Publication: Gallaun, H.; Zanchi, G.; Nabuurs, G.-J.; Hengeveld, G.; Schardt, M. & Verkerk, P. J., 2010. EU-wide maps of growing stock and above-ground biomass in forests based on remote sensing and field measurements Forest Ecology and Management, 2010, 260 Issue 3, 252-261

EU-wide maps of growing stock and above-ground biomass in forests based on remote sensing and field measurements. Heinz Gallaun

a,*, Giuliana Zanchi

a, Gert-Jan Nabuurs

b, Geerten Hengeveld

b, Mathias Schardt

a, Pieter J. Verkerk

c

a Joanneum Research, Steyrergasse 17, 8010 Graz, Austria

b Alterra b.v., PO Box 47, 6700 AA Wageningen, Netherlands

c European Forest Institute, Torikatu 34, 80100 Joensuu, Finland

Abstract The overall objective of this study was to combine national forest inventory data and remotely sensed data to produce pan-European maps on growing stock and above-ground woody biomass for the two species groups “broadleaves” and “conifers”. An automatic up-scaling approach making use of satellite remote sensing data and field measurement data was applied for EU-wide mapping of growing stock and above-ground biomass in forests. The approach is based on sampling and allows the direct combination of data with different measurement units such as forest inventory plot data and satellite remote sensing data. For the classification, data from the Moderate Resolution Imaging Spectroradiometer (MODIS) was used. Comprehensive field measurement data from national forest inventories for 98979 locations from 16 countries were used for which tree species and growing stock estimates were available. The classification results were evaluated by comparison with regional estimates derived independently from the classification from national forest inventories. The validation at the regional level shows a high correlation between the classification results and the field based estimates with correlation coefficient r=0.96 for coniferous, r=0.94 for broadleaved and r=0.97 for total growing stock per hectare. The mean absolute error of the estimations is 25 m³/ha for coniferous, 20 m³/ha for broadleaved and 25 m³/ha for total growing stock per hectare. Biomass conversion and expansion factors were applied to convert the growing stock classification results to carbon stock in above-ground biomass. As results of the classification, coniferous and broadleaved growing stock as well as carbon stock of the above-ground biomass is mapped on a wall-to-wall basis with a spatial resolution of 500m by 500m per grid cell. The mapped area is 5 million km², of which 2 million km² are forests, and covers the whole European Union, the EFTA countries, the Balkans, Belarus, the Ukraine, Moldova, Armenia, Azerbaijan, Georgia and Turkey. Keywords: Biomass, growing stock, carbon, remote sensing, forest inventory, MODIS 1. Introduction The overall objective of this study was to develop EU-wide maps on growing stock and above-ground woody biomass for the two species groups “broadleaves” and “conifers”, by combining national forest inventory data and remotely sensed data. Forests cover 44% of land area in Europe and are therefore of major ecological, political and economic importance. Whereas accurate estimations of forest parameters are available at the local level, especially for growing stock, European-wide assessments with high spatial resolution are not common. The recent development of strategies for renewable energy production, the progressive globalisation of the markets, including the market of wood products, and carbon stock estimation for climate change modelling, have increased the demand for spatial explicit information on forest resources. Two important parameters that well describe the spatial distribution of the forest resources are maps on growing stock and the biomass stock. For the estimation of growing stock and above-ground biomass, different approaches have been developed, based on field measurement, modelling, and remote sensing. A good overview is given by Lu (2006). Field measurements are used for accurate estimation of growing stock and woody biomass from the allometric relationship with measured diameter at breast height and tree height. Mostly, only small areas are measured, as the field measurements are very labour- and cost-intensive. The typical size of plots measured in national forest inventories is between 200m² to 500m². In general, remote sensing approaches for estimating forest parameters are based on reference data such as e.g. plot data, for which the forest parameters are measured or estimated. The reference data are then combined with the remote sensing data to derive statistical parameters that are used for the estimation of the forest parameters for the whole area. However, this direct comparison requires that the reference data and the picture elements of the remote sensing data cover the same area on the ground. The remotely sensed imagery currently available for regional to global vegetation studies are often at a much lower resolution than reference data (i.e. 25 ha minimum pixel size for MODIS satellite data and 200m² to 500 m² for field plot measurements). Therefore, and because different land cover categories arc covered by single picture

elements (mixed pixels), the integration of sample data and remote sensing-derived variables is still difficult (Lu, 2006; Anaya et al., 2009). Blackard et al. (2008) mapped forest biomass in the United States using nationwide forest inventory plot data and moderate resolution information by modelling. The geospatial predictor variables included: MODIS-derived image composites and percent tree cover, land cover proportions, topographic variables, monthly and annual climate parameters and other ancillary variables. The estimations were performed separately for 65 strata. Only for 75% of the strata, relative errors were below 1.0, indicating gains in the modelling process. Päivinen et al. (2009) generated pan-European growing stock maps based on the forest proportion map of Häme et al. (2001) and regional statistics derived from national forest inventories. The forest proportion map was derived from NOAA-AVHRR satellite data which has a spatial resolution of 1.1 km by 1.1 km and CORINE 1990 land cover data. The regional to national forest inventory estimates were distributed according to the forest cover proportion. They transferred the mean growing stock volume for the region to the pixels by multiplying with the share of forest area within a pixel. They analysed rounding errors but do not report on an independent accuracy assessment. Stümer et al. (this issue, 2009) applied a self-organizing map algorithm for the assessment of carbon stock in forests for Thuringia in Germany. The approach is based on national forest inventory data on plot level and remote sensing data from Landsat. They compared this approach with a k-nearest neighbour approach as well as with results of the present study. The results were compared with estimates derived from national forest inventory plots and all three approaches show good agreement at the regional level (Stümer et al., this issue, 2009). Muukkonen and Heiskanen (2007) combined ASTER and MODIS satellite data to estimate growing stock and above-ground biomass in southern Finland. They used regression models based on stand-wise forest inventory data as intermediate step between the field data and the MODIS data. The estimates obtained were close to the district-level mean values provided by the Finnish National Forest Inventory with relative root mean square error of 9.9%. Tomppo et al. (2002) combined Landsat Thematic Mapper and IRS-1C Wide Field Sensor (WiFS) data to estimate tree stem volume and above-ground biomass in Finland and Sweden. The Landsat Thematic Mapper data were used as an intermediate step between field data and WiFS data. The nonparametric k–nearest neighbour method was used to analyse relationships between Landsat Thematic Mapper and field data, and nonlinear regression analysis was used to develop models for predicting volume and biomass for WiFS pixels. Validation was performed by comparing stem volume at the municipality level with the Finnish National Forest Inventory. Only WIFS pixels which were totally on forestry land were used. The reported relative differences range from -24.9% to 28.5%. Regarding growing stock, European-wide mapping was performed by Päivinen et al. (2009) as described above. A main difference of the current study is that we use forest inventory plot data as reference data, whereas Päivinen et al. (2009) use regional to national data. Furthermore, our maps are based on MODIS data with a spatial resolution of 25 ha, whereas the work of Päivinen et al. (2009) is based on NOAA-AVHRR data with a spatial resolution of 120 ha. The area covered by the map of Päivinen et al. (2009) includes the European part of Russia, which is not covered by our work, whereas we cover Turkey, Armenia, Azerbaijan and Georgia, which is not covered by Päivinen et al. (2009). EU-wide mapping of above ground forest biomass by direct combination of remote sensing and forest inventory plot data has not been published up to now. The main reason for this is that appropriate ground measurement plot data are mostly only available at the regional to national level, whereas for the current study, plot-level data from 16 national forest inventories, distributed over the main forest ecosystem regions of Europe, were provided by national forest inventory institutes. The work for this study was performed within the research project “CarboEurope-IP” which aimed to understand and quantify the terrestrial carbon balance of Europe. The mapped area is 5 million km² of which 2 million km² are forests. 2. Material and methods For the production of the pan-European maps, comprehensive field measurement data from national forest inventories were used (Nabuurs et al., this issue, 2009), which were provided by national forest inventory institutes. An automatic up-scaling approach was applied to allow the direct use of the field inventory plots for the classification of the MODIS data. The estimates refer to the year 2000 for which the MODIS data, the calibrated inventory data at the regional level, and the mean biomass expansion factors at the regional level were available. The estimations were performed on a wall-to-wall basis for the whole area with a spatial resolution of 500 meters by 500 meters. The general workflow is outlined in Figure 1.

Fig. 1: General workflow 2.1 Field data from national forest inventories The data were selected from a collection of National Forest Inventory (NFI) plot data (Nabuurs et al., this issue, 2009). From this database, 98979 locations in 16 countries were selected where data on tree species and growing stock were provided by national forest inventory institutes. The data were classified and aggregated to the classes ”broadleaved” and ”conifer”. For each location the volumes were reported in m

3/ha. All locations were geo-

referenced. 2.2 Remote sensing and ancillary data We used four datasets from remote sensing: MODIS satellite data, meteorological data, CORINE Land Cover 2000 data and MODIS Vegetation Continuous Fields. We only considered data that are freely available for map production. Repeated application of this approach therefore allows a cost effective monitoring of the forest resources over time. MODIS satellite data The Moderate Resolution Imaging Spectroradiometer (MODIS) was launched on board Terra satellite in 1999 followed by Aqua satellite in 2002. It is a 36-channel visible to thermal infrared sensor with a scene width of 2230 km and a temporal resolution (repeated coverage) of 1-2 days. For the classification, the product “Nadir BRDF-Adjusted Reflectance 16-Day L3 Global 500m” (MCD43A4) was used. This product is based on multi-date, atmospherically corrected, cloud-cleared input data measured over 16-day periods. For the classification, the six spectral bands green (0.545–0.565 µm), red (0.620–0.670 µm), near-infrared (0.841–0.876 µm) and middle-infrared (1.628–1.654 µm and 2.105–2.155 µm) were used. Band 3 in the blue spectral region (0.459–0.479 µm) was not used because calibration uncertainty can lead to large errors in this band. For the classification, data from 9

th June to 24

th June 2000 (days of year 161 to 176) were used. For small areas that are not properly covered by

this dataset, e.g. cloud covered region in Scotland, data from July were additionally incorporated. The so-called Collection 5 was used as it reflects the recent (2009) improvements to the MODIS science algorithms and has a spatial resolution of 500 m rather than 1 km as in previous collections. Meteorological data In addition to the MODIS reflectance data, meteorological data from Hijmans et al. (2005) were used as geospatial predictor layers. Hijmans et al. (2005) interpolated climate surfaces for global land areas at a spatial resolution of

30 arc seconds, often referred to as 1-km spatial resolution. The mean temperature of the warmest quarter and the precipitation of the warmest quarter were incorporated in the classification. CORINE Land Cover 2000 CORINE Land Cover 2000 (CLC2000) data were used for stratification for the unsupervised clustering approach and as a basis to generate the forest/non-forest map. CORINE Land Cover 2000 data is an update for the reference year 2000 of the first CORINE Land Cover database which was finalised in the early 1990s as part of the European Commission programme to coordinate information on the environment (CORINE). It provides consistent information on land cover and land cover changes during the past decade across most European countries. MODIS Vegetation Continuous Fields The Vegetation Continuous Fields (VCF) data contains proportional estimates for vegetative cover types: woody vegetation, herbaceous vegetation, and bare ground. The product is derived from all seven bands of the Moderate Resolution Imaging Spectroradiometer sensor (Hansen et al., 2003). For the current study, the vegetative cover of woody vegetation was taken as input for the generation of the forest/ non-forest map. 2.3 Forest/ non-forest map An EU-wide forest/ non-forest map was generated, consistent with the Temperate and Boreal Forest Resource Assessment –TBFRA 2000 (UNECE-FAO, 2000) at the national level. For areas where CORINE land cover data are available, the CORINE dataset was aggregated from the original 100 meters to 500 meters spatial resolution. Firstly, the number of forest pixels within each 5 by 5 pixel aggregation unit was calculated. Secondly, a threshold with the minimum number of forested pixels within the aggregation units was determined for each country. This threshold was selected accordingly, to generate a forest map in agreement with the total forest area given by TBFRA 2000 at the national level. For areas not covered by CORINE data, a similar approach was applied with Vegetation Continuous Fields data. The area covered with woody vegetation in the VCF data is given in percent. A percentage threshold of the minimum area covered by woody vegetation was defined for each country in order to produce the forest/ non-forest map. Also in this case, the total forest area is coherent with TBFRA 2000 data. The result is an EU-wide forest/ non-forest map in agreement with the Temperate and Boreal Forest Resource Assessment –TBFRA 2000 on total forest areas at the national level. 2.4 Growing stock estimation Growing stock or stem volume is defined in this study as volume of standing trees, above stump measured over bark. The data sources described above were used for the EU-wide estimation of the growing stock of the two species groups “conifers” and “broadleaves”. The comprehensive field measurement plot data were divided in two parts: one part was taken for training and the other part was used for an independent accuracy assessment. An automatic up-scaling approach was applied to allow the up-scaling of plot measurements and remote sensing data of a different spatial resolution, e.g. up-scaling of forest inventory plot data with typical measurement area below 1 ha and MODIS satellite data with a pixel size of 25 ha. As opposed to conventional classification methods such as k-NN classification (Tomppo et al., 2002) which relate the plot measurements directly to remote sensing data, this approach has the main advantage that it is not sensitive to scale mismatch between the area covered by the individual plots and the area covered by each pixel. The main processing steps of the automatic up-scaling approach are as follows:

1. Clustering is performed which separates the remote sensing data into classes such that the between-class variance of the specified number of classes is maximized. The determination of the clusters is not performed for the whole area, but separately for selected land cover types, based on ancillary data.

2. For each class, fractional cover maps or posterior probability of membership are calculated for each pixel.

3. For each class, a training set consisting of “pure” pixels which show a high posterior probability of membership to the respective class is selected.

4. The sampling plot or point measurements located within each training set are used to estimate the respective parameter, e.g. mean growing stock for each class.

5. The classification of the whole area is then performed by weighting the class mean values with the fractional cover maps or posterior probability of membership of the respective class.

The detailed processing steps as applied for the generation of the EU-wide mapping of growing stock are described in the following paragraphs (compare also Figure 1, General workflow). Step 1 – Clustering For clustering, the Iterative Self-Organizing Data Analysis Techniques (ISODATA) algorithm was applied. CORINE land cover data was used as ancillary information to avoid selection of clusters which represent composed land cover types (e.g. cluster which represent mixed agriculture / forest land cover). For each selected land cover category, only those MODIS pixels were utilised which show a homogeneous land cover (all CORINE pixels within the respective MODIS pixel belong to the same land cover type). Step 2 – Calculation of fractional cover maps for each category To measure the fractional coverage of the different land cover categories within each pixel, the posterior probability of class membership was calculated (Foody et al., 1992; Foody and Cox, 1994; Häme et al., 2001). The posterior probability was calculated according to Strahler (1980):

The result of this processing step is one layer for each land cover class with posterior probability of membership to the respective class. Step 3 – Automatic selection of training sets For each class, the pixels selected for training are the ones that include NFI sample plot data and that show a high posterior probability of membership to the respective class (Foody and Arora, 1996). Contrary to conventional classification, where homogeneous training areas are often used as reference data, this approach produces one training set for each class composed of dispersed pixels. Step 4 – Calculation of mean growing stock for each class The NFI plots located within each training set are used to calculate mean growing stock for each class, separately for the species groups “conifers” and “broadleaves”. Since the mean values are determined by sampling, the method is not sensitive to mismatch in spatial scale between the plot measurements and the pixel size. This is the main advantage of this approach compared to conventional classification methods. Step 5 - Classification of the whole area The classification of the growing stock is performed by weighting the class mean values with the posterior probability of membership of the respective class (Foody et al., 1992; Foody and Arora, 1996; Häme et al., 2001). The classification procedure was performed separately for the three strata: Northern, Central and Southern Europe. The result of these processing steps is an estimation of the growing stock for each pixel, separately for the species groups “conifers” and “broadleaves”. Accuracy assessment For assessing the accuracy of the results, 24 regions were selected that included at least 250 field measured plots located within each region. To allow an accuracy assessment independent of the classification, these plots were not used in the previous classification process. For each region, the mean growing stock per hectare forest for conifers and broadleaves was calculated with the field measured plot data. These estimates were compared with the mean growing stock per hectare of forest derived from the classification. Areas at the forest borders were excluded for the calculation of the mean values from the classification, because they are composed of forest and non-forest areas. The results of the comparison are presented in Table 1 and in Figure 2. The accuracy assessment is representative for those regions for which appropriate ground reference plot data were provided by the national forest inventory mapping agencies (Belgium, Croatia, Estonia, Finland, France, Germany, Italy, Lithuania, Netherlands, Norway, Portugal, Slovak Republic, Slovenia, Spain, Sweden).

Fig. 2: Comparison of coniferous, broadleaved and total growing stock estimates based on remote sensing, with estimates based on field measurements for n=24 selected regions. The correlation coefficient R is between 0.94 and 0.97. Table 1: Comparison of remote sensing based growing stock estimates (m³/ha) and field based growing stock estimates for n=24 selected regions.

Total growing stock

Broadleaved growing stock

Coniferous growing stock

Correlation coefficient 0.97 0.94 0.96

RMSEa 32 m³/ha 25 m³/ha 32 m³/ha

MAEb 25 m³/ha 20 m³/ha 25 m³/ha

MAErc 11 % 23 % 17 %

Despite the high correlations, the accuracy assessment shows a slight underestimation of the growing stock in regions characterised by high growing stock volume. This may be caused by saturation effects, as optical sensors such as the utilised MODIS bands measure only the upper layers of vegetation cover, and is in agreement with analysis of others, e.g. Tomppo et al. (2002). As the comparison was performed at the regional level, this assessment shows the accuracy at the regional level, but not at the pixel level. The accuracy assessment was performed prior to the calibration at the regional level, which is described below. Calibration based on regional EFISCEN data: At the regional level, the accuracy of the mean growing stock estimates from field based forest inventories is generally high. The estimates of growing stock from forest inventories are usually included in a 5% range of the mean (Liski et al. 2006). As shown in Table 1, the relative mean absolute error of the remote sensing based growing stock estimate is 11 % at the regional level. The inventory-based regional data from EFISCEN were therefore used to calibrate the remote sensing based estimates in those areas which are covered by EFISCEN data. EFISCEN is an inventory-based model that projects the development of forest resources under different management and wood demand scenarios (Sallnäs 1990; Schelhaas et al., 2007). The model results are projected on a five-year time scale at the regional, national and European level (Eggers et al., 2008; Nabuurs et al., 2007). National forest inventories are used as input data sources to define the state of forests and their future development. In the model, the forest is structured as a matrix in which the forest area is distributed over age and volume classes for each tree species. Growth dynamics are simulated by shifting forest area between matrix cells and the growth is driven by functions based on inventory data or yield tables. The development of forest resources is also influenced by the management regimes (thinnings and fellings) and by the increase or decrease of forest area. Some of the outputs of the model are the projection of forest area, growing stock, increment, harvest level, age class distribution and biomass carbon stock for different tree compartments. The results are aggregated at the regional level covering 27 European countries (EU27, Greece and Malta excluded and Switzerland and Norway included).

Growing stock and above-ground biomass were estimated for the year 2000 for the European Union (excluding Cyprus, Greece and Malta), Norway and Switzerland. For Bulgaria, the Czech Republic, Estonia, Latvia and Lithuania the initial inventory data in the EFISCEN dataset represented the year 2000. For the other countries with forest inventory data, we simulated forest resource development until 2000 based on (i) conventional forest management regimes (Nabuurs et al. 2007), (ii) historical roundwood production converted to overbark volumes (UNECE-FAO, 2000) from coniferous and broadleaved species separately, and (iii) used harvest residues. To capture regional differences in management intensity, we regionalised the harvest volumes for Austria, Finland, France, Germany, Norway, Sweden and Switzerland based on national reports. We scaled the Forest available for wood supply in our database to the values reported by MCPFE (2007) to correct for small deviations in the area. To calibrate the classification results, the EFISCEN data on the growing stock in 2000 were aggregated in the two species groups “broadleaves” and “conifers”. The regional mean values on growing stock derived from the classification results were then adjusted to correspond to the regional EFISCEN data by multiplication. For areas not covered by the EFISCEN model, no calibration was performed. 2.5 Carbon stock of the above-ground woody biomass Biomass conversion and expansion factors (BCEFs) were applied to convert the growing stock results from remote sensing to the carbon stock in above-ground woody biomass. Mean BCEFs were calculated for conifers and broadleaves at the regional level based on EFISCEN results. In EFISCEN, age- and species-dependent BCEFs are applied to convert the growing stock to biomass over bark for the different tree compartments (stem, branches, foliage, coarse roots and fine roots). The factors are calculated on the basis of species-specific allometric equations and yield tables (Vilén et al., 2005; Mokany et al., 2006; Gasparini et al., 2006). For this study, aggregated BCEFs for broadleaves and conifers were calculated by dividing the mean above-ground woody biomass (stem and branches) at the regional level for the mean growing stock in the same region. In the countries not covered by EFISCEN projection, the expansion factors were calculated in a similar way based on the data published in the Temperate and Boreal Forest Resource Assessment – TBFRA 2000 (UNECE-FAO, 2000) or by applying the aggregated EFISCEN BCEFs from neighbouring countries where no data were available. The woody biomass was converted to carbon stock (tC ha

-1) by applying a constant carbon fraction of 50% (IPCC,

2003). 3. Results and discussion 3.1. Growing stock Maps of retrieved coniferous, broadleaved and total growing stock per hectare are presented in Figures 3, 4 and 5. The forest type map, based on the share of coniferous growing stock in the total growing stock is shown in Figure 6. The spatial distribution of the assessed data is determined with a pixel size of 25 ha. The higher growing stocks are concentrated in the Central European countries, in particular in mountain areas. Coherently with the ecological distribution of tree species, the growing stock of broadleaves increases at lower latitudes, while conifers have higher stocks at high latitude and altitude.

Fig. 3: Growing stock, coniferous per hectare of forest.

Fig. 4: Growing stock, broadleaved per hectare of forest.

Fig. 5: Total growing stock per hectare of forest.

Fig. 6: Forest type map, based on share of coniferous growing stock in total growing stock as defined in the legend.

The accuracy assessment shows a high correlation between the remote sensing based estimates and the estimates based on field inventories at the regional level with correlation coefficients of r = 0.96 for coniferous, r=0.94 for broadleaved and r=0.97 for total growing stock per hectare of forest. The mean absolute error of the estimations is 25 m³/ha for coniferous, 20 m³/ha for broadleaved and 25 m³/ha for total growing stock per hectare. Despite the high correlations, the accuracy assessment shows a slight underestimation of the growing stock in regions characterised by high growing stock volume. This may be caused by saturation effects, as optical sensors such as the utilised MODIS bands measure only the upper layers of vegetation cover. For those regions for which EFISCEN data was available for recalibration of the classification result, this underestimation is corrected. The accuracy assessment was performed at the regional level, but not at the pixel level. For future applications the accuracy should also be assessed for smaller reference units, up to the pixel level. However, this necessitates comprehensive data with area-wide coverage of reference sites, e.g. derived from laser scanning, which were not available for the current study. The results give area-wide spatial explicit information on the forest resources for the whole European Union, EFTA countries, the Balkans, Belarus, the Ukraine, Moldova, Armenia, Azerbaijan, Georgia and Turkey with a high spatial resolution of 25 ha per grid cell which was not available up to now. 3.2. Carbon stock in the above-ground biomass Carbon stock in the above-ground forest biomass was calculated separately for broadleaves and conifers with a spatial resolution of 500 meters, according to the spatial resolution of the MODIS satellite image data. Aggregated biomass conversion and expansion factors (BCEFs) were applied to convert the remote sensing based growing stock classification results to carbon stock of the above-ground forest biomass. As the accuracy of the aggregated BCEFs is not known for all regions, no separate validation of the carbon stock estimates was performed. The uncertainty of the biomass carbon stock using EFISCEN, which includes the uncertainty of BCEFs, was estimated to be quite low. The calculated coefficient of variation in 4 pilot countries was equal to 2-5% (Meyer et al. 2005). The biomass expansion factors in Great Britain, converting stem biomass to above-ground biomass, show a coefficient of variation of 4-14% (Levy et al. 2004). In Lethonen et al. (2004) a relative standard error of 3-21% was calculated for the BCEF converting growing stock to total biomass for different species and age classes. The higher uncertainties refer only to BCEFs for young age classes, while the relative standard error is about 3-5% for factors of mature stands. Similar ranges are also reported for the Czech Republic (Lethonen et al. 2007). The total sum of carbon stocks in above-ground forest biomass within a raster of 10km by 10km is shown in Figure 7. Whereas Figures 3 and 4 show the mean values of growing stock per hectare forest within 500m by 500m grid cells, Figure 7 shows the spatial distribution of the carbon stock aggregated within 10km by 10km grid cells.

Fig. 7: Total carbon stock of above-ground forest biomass within a 10km by 10km raster.

As for the growing stock, the highest densities of tree carbon stock in Europe can be found in mountain areas and in the southern part of Fenno-Scandia. A relevant amount of carbon stock also exists in some parts of the Baltic countries and Belarus. The lowest carbon stocks were calculated for the Mediterranean countries (Greece, Spain and Turkey) because of the low productivity of forests and in Great Britain and Ireland because of the limited extent of forest area. 4. Conclusions This study presents the estimates of coniferous and broadleaved growing stock as well as carbon stock of the above-ground biomass with a spatial resolution of 500m by 500m per grid cell. The mapped area is 5 million km² of which 2 million km² are forests. It covers the whole European Union, the EFTA countries, the Balkans, Belarus, the Ukraine, Moldova, Armenia, Azerbaijan, Georgia and Turkey. The classification results were evaluated by the comparison with regional estimates derived from national forest inventories independently of the classification. The validation shows a high correlation between the classification results and the field based estimates with correlation coefficients of r = 0.96 for coniferous, r=0.94 for broadleaved and r=0.97 for total growing stock per hectare. The mean absolute error of the estimations is 25 m³/ha for coniferous, 20 m³/ha for broadleaved and 25 m³/ha for total growing stock per hectare. Biomass conversion and expansion factors were applied to convert the growing stock classification results to carbon stock in above-ground biomass. An automatic up-scaling approach was applied to allow the up-scaling of forest inventory plot measurements with a typical area below 1 ha and MODIS satellite data with a pixel size of 25 ha. Contrary to conventional classification methods such as k-NN classification which relate the plot measurements directly to remote sensing data, this sampling based approach has the main advantage that it is not sensitive to scale mismatch between the area covered by the field measurements and the area covered by each pixel. This approach can be applied when sampling based reference data are used for automatic classification of remote sensing imagery. In the current study, the estimates are for the year 2000. In the future, the method will be applied to evaluate the growing stock and the carbon stock of European forests on a yearly basis. The development of similar estimates for different points in time will give the opportunity to analyse forest dynamics all over Europe by using satellite images and to fill information gaps in areas or in years where/when no inventory data exist. In addition the maps will provide a harmonised data source to be compared with other international datasets like forest statistics from FAO with the advantage of adding spatial information on forest resources. Acknowledgements The work was financed by the European Union within the project “CarboEurope-IP - Assessment of the European Terrestrial Carbon Balance”, Framework Progamme 6 (GOCE-CT-2003-505572), and by JOANNEUM RESEARCH Forschungsgesellschaft mbH., Austria. We’re greatly indebted to the Forest Focus programme and the national forest inventory institutes’ correspondents. Special thanks go to the many NFI field crews in the countries. NFI plot data were received from: Jacques Rondeux and Martine Waterinckx, Belgium; Juro Cavlovic, Croatia; Veiko Aderman, Estonia; Kari Korhonen, Finland; Thierry Bélouard, France; Heino Polley, Germany; Marino Vignoli, Remo Bertani, Giorgio Dalmasso & Maurizio Teobaldelli, Italy; Andrius Kuliesis, Lithuania; Wim Daamen and Henny Schoonderwoerd, Netherlands; Stein Tomter, Norway; Susanna Barreiro & Margarida Tomé, Portugal; Olivier Bouriaud, Romania; Vladimir Seben, Slovak Republic; Gal Kusar, Slovenia; J. Villanueva and Antoni Trasobares, Spain; Göran Kempe, Sweden; Bill Mason & Shona Cameron, United Kingdom; Igor Buksha, Ukraine. The collection of national forest inventory plot data was carried out under Eforwood-IP, Carbo Europe IP, ADAM-IP and in connection with COST E43. Gert-Jan Nabuurs and Geerten Hengeveld were partially funded under the BSIK-IC2 Programme. References: Anaya, J.A., Chuvieco, E., Palacios-Orueta, A., 2009. Aboveground biomass assessment in Colombia: A remote sensing approach. Forest Ecology and Management 257, 1237-1246. Blackard, J. A., Finco, M. V., Helmer, E. H., Holden, G. R., Hoppus, M. L., Jacobs, D. M., Lister, A.J., Moisen, G.G., Nelson, M.D., Riemann, R., Ruefenacht, B., Salajanu, D., Weyermann, D.L., Winterberger, K.C., Brandeis, T.J., Czaplewski, R.L., McRoberts, R.E., Patterson, P.L., Tycio, R.P., 2008. Mapping U.S. forest biomass using nationwide forest inventory data and moderate resolution information. Remote Sensing of Environment, 112, 1658−1677. Eggers, J., Lindner, M., Zudin, S., Zaehle, S., Liski, J., 2008. Impact of changing wood demand, climate and land use on European forest resources and carbon stocks during the 21st century. Global Change Biology 14, 1-16.

Foody, G.M., Campbell, N.A., Trodd, N.M.,Wood, T.F., 1992. Derivation and applications of probabilistic measures of class membership from the maximum likelihood classification. Photogrammetric Engineering and Remote Sensing 58 9, 1335–1341. Foody, G.M., Cox, D.P., 1994. Sub-pixel land cover composition estimation using a linear mixture model and fuzzy membership functions. International Journal of Remote Sensing 15 3, 619–631. Foody, G.M., Arora, M., 1996. Incorporating mixed pixels in the training, allocation and testing stages of supervised classifications, Pattern Recognition Letters 17 (1996), 1389–1398. Gasparini, P., Nocetti, M., Tabacchi, G., Tosi, V., 2006. Biomass equations and data for forest stands and shrublands of the Eastern Alps (Trentino, Italy). In: Sustainable Forestry in Theory and Practice: Recent Advances in Inventory and Monitoring, Statistics and Modeling, Information and Knowledge Management, and Policy Science (ed. Reynolds KM). US Department of Agriculture, Forest Service, Pacific Northwest Research Station Portland, OR, USA. Hansen, M. C., DeFries, R. S., Townshend, J. R. G., Carroll, M., Dimiceli, C., Sohlberg, R. A., 2003. Global percent tree cover at a spatial resolution of 500 meters: First results of the MODIS Vegetation Continuous Fields algorithm. Earth Interactions, 7,1– 15.

Häme, T., Stenberg, P., Andersson, K., Rauste, Y., Kennedy, P., Folving, S., Sarkeala, J., 2001. AVHRR-based forest proportion map of the Pan-European area. Remote Sensing of Environment, 77 1, 76-91. Hijmans, R.J., S.E. Cameron, J.L. Parra, P.G. Jones, A. Jarvis, 2005. Very high resolution interpolated climate surfaces for global land areas. International Journal of Climatology 25, 1965-1978. IPCC, 2003. Good Practice Guidance for Land Use, land-Use Change and Forestry. Penman, J., Gytarsky, M., Hiraishi, T., Kruger, D., Pipatti, R., Buendia, L., Miwa, K., Ngara, T., Tanabe, K., Wagner, F. (Eds). Intergovernmental Panel on Climate Change (IPCC), IPCC/IGES, Hayama, Japan. Lehtonen, A., Cienciala, E., Tatarinov, F., Mäkipää, R., 2007. Uncertainty estimation of biomass expansion factors for Norway spruce in the Czech Republic. Annals of Forest Science 64(2), 133-140. Lehtonen A., Mäkipää R., Heikkinen J., Sievänen R., Liski J., 2004. Biomass expansion factors (BEFs) for Scots pine, Norway spruce and birch according to stand age for boreal forests. Forest Ecology and Management 188, 211-224. Levy P.E., Hale S.E., Nicoll B.C., 2004. Biomass expansion factors and root:shoot ratios for coniferous tree species in Great Britain. Forestry 77 (5), 421-430. Liski J., Lehtonen A., Palosuo T., Peltoniemi M., Eggers T., Muukkonen P., Mäkipää R., 2006. Carbon accumulation in Finland’s forest 1922-2004 – an estimate obtained by combination of forest inventory data with modelling of biomass, litter and soil. Annals of Forest Science 63, 687-697. Lu, D., 2006. The potential and challenge of remote sensing-based biomass estimation. Int. Journal of Remote Sensing 27 7, 1297-1328. MCPFE, 2007. State of Europe's forests 2007. The MCPFE report on sustainable forest management in Europe. Ministerial Conference on the Protection of Forests in Europe, Warsaw, pp. 247. Meyer J., Vilén T., Peltoniemi M., Faubert P., Thürig E., Lindner M., 2005. Uncertainty estimate of the national level biomass and soil carbon stock and stock change. Final Report - CarboInvent Project, WP6-D6.3-EFI, Joanneum Research, http://www.joanneum.at/carboinvent/topdown_integration.php Mokany K., Raison R.J., Prokushkin A.S., 2006. Critical analysis of root: Shoot ratios in terrestrial biomes. Global Change Biology, 12, 84-96. Muukkonen, P., Heiskanen, J., 2007. Biomass estimation over a large area based on stand wise forest inventory data and ASTER and MODIS satellite data: a possibility to verify carbon inventories. Remote Sensing of Environment 107, 617–624. Nabuurs, G.J.M.M., Pussinen, A., Van Brusselen, J. Schelhaas, M., 2007. Future harvesting pressure on European forests. European Journal of Forest Research, 126, 391-400.

Nabuurs, G.J.M.M., G.M.Hengeveld, B. Van der Werf, N. Heidema, this issue, 2009. European forest carbon balance assessed with inventory based methods - an intro to the special feature. Päivinen R., Brusselen J., Schuck A., 2009. The growing stock of European forests using remote sensing and forest inventory data. Forestry Advance Access published online on May 27, 2009, Forestry, doi:10.1093/forestry/cpp017. Sallnäs, O., 1990. A matrix model of the Swedish forest. Studia Forestalia Suecica 183, 23. Schelhaas, M.-J., Eggers-Meyer, J., Lindner, M., Nabuurs, G.J.M.M., Päivinen, R., Schuck, A., Verkerk, P.J., Werf, D.C.v.d., Zudin, S., 2007. Model documentation for the European Forest Information Scenario model (EFISCEN 3.1.3). Alterra report 1559 and EFI technical report 26. Alterra and European Forest Institute, Wageningen and Joensuu, 118 pp. Strahler A.H., 1980. The use of prior probabilities in maximum likelihood classification of remotely sensed data, Remote Sensing of Environment 10 (1980), 135–163. Stümer W., Kenter B., Köhl M., this issue, 2009. Spatial interpolation of in-situ data by self-organizing map algorithms (neural network) for the assessment of carbon stock. Tomppo E., Nilsson M., Rosengren M., Aalto P., Kennedy P., 2002. Simultaneous use of Landsat-TM and IRS-1C WiFS data in estimating large area tree stem volume and aboveground biomass. Remote Sensing of Environment 85, 156-171. UNECE-FAO, 2000. Forest Resources of Europe, CIS, North America, Australia, Japan and New Zealand. Geneva Timber and Forest Study Papers, No. 17, United Nations, New York and Geneva, pp. 445. Vilén T., Meyer J., Thürig E., Lindner M., Green T., 2005. Improved regional and national level estimates of the carbon stock and stock change of tree biomass for six European countries. Final Report - CarboInvent Project, WP6-D6.2-EFI, Joanneum Research, http://www.joanneum.at/carboinvent/topdown_integration.php