their relative frequencies may serve as an early warning system for ensuing changes in C-sequestration. In addition the recognition of pure Sphagnum patches in a peatland vegetation may be used in the assessment of near-surface wetness (Harris et al., 2006), which exerts an important control on the 25 carbon balance of peatland soils. Dry conditions may lead to dramatic reductions in Sphagnum growth rates and carbon fixation. Near-surface water content is an indicator of the thickness of the unsaturated zone in peat soils and thus relates to the aerobic (fast) versus anaerobic (slow) decay rates in peatland soils. Finally the near-surface wetness is related to the water table below the peatland surface, determining the size of the unsaturated zone. A thin unsaturated zone generally leads to a higher CH 4 fluxes from the peatland surface.
The main three plant functional types (PFT) in peatlands (mosses, grasslike species (graminoids), shrubs) represent groups of di◆erent structure and biochemistry, such 5 as leaf construction (mosses lacking a vascular system, versus monocots, and dicots), water and pigment content, and architecture. These di◆erences result in distinct spectral characteristics at leaf and canopy level, such as the reflectance peak around 640 nm for red moss species and the low NIR reflectance of the ericoids (shrubs). Various studies highlight the spectral reflectance properties of Sphagnum mosses and 10 the influence of the water content and moisture conditions on their reflectance under laboratory conditions (Bryant and Baird, 2003;Bubier et al., 1997;Harris et al., 2005;Vogelmann et al., 1993). Only few studies analysed spectrometer data of peatlands at the ecosystem scale, thus including mixtures of Sphagnum, shrub, graminoid, and tree cover. Harris et al. (2006) inferred proxies for near-surface wetness to map e◆ects of 15 water stress on Sphagnum, based on a variety of field and airborne sensors (Milton et al., 2008), including an Analytical Spectral Device (ASD) FieldSpec Pro spectroradiometer and the Compact Airborne Spectrographic Imager (CASI). Recently, tree and shrub leaf area indices were inferred for a peatland through an approach combining in situ LAI measurements, field spectrometer data, and Landsat TM imagery (Sonnentag 20 et al., 2007). Spectral di◆erences between vegetation types were successfully used in other ecosystems, such as floodplains, to produce vegetation maps of structurally or chemically divergent vegetation based on imaging spectrometer data . Plant pigment and non-pigment retrieval at leaf and canopy level are becoming increasingly possible (Ustin et al., 2008), resulting in spatially distributed 25 and continuous vegetation characteristics that can then be assimilated into processoriented ecosystem models or regional climate models to overcome the limitations as given by maps with discrete land cover classes (Schaepman, 2007). Before employing coarser resolution imaging spectrometer over peatland areas to map the vegetation BGD 5,2008  Interactive Discussion composition, it is therefore useful to compile a spectral library in situ, allowing to constrain model inversion approaches for inferring fractional abundance or biochemistry, or assess the scalability of the approach from plant to canopy level (e.g. Kalacska et al., 2007;Bojinski et al., 2003). In this study we establish a non-destructive methodology to derive fractional cover 5 and biomass of the main plant functional groups within peatlands (Sphagnum, ericoid shrubs and graminoids) using field spectroscopy. To this end we selected field plots di◆ering in PFT fractional cover and related spectral reflectance data of these plots with fractional cover estimates inferred from point intercept data and destructive measurements on dry biomass. heath with patches of ombrotrophic (rain-fed) peatland vegetation. Five plots (1-5) were selected along five transects (A-E), situated along the water table gradient, going from relatively wet plots that were mostly dominated by Sphagnum with sparse graminoid cover to relatively dry plots with a continuous Sphagnum cover that were dominated by ericoid vegetation (Fig. 1). In between these extremes we selected plots 20 co-dominated by ericoids and graminoids. All plots had a (near) continuous Sphagnum layer (>95% cover), with one exception with ca. 80% Sphagnum cover (plot A5 was mounted at 65 cm height on a tripod; care was taken that no shade was cast on the measurement area. Using the instrument's 25 field of view (FOV), the ground field of view (GFOV) covered was approximately 29 cm in diameter. To reduce the impact of changing irradiance conditions throughout the day, a reference spectrum was collected from a white spectralon panel before the measurement of each plot. Following the 5 terminology of Schaepman-Strub et al. (2006), the data acquired with the field spectrometer correspond to hemispherical-conical reflectance factors (HCRF). No further corrections for the changes in solar zenith angle, or partitioning of the direct to di◆use irradiance were performed. On 8 June 2006 spectral reflectance measurements were taken from the dominant 10 species to obtain pure endmember spectra, following the same measurement protocol as 4 June 2006. To this end homogeneous vegetation outside the previously measured mixed plots was selected with a high fractional cover (>80%) of the target species. The spectra were visually controlled for their quality and measurements outside one standard deviation of the five measurements per plot were excluded. The remaining 15 measurements of each plot were averaged to the final plot spectrum. Wavelength regions influenced by atmospheric water vapour were excluded from further analysis. This resulted in a wavelength range of 400-1360 nm, 1410-1800 nm, and 1960-2400 nm (1793 spectral bands in total) used in the data analysis.
2.5 Multiple endmember spectral mixture analysis (MESMA) 20 Spectral mixing analysis is based on the assumption that the reflectance of an observed surface is a combination of the reflectance of the single components of the surface, weighted according to their abundance (Adams et al., 1986). In most studies it is further assumed that this mixture is linear and that the multiple scattering is negligible. The model can be expressed as BGD 5,2008 Fcover estimation of plant functional types in peatlands where R ,obs is the observed reflectance factor of the surface in the spectral band , f i is the fraction of the endmember, N is the number of endmembers, and " is the residual error. The quality of the unmixing is indicated by the root mean square error (RMSE) of the model fit over all spectral bands (M), calculated as Based on the above assumptions, we hypothesize that the linear combination of the reflectance of the single species according to their fractional cover would equal the fractional cover as recorded by the first-hit data using the point intercept methodology. We used the multiple endmember spectral mixture analysis (MESMA) algorithm to infer the fractional cover of the three plant functional types based on the measured 10 plot hemispherical-conical reflectance factors (Roberts et al., 1998). Therefore, the MESMA algorithm was selected where three endmember groups can be indicated containing a set of pure spectra. The algorithm searches for the best spectrum within each endmember group. The minimum non-shade fraction was constrained to 0, the maximum non-shade fraction to 1. Negative fractions were allowed for the shade end-15 member. No further restrictions were applied (e.g. RMSE, max. shade fraction, max. residual). The MESMA output contains the model assigned according to the best fit to the plot level spectrum (i.e. the name of the endmember spectra chosen for the best fit), the abundances of each of the endmember spectra, as well as the mean error of the model fit (RMSE). The algorithm also indicates cases for which no model could be 20 assigned with the indicated constraints.
Slightly di◆ering implementations of the unmixing approach exist (Plaza et al., 2004). We also evaluated the linear unmixing algorithm as implemented in ENVI (ITT Vis, Boulder, USA) which did not reveal satisfactory results. The main reason being that the endmember selection is not varying with the observed spectrum, but is fixed to 25 the total number of available endmember spectra, thus all endmembers get a fraction 1300 assigned. MESMA selects a subset of endmember spectra which best match up best the plot level spectrum and only assigns and abundances to this selection.

Endmember selection and aggregation
From the five measurement locations containing single dominant species the most representative were selected to obtain pure reflectance spectra. For Sphagnum we 5 selected two homogenous plots of the red species Sphagnum magellanicum and one of the green species S. fallax (Fig. 2a). The S. magellanicum endmember clearly showed the red reflectance peak around 636 nm, while the green peak of the S. fallax endmember was very broad compared to vascular plants. Further, the Sphagnum mosses did not show a steep red edge with a clear NIR shoulder, showed a higher reflectance peak in the near infrared, and had a lower reflectance in the covered shortwave infrared (1400-2400 nm) than green vascular plants (e.g. Molinia green). Strong water related features were seen at 970 nm, 1200 nm, 1450 nm. Because the measurements were performed in late spring, the dead fraction in graminoid-dominated plots was still very high. We therefore selected one spectrum for pure Molinia green 15 biomass and one for Molinia litter. Additionally we had two spectra for Eriophorum, combining standing dead and living biomass (Fig. 2b). Except for the green Molinia, all graminoid endmembers contain dead biomass expressed by the relatively weak absorption around 680 nm which is usually more pronounced by chlorophyll a absorption, the low near-infrared reflectance, and the almost missing water absorption features. 20 For the ericoid shrubs we obtained two representative spectra for all ericoids (mixed vegetation of Erica, Vaccinium and Empetrum) and two pure spectra from relatively high Erica vegetation (Fig. 2c). The pure ericoid endmember spectra show a very low reflectance in the near infrared compared to Sphagnum and graminoids. The single species reflectance spectra were grouped by their plant functional type,  (Table 1). The set of provided endmembers (3 for mosses, 4 for graminoids, 3 for shrubs, one for shade) leads to a total of 36 potential endmember models (3⇥4⇥3=36) which can be assigned to a plot level spectrum. We thus assumed that in each plot one species is representing the fractional cover of the entire plant functional type. This may lead to higher RMSE values 5 of the model fit when the reflectance of the species within one PFT varies significantly. The shrubs were casting most of the shadow due to their height and loose canopy, further the ericoids showed the lowest HCRF values, thus being closest to the zero reflectance spectrum of the shade. We therefore added the shade abundance results to the shrub abundances for the comparison with the point intercept methodology data. 10 Three main scenarios are discussed in the results, namely: Scenario 1 including all plot results where an endmember model was assigned. Scenario 2 excluding plot results with a high RMSE. The corresponding threshold is set as the mean RMSE of all plots plus 2 times the standard deviation.
Scenario 3 results are inferred from a MESMA run restricted to the wavelength 400-15 1000 nm. This reveals whether the visible spectral range contains enough information compared to the full wavelength range (400-2400 nm) of Scenario 1.

MESMA results
3.1.1 Scenario 1 results (plots with assigned endmember model) 20 Out of 36 potential endmember models, 13 were identified by the MESMA algorithm to fit the 25 observed plot HCRF data. No specific pattern in the model selection could be found, apart from the fact that the green Molinia caerulea and the Vaccinium oxycoccus were only included in 4 and 3 cases, respectively. For two plot spectra, no appropriate model could be assigned within the given constraints (transect A, plot 1 and 4), leading BGD 5,2008 Fcover estimation of plant functional types in peatlands to a final number of 23 plots with successfully assigned abundances. The reason for A1 (Sphagnum cover of 0.68 by point intercept method) was that the observed plot level spectrum had higher values than the available Sphagnum fallax endmember, while the S. magellanicum endmember showed a higher reflectance but did not fit the shape in the green and red spectral bands due to the red reflectance peak. Generally, 5 the obtained fits were very good, with a low RMSE ranging from 0.0027 to 0.0118, exceeding 0.01 in two cases only. The mean RMSE over all assigned plots (23) was 0.0056.

Scenario 2 results (plots with assigned endmember model and high RMSE)
Based on the mean (0.0056) and standard deviation (0.0027) of the RMSE of all plots, 10 the RMSE threshold (mean RMSE of all plots plus 2 times the standard deviation) was set to 0.011, leading to the exclusion of two plots. The mean RMSE over the remaining 21 plots was 0.0050.

Scenario 3 results (MESMA run based on visible spectral range)
Scenario 3 revealed how important the infrared spectral range is for inferring the frac- 15 tions of the three plant functional types within the test site. Only for a single plot spectrum no corresponding endmember model was assigned, whereas the RMSE was higher than 0.01 for three plots. The mean RMSE over all assigned plots (24) was 0.0045, thus slightly lower than for Scenario 1 and 2.

Scenario 1 results (plots with assigned endmember model)
For Sphagnum, the fractional cover estimate using field spectroscopy was in agreement with the point intercept method (R 2 =0.76) and close to the 1:1 line, yielding a high Sphagnum cover under open vascular plant canopy and a low Sphagnum cover under 1303 dense vascular plant canopy. The absolute Sphagnum cover in all plots was generally >90% (see Sect. 2.1). The above result shows that the fractional cover as inferred from the MESMA procedure corresponds to the first-hit data. As vascular plant cover increases, their signal obscures the Sphagnum signal. For graminoids, the agreement between both methods was less strong (R 2 =0.54), 5 but still highly significant (Table 2), with the field spectroscopy method giving a higher estimate at low covers and a lower estimate at dense cover. For the ericoid shrubs, the agreement between the two methods was weak (R 2 =0.19) when the shade fraction was not assigned to the shrub fraction. When summing up the shrub and the shade fraction, the agreement between point intercept 10 and field spectroscopy inferred results was much higher (R 2 =0.43).
The field spectroscopy method generally yielded a higher estimate at low cover and a lower estimate at dense cover, although the deviation from the 1:1 line was less than for the graminoids.

15
The exclusion of plot results showing a model fit significantly deviating from the mean (i.e. RMSE>0.011) lead to an improved fit with the point intercept methodology (Table 2). The correlations improved to R 2 =0.81 for Sphagnum, R 2 =0.57 for graminoids, and R 2 =0.54 for shrubs (Fig. 3). Further, the fit equations move closer to a 1:1 relationship for all three plant functional types (Table 2). When summing up the shrub 20 and graminoid fractions to a vascular plant fraction, the point intercept and MESMA inferred results correlate with a similar fit (R 2 =0.82). The small di◆erence between the Sphagnum and the vascular plant fit quality can be explained by the RMSE. It should be noted that the exclusion of the two data points relies on a statistical quality measure of the fit of the MESMA procedure and is not based on any information 25 inferred from the point intercept methodology. This indicates that results based on an RMSE significantly deviating from the mean RMSE can be identified as outliers due to methodological problems (i.e. plot spectrum cannot be reasonably explained by 1304 infrared and shortwave infrared) are to distinguish between plant functional types in peatlands. The mosses show distinct water absorption features in the longer wavelength range due to their high water holding capacity, and pigmentation alone in the visible range is not sucient to distinguish between the three plant functional types.

10
The MESMA fractional cover based on Scenario 2, as well as the fractional cover inferred from point intercept data were finally correlated to the biomass. For Sphagnum there was no relationship (R 2 0.03) between fractional cover and biomass, irrespective of the method used (Fig. 4 top). This result is expected as the Spagnum shieldedr by the graminoid and ericoid leaves were not detected by both 15 methods, while it contributes considerably to the biomass.
For graminoids however, biomass was surprisingly well related (R 2 =0.57) to the fractional cover estimate derived from field spectroscopy, but not to the estimate derived from the point intercept method (R 2 =0.26). The graminoid fractional cover as inferred by MESMA varies between 0.02 and 0.7, with an average value of 0.28. This interme-20 diate vegetation density allows getting a rough approximation of the biomass through the fractional cover as inferred from MESMA. For the ericoid shrubs, there was an opposite response. Here biomass was better predicted by the fractional cover estimate derived from the point intercept method (R 2 =0.44), than by the estimate derived from field spectroscopy (R 2 =0.14).

BGD
site. The low RMSE values obtained show that generally the selected endmembers and corresponding abundances fit well with the plot level spectrum. This result is certainly influenced by the fact that all spectra were obtained with the same instrument and thus same spatial and spectral resolution and under comparable solar angles and weather conditions. Unlike earlier studies, we present reflectance spectra of all plant functional 10 types of peatlands under natural illumination conditions as mixture and in pure species coverage.
Compared to the endmembers used by Sonnentag et al. (2007), the Sphagnum versus vascular plant endmembers of this study show di◆erent reflectance behaviour. In our dataset, the Sphagnum reflectance is generally higher than most vascular plant spectra in the spectral range of the Landsat TM band 4 (760-900 nm), and always lower in band 5 (1550-1750 nm) and 7 (2080-2350 nm), while it is the opposite in the Sonnentag et al. (2007) study. This opposite reflectance behaviour may partly stem from the di◆erent species and structure of the canopy, because the study site of Sonnentag et al. (2007) included also tree species which were not present in our case. 20 On the other hand it should be noted that they inferred the moss endmembers from field spectrometer data while the vascular plant endmember is based on the TM image. Di◆erent scattering and adjacency e◆ects present in field and satellite measurements, even when compensated properly for atmospheric e◆ects and illumination geometries, may result in biases which do not relate to inherent properties of the plant functional 25 types. As the endmembers are crucial for the MESMA procedure, it is recommended to carefully investigate their consistency and reliability for the corresponding test site. 1306 The wavelength test shows that with visible and NIR wavelengths only, the MESMA algorithm finds suitable solutions, expressed by the relatively low RMSE. However, the solutions do not correlate well with the fractional cover as inferred from the point intercept methodology. This is an important finding for upscaling purposes, as many of the operationally operated and higher resolution satellites (e.g. Landsat ETM+ and TM) 5 do not cover the spectral range of the well-defined water absorption bands shown by the Sphagnum endmember. Nevertheless a combination of the green, NIR and SWIR band might still lead to satisfactory results. However, this means that the absolute reflectance values of the endmembers have to be determined very carefully as the water absorption feature cannot contribute to identify the properly matching endmembers.
This shows the need to perform more detailed analysis when reducing the spectral resolution for the MESMA procedure.
This study demonstrated that based on spectrometer data, Sphagnum fractional cover can successfully be distinguished from vascular plant fractional cover. However, the partitioning of the vascular plant fractional cover into a shrub and graminoid 15 fraction was less successful and needs further work with respect to endmember measurements. This includes the measurement of shade endmembers to replace the photometric shade which is wavelength independent, as well as a careful experimental setup in terms of solar angle and corresponding reflectance anisotropy e◆ects.
The above results lead to the following conclusions for the application to assess and 20 predict the carbon cycle in peatlands: 1. The presented methodology can be used to map open Sphagnum patches and consequently apply a water-based index on the pure Sphagnum spectra as a proxy for the near-surface hydrological conditions (methodology as presented in Harris et al., 2006). Reduction of the near-surface water content can lead to a reduction in rates of 25 Sphagnum growth and therefore, carbon fixation. Harris et al. (2006) mentioned that their Sphagnum mapping procedure, based on a mixed tuned match filtering approach, was largely an iterative process and may increase user bias and reduce accurate repeatability due to the manual adjustment of the threshold. After the careful selection of BGD the endmembers, the MESMA approach applied in this study proved to be very stable, and revealed reasonable results even without application of a threshold. Applying a standard threshold to the product quality output (mean RMSE plus 2 standard deviations) resulted in an R 2 =0.81 for Sphagnum and R 2 =0.82 for vascular plants (sum of shrub plus ericoid fractional cover) when compared to point intercept methodology 5 data.
2. The partitioning of the vascular plant cover into shrub and graminoid fraction is still limited. Although this is not such a problem for boreal peatlands, where litter decomposition rates are less defined along graminoid-shrub lines, results for more temperate peatlands would su◆er (Dorrepaal, 2007). We see the potential to improve the pure endmembers of the shrubs as the mixed shrub endmember we used showed the influence of the Sphagnum reflectance. Further, the shade of the graminoids and the ericoids should be measured separately, replacing the photometric shade by a wavelength dependent shade. Plant functional type specific shade would also partially compensate for the non-linear scattering contributions of the vegetation canopy which 15 is currently neglected by the MESMA approach. Once the partitioning into shrubs and graminoids has improved, the fractions of the 3 plant functional types can directly be used to assess the carbon sequestration potential of the observed peatland. The observation of the vegetation composition in combination with wetness indicators will lead to a better estimation of the net ecosystem carbon dioxide exchange as suggested 20 by Strack et al. (2006), who emphasize the importance to combine ecological and hydrological indicators.
3. The spatially continuous mapping of the Sphagnum versus vascular plant fractional cover over time can be used to monitor potential vegetation succession as induced by changes in hydrology or nutrient availability (e.g. nitrogen deposition). Table 1. Endmember provided for the MESMA procedure. From each endmember group (i.e. mosses, graminoids, and shrubs), one endmember is selected by MESMA leading to the best plot level spectrum fit. The forth endmember is a photometric shade (zero reflectance in all spectral bands).   Transect A B C D E Regression 95 % conf limits Fig. 4. Relationship for three plant functional types between fractional cover estimates and biomass for point intercept method (left) and field spectroscopy (right). Biomass refers to aboveground dry biomass for graminoids and shrubs and upper 7 cm for Sphagnum.