Do we miss the hot spots? ? The use of very high resolution aerial photographs to quantify carbon fluxes in peatlands

Accurate determination of carbon balances in heterogeneous ecosystems often re-quires the extrapolation of point based measurements. The ground resolution (pixel size) of the extrapolation base, e.g. a land-cover map, might thus inﬂuence the calculated carbon balance, in particular if biogeochemical hot spots are small in size. In 5 this paper, we test the e ﬀ ects of varying ground resolution on the calculated carbon balance of a boreal peatland consisting of hummocks (dry), lawns (intermediate) and ﬂarks (wet surfaces). The generalizations in lower resolution imagery led to biased area estimates for individual micro-site types. While areas of lawns and hummocks were stable below a threshold resolution of ∼ 60 cm, the maximum of the ﬂark area was 10 located at resolutions below 25 cm and was then decreasing with coarsening resolution. Using a resolution of 100 cm instead of 6 cm led to an overestimation of total CO 2 uptake of the studied peatland area (approximately 14 600 m 2 ) of ∼ 6% and an underestimation of total CH 4 emission of ∼ 11%. To accurately determine the surface area of scattered and small-sized micro-site types in heterogeneous ecosystems (e.g. ﬂarks in 15 peatlands), a minimum ground resolution appears necessary. In our case this leads to a recommended resolution of 25 cm, which can be derived by conventional airborne imagery. The usage of high resolution imagery from commercial satellites, e.g. Quickbird, however, is likely to underestimate the surface area of biogeochemical hot spots. It is important to note that the observed resolution e ﬀ ect on the carbon balance estimates 20 can be much stronger for other ecosystems than for the investigated peatland where the relative hot spot area of the ﬂarks is very small and their hot spot characteristics with respect to CH 4 and CO

this paper, we test the effects of varying ground resolution on the calculated carbon balance of a boreal peatland consisting of hummocks (dry), lawns (intermediate) and flarks (wet surfaces). The generalizations in lower resolution imagery led to biased area estimates for individual micro-site types. While areas of lawns and hummocks were stable below a threshold resolution of ∼60 cm, the maximum of the flark area was 10 located at resolutions below 25 cm and was then decreasing with coarsening resolution. Using a resolution of 100 cm instead of 6 cm led to an overestimation of total CO 2 uptake of the studied peatland area (approximately 14 600 m 2 ) of ∼6% and an underestimation of total CH 4 emission of ∼11%. To accurately determine the surface area of scattered and small-sized micro-site types in heterogeneous ecosystems (e.g. flarks in tem are selected, which cover the spatial heterogeneity of the study site. There, fluxes are measured, and the modeled seasonal gas exchange fluxes from these plots are extrapolated to larger areas or the whole ecosystem. Extrapolation is usually done based on the spatial representation of each measured micro-site within the ecosystem: a modeled flux of a particular representative micro-site is usually multiplied by the 5 area that particular micro-site type occupies (Schimel and Potter, 1995). The exact spatial distribution of micro-sites is in particular important, if micro-site size is small and the ecosystem surface strongly heterogeneous, e.g. in many peatland ecosystems. Spatial information on micro-site distribution can be obtained by rough estimation, vegetation mapping in a smaller area e.g. Riutta et al. (2007), along transects e.g. Alm et al. (1997) and Laine et al. (2006), or with a land-cover map of the complete area under study e.g. Bubier et al. (2005). While this last approach promises the most reliable spatial estimates and thus the most reliable flux extrapolation, it depends entirely on the relationship between the ground resolution of the imagery and the size of the micro-sites. Here, we show that ecosystem trace gas flux estimates, especially for methane, depend significantly on the resolution of the underlying landcover map. We further develop recommendations for a reasonable ratio between size of micro-sites and resolution of the underlying landcover map.

Study site
The peatland "Salmisuo" is located at 62 • 47 ′ N, 30 • 56 ′ E, in Eastern Finland (Fig. 1), 20 and is generally classified as an oligotrophic low-sedge pine fen (Saarnio et al., 1997). Climatic conditions represent the boreal forest climate (Strahler and Strahler, 2005) with a mean annual air temperature of + 2.1 • C and a mean annual precipitation of 667 mm (years: 1971-2000in Finnish Meteorological Institute, 2002

Methods
The calculated carbon balance for this study is based on 1) plot-scale quantification of CO 2 and CH 4 exchange fluxes using closed chambers over 50 days, 2) a hydrological 5 part to estimate the lateral carbon losses by dissolved organic carbon (DOC) and 3) a remote sensing part to map the spatial distribution of micro-sites.

Gas flux measurements
For this study, we analyzed CO 2 and CH 4 emission for the time period 26 July 2005-13 September 2005 (50 days): Fluxes of CO 2 and CH 4 were measured with the closed 10 chamber technique (Alm et al., 2007). Sample plots have been choosen by the three dominant types of micro-sites (flarks, lawns and hummocks). For every micro-site type four replicate sample plots have been used to achieve a representive mean value of the appropriate vegetation type.
CO 2 and CH 4 fluxes were measured once a week. The CO 2 measurements were 15 performed over 24 h. For determination of net ecosystem CO 2 exchange, we employed a vented transparent chamber (60 cm×60 cm×32 cm) with an automatic cooling system which kept the headspace air temperature within approximately 1 • C of the ambient temperature. The dark respiration and CH 4 flux measurements were conducted using vented aluminum chambers. The CO 2 concentrations were measured using a 20 CO 2 /H 2 O infrared gas analyzer (LI-840, Licor, USA). CO 2 readings were taken at 1 s intervals over 180 s. During the CH 4 flux measurements, four headspace samples were taken every 4 min from the chamber in a 16 min time period. CH 4 concentration in the syringes were analyzed one day later with a gas chromatograph (Shimadzu 14-A) equipped with a flame ionisation detector. The gas fluxes were calculated from the concentration increase in the chamber headspace over time applying nonlinear regression for CO 2 (Kutzbach et al., 2007) and linear regression for CH 4 . The seasonal time series of CO 2 and CH 4 exchange fluxes over the investigation period were modeled on a temporal resolution of 0.5 h for CO 2 and 1.0 h for CH 4 using multilinear regression models with photosynthetically active radiation, air temperature, peat tem-5 perature in 5 cm, air pressure, wind speed and water table as predictors of CO 2 fluxes and groundwater table, peat temperature in steps of 5 cm, 10 cm, 20 cm and 50 cm and wind speed as predictors for CH 4 fluxes. Then, the modeled time series of CO 2 and CH 4 fluxes were integrated to derive the total amount of CO 2 and CH 4 exchanged over the investigation period. The flux value for each micro-site type was then calculated as 10 the mean of the four replicates.

Hydrology
Dissolved organic carbon (DOC) export was calculated by multiplying daily surface runoff with average daily DOC mass per volume concentrations ([DOC]); measurements were undertaken at a ditch collecting the peatland outflow.

Remote sensing
The remote sensing task was covered by very high resolution imagery taken from a helium filled dirigible. The dirigible with a volume of 2 m 3 was capable to lift 1 kg of payload and was with his tail fins well equipped to be more stable in the air than a balloon. At the bottom of the dirigible, a camera rig was attached that held the camera 5 in an almost nadir position.
To obtain the imagery, we utilized a 7 megapixel point and shoot camera (Canon Powershot G6) combined with a 2 gigabyte storage medium. This setting provided us with the ability to obtain 100 raw data images (*.crw) per flight session with a resolution of 3072×2304 pixels and a shooting frequency of one image per minute. The restriction 10 of 100 images was given by the software of the camera. The ground resolution of these imagery depends very much on the flying height of the platform (e.g. ∼5 cm at a flying height of 130 m above the ground).
For further processing the imagery was georectified using a grid of ground control points (GCPs). The grid had a cellwidth of about 50 m, and the position of every GCP 15 was measured with a differential global positioning system. The average horizontal accuracy of these measurements was 35 cm.
In order to get a reasonable amount of GCPs for georectification and at the same time a very high ground resolution, a flying height of ∼150 m above the ground was chosen, offering a ground resolution of about 6 cm and a minimum of 6 GCPs in every 20 image.
To simulate different flying heights of the dirigible, we coarsened the ground resolution from 6 cm to 10 cm and further in steps of 5 cm up to a resolution of 100 cm. By coarsening the resolution up to 100 cm we cover the range from very high resolution airborne imagery to very high resolution commercial satellite imagery (e.g. QuickBird Resource Mapping, 2006).
The georectified imagery was classified in the next step, defining regions to represent the micro-site types and using a supervised classification with the maximum likelihood algorithm in ER Mapper 7.1. The resulting land cover map (Fig. 2) was vectorized, using the Raster-To-Polygon function in ArcGIS of ESRI to proceed to the statistical analysis (ESRI, 2004).
Furthermore we calculated total area and average size of each micro-site type for each resolution (Table 1, Fig. 3).
To locate discontinuities in the data we conducted a moving split window analysis (Johnston et al., 1992). Using the moving split window a changing of the observed 10 attribute is indicated by maximum values in the graphs. A four-sample window width was applied to find possible thresholds while coarsening the ground resolution.

Results
Highest obtained ground resolution was 6 cm and subsequent coarsening resulted in 20 area estimates (Fig. 3) for each micro-site type. Flark area was stable (∼200- Do we miss the hot spots? T. Becker et al. total area, estimates of lawns and hummocks behave nearly as mirror images of each other (Fig. 3). This effect is probably also related to the resampling and classification method. The amount of single objects in the classes of lawns and hummocks and their close spatial relationship is causing a give-and-take between these two classes at their common border. Hence the spatial representation of the two major classes depend on 5 each other and a changing of much smaller classes has no reasonable effect. Seasonal gas fluxes differed between micro-site types (Tab. 2) with flarks emitting the most CH 4 per area and hummocks taking up most of the CO 2 per area. Seasonal DOC export was calculated as 0.09 ± 0.02 g C/m 2 , representing only 0.44 % of the seasonal carbon balance. Taken together, the generalizations in lower resolution imagery 10 lead to biased area estimates for the individual micro-site types (Fig. 3), and thus at a resolution of 100 cm to an overestimation of total CO 2 uptake of ∼5.5 % (Fig. 5a) and an underestimation of total CH 4 emission of ∼11 % (Fig. 5b).
The accuracy of gas flux estimations in this approach is highly related to the ground resolution of the imagery used for the classification. Due to stronger generalization at 15 a smaller scale the loss of small objects is increasing by coarsening the pixel size.
To identify possible thresholds for the detection of large changes in the calculated area during the coarsening process and thus reasonable object sizes at the particular resolution (Fig. 4), we used the moving split window analysis (MSWA) e.g. Johnston et al. (1992). For every micro-site the lowest possible detection threshold, indicated by 20 the peak, is located at a ground resolution of 25 cm. The next possible threshold for every micro-site is at a ground resolution of 60 cm.
Based on the results of the MSWA (Fig. 4) we calculated the mean object size for every micro-site type at ground resolutions of 25 cm and 60 cm (Table 3) to propose ratios for each micro-site type for the identification of objects in similar heterogeneous 25 environments like the observed peatland (Table 4). We have choosen the mean object size to minimize influence of a dominating number of small objects at all resolutions.

Discussion
The underestimation of methane fluxes at lower resolution, caused by the underestimated area of flarks and lawns, leads to a conservative approximation of the methane fluxes in the particular area. Using a ground resolution of 100 cm the total carbon budget is underestimated by ∼1.18 g/m 2 in the sample area, compared to the highest 5 resolution of 6 cm. The total amount of effective greenhouse gases would be underestimated by ∼9.3 % between a ground resolution of 6 cm and 100 cm. Using land-cover maps with even lower resolutions (Takeuchi et al., 2003), would very likely increase this effect.
As shown in Fig. 3 the total area of individual micro-site types, depending entirely 10 on the size and number of the associated polygones, is altered at a changing resolution. On the one hand this is caused by the generalization of details from high to lower resolution data (Jensen, 2000). On the other hand it is more difficult to identify smaller objects at lower resolutions, leading to errors during the classification process (Markham and Townshend, 1981). It is also possible, that the classification result is 15 influenced by the data distribution, considering that the maximum likelihood algorithm assumes a normal distribution of the band data (Leica Geosystems GIS and Mapping, 2003). The result of the MSWA indicates possible thresholds for the resolution of the imagery (Fig. 4). To achieve reasonable classification results in a peatland like Salmisuo 20 a ground resolution of 25 cm is recommended to analyze small micro-sites (e.g. flarks).
To analyze micro-sites as lawns and hummocks a ground resolution of 60 cm seems to be adequate. Both thresholds show that very high satellite imagery still tends to misjudge the distribution of the micro-sites (plant communities) in small patterned peatlands. 5,2008 Do we miss the hot spots?

BGD
T. Becker et al.

Conclusions
We show that based on differing ground resolution of the land-cover map, substantially different areas for individual micro-site types are calculated. This influences the calculation of the carbon balance, since gas fluxes between the ecosystem and the atmosphere are measured at representative spots of each micro-site type and then 5 multiplied by the micro-site area. In particular small micro-sites, which are often biogeochemical hot-spots, (e.g. wet areas emitting CH 4 ), tend to be affected. In our field site, a ground resolution of 25 cm seems to be necessary for the detection of these biogeochemical hot-spots with respect to CH 4 emission. A resolution of 60 cm seems sufficient for a representative detection of larger micro-site types as well as with respect to CO 2 fluxes for all micro-sites types. To successfully detect small micro-site types (e.g. flarks), we thus recommend a ratio of 1:2 of mean object size to image ground resolution and for larger micro-site types (e.g. lawns and hummocks) a ratio of Interactive Discussion variation in CH4 emissions and production and oxidation potentials at microsites on an oligotrophic pine fen, Oecologia, 110, 414-422, doi:10.1007/s004420050176, 1997. 1099