Remote Sens. 2013, 5, 4503-4532; doi:10.3390/rs5094503 OPEN ACCESS Remote Sensing ISSN 2072–4292 www.mdpi.com/journal/remotesensing Article Estimates of Forest Growing Stock Volume for Sweden, Central Siberia, and Québec Using Envisat Advanced Synthetic Aperture Radar Backscatter Data Maurizio Santoro 1,*, Oliver Cartus 2, Johan E.S. Fransson 3, Anatoly Shvidenko 4, Ian McCallum 4, Ronald J. Hall 5, André Beaudoin 6, Christian Beer 7 and Christiane Schmullius 8 1 Gamma Remote Sensing, Worbstrasse 225, 3073 Gümligen, Switzerland; E-Mail: [email protected] 2 Woods Hole Research Center, Falmouth, MA 02540, USA; E-Mail: [email protected] 3 Department of Forest Resource Management, Swedish University of Agricultural Sciences, SE-901 83 Umeå, Sweden; E-Mail: [email protected] 4 International Institute of Applied Systems Analysis, A-2361 Laxenburg, Austria; E-Mails: [email protected], [email protected] 5 Northern Forestry Centre, Canadian Forest Service, Natural Resources Canada, Edmonton, AB T6H 3S5, Canada; E-Mail: [email protected] 6 Laurentian Forestry Centre, Canadian Forest Service, Natural Resources Canada, Sainte-Foy, QC G1V 4C7, Canada; E-Mail: [email protected] 7 Department of Applied Environmental Science (ITM) and Bert Bolin Centre for Climate Research, Stockholm University, SE-106 91 Stockholm, Sweden; E-Mail: [email protected] 8 Department of Earth Observation, Friedrich-Schiller University, D-07743 Jena, Germany; E-Mail: [email protected] * Author to whom correspondence should be addressed; E-Mail: [email protected]; Tel.: +41-31-951-7005; Fax: +41-31-951-7008. Received: 25 June 2013; in revised form: 29 August 2013 / Accepted: 5 September 2013 / Published: 12 September 2013 Abstract: A study was undertaken to assess Envisat Advanced Synthetic Aperture Radar (ASAR) ScanSAR data for quantifying forest growing stock volume (GSV) across three boreal regions with varying forest types, composition, and structure (Sweden, Central Siberia, and Québec). Estimates of GSV were obtained using hyper-temporal observations of the radar backscatter acquired by Envisat ASAR with the BIOMASAR algorithm. In total, 5.3⋅106 km2 were mapped with a 0.01° pixel size to obtain estimates representative for the year of 2005. Comparing the SAR-based estimates to spatially explicit datasets of Remote Sens. 2013, 5 4504 GSV, generated from forest field inventory and/or Earth Observation data, revealed similar spatial distributions of GSV. Nonetheless, the weak sensitivity of C-band backscatter to forest structural parameters introduced significant uncertainty to the estimated GSV at full resolution. Further discrepancies were observed in the case of different scales of the ASAR and the reference GSV and in areas of fragmented landscapes. Aggregation to 0.1° and 0.5° was then undertaken to generate coarse scale estimates of GSV. The agreement between ASAR and the reference GSV datasets improved; the relative difference at 0.5° was consistently within a magnitude of 20–30%. The results indicate an improvement of the characterization of forest GSV in the boreal zone with respect to currently available information. Keywords: SAR backscatter; Envisat ASAR; growing stock volume; boreal forest; Sweden; Siberia; Québec; BIOMASAR algorithm 1. Introduction The amount and spatial distribution of forest resources in the boreal zone are highly debated as they are often roughly quantified and seldom verified [1,2]. Continuous changes to forest resource distribution, due to changes in land use, natural and anthropogenic disturbances and recovery, and physiological or metabolic changes result in a need for periodic updating [3]. Gaps and errors in available datasets imply that investigations based on such records (e.g., for assessing carbon stocks) can suffer from substantial uncertainties. For the boreal zone, traditional field survey techniques have substantial limitations to achieve accurate and repeated characterization of forest conditions [4–6]. In addition, financial and logistical constraints may lead to inconsistent quality of field measurements of forest resources. The Nordic European countries can afford accurate forest inventories undertaken on a regular basis with short cycles, such as five year cycles [7]. Advanced field survey techniques are implemented to increase coverage and timeliness of the collected data. In contrast, the completion of a country-wide inventory of Russia spans at least a decade and the precision of the field survey depends to a large degree on land accessibility and financial means to support such surveys. In Siberia and the Far East, more than 50% of the forests have not been revisited for more than 20 years after the last inventory due to budget restrictions [8]. In Québec, the temperate and southern boreal forest areas (roughly 40% of the provincial forest land) are systematically and intensively inventoried each decade [9]. Satellite remote sensing supports mapping and monitoring of forest resources on a large scale because of its synoptic view, frequent revisit capability, relative low cost, and sensitivity of the observable to either biophysical or structural properties of forests. Satellite optical images in conjunction with plot-wise measurements from forest field inventory have been used to generate country-wide [10–12], continental [13,14], and global [15] data products of above-ground biomass, growing stock volume and canopy cover fraction. LiDAR (Light Detection And Ranging) point-wise measurements have been implemented in combination with either in situ or two-dimensional data from optical imagery to provide spatially explicit estimates of forest canopy height [16,17], above-ground Remote Sens. 2013, 5 4505 biomass [18–22], and above-ground carbon [23] at regional and global scales. Mapping efforts based on spaceborne synthetic aperture radar (SAR) are instead limited to investigations based on images acquired between 1994 and 2000 to obtain estimates of above-ground biomass in the boreal zone of Canada [24] and the Amazon basin [25], canopy height for the conterminous US [26], growing stock volume in Central Siberia [27–29] and Northeast China [30], and forest age for UK [31]. Despite the suitability of active microwaves for the estimation of forest structural parameters and the wealth of images acquired especially during the last decade by several spaceborne SAR sensors, there is minimal use of SAR images for extensive mapping of forest resources when compared to similar applications of optical and LiDAR data. Global datasets of SAR images, repeated acquisitions, and systematic availability of the data to the scientific community have only been guaranteed from the Envisat Advanced Synthetic Aperture Radar (ASAR) instrument operating at C-band, in the large swath (400 km) ScanSAR mode [32]. At C-band, the limited penetration of the microwaves into the canopy results in weak sensitivity to structural parameters such as growing stock volume (GSV) and implies significant retrieval errors. The error associated with the retrieval based on a single backscatter observation was reported on the order of 100% in boreal forest [33–35]. Only the combination of a large number of estimates of GSV from individual measurements of the ASAR backscatter could improve the retrieval results [36]. For three test sites in the boreal zone, the estimation error was on the order of 35–40% for the two spatial resolutions of the data when ASAR operated in ScanSAR mode (100 m and 1,000 m). Furthermore, the error decreased when aggregating GSV estimates at lower resolution, being below 25% when averaging over at least 100 pixels (e.g., from 1 km to 10 km). While such retrieval accuracy and the coarse resolution are not sufficient for operational mapping, estimates of GSV from Envisat ASAR ScanSAR have the potential to provide a first-order assessment in areas where GSV is poorly quantified (e.g., Northern Canada) or cannot be updated according to plan (e.g., Siberia). Accuracy and spatial resolution of the ASAR GSV estimates are also sufficient for the initialization of global carbon/biosphere models in view of more accurate modeling of the land-atmosphere CO exchange compared to the current approaches implemented in these models [37,38]. 2 The objective of this study was to assess the contribution of Envisat ASAR ScanSAR hyper-temporal datasets of the backscatter for generating spatially explicit estimates of forest GSV at regional scale in the boreal zone. For this, the retrieval approach referred to as the BIOMASAR algorithm [36] has been applied across three large boreal regions with varying forest types, composition and structure (Sweden, Central Siberia, and Québec). The GSV was retrieved and evaluated at the native resolution of the ASAR data (0.01° corresponding approximately to 1 km). Furthermore, aggregated estimates at the scale of 0.1° and 0.5° were obtained and evaluated. These latter scales are commonly used by biogeochemical models to predict carbon pools [39]. The contribution of ASAR estimates of GSV in the context of current knowledge of carbon stocks in the three study regions was assessed by comparing the results against several datasets of spatially explicit estimates of GSV or above-ground biomass, and official statistics at regional and national level. Remote Sens. 2013, 5 4506 2. Study Regions 2.1. Sweden Sweden covers 450,295 km2 between 56°N and 69°N, and 10°E and 24°E. Forest is the predominant land cover (55%), with wetlands, grassland, arable land, and water bodies representing other major land cover classes [40] (Figure 1). Topography is mostly flat to moderately flat; in the northwest, mountain ridges run along the border with Norway. Forests are primarily boreal and hemi-boreal coniferous species (Norway spruce and Scotch pine). Birch is the main deciduous tree species. Forest cover, growth, and GSV are strongly related to climate. The GSV is highest at the southern latitudes, corresponding to the transition from boreal to hemi-boreal forests. Above 60°N, GSV is higher along the coast as compared to inland. 2.2. Central Siberia Central Siberia is a 3.28 million km2 region between 52°N and 72°N, and 88°E and 110°E, containing the Krasnoyarsk Kray and the Irkutsk Oblast administrative regions. These regions include almost all vegetation zones located within the northern hemisphere with forested areas corresponding to the predominant land cover (59% in Krasnoyarsk Kray and 82% in Irkutsk Oblast) [41] (Figure 1). The remaining land cover classes consist of pasture/grassland, agriculture in the southern plains, wetlands, water bodies, and urban settlements. The topography in these regions ranges from mostly flat to moderately hilly. The southernmost part of the region includes the Sayan Mountains range, with elevations between 2,000 m and 2,700 m and areas of steep slopes. The study region is completely within the boreal zone with the exception of a relatively narrow polar belt along the Arctic shoreline. Forest cover, growth, and GSV decrease with increasing latitude. At the northernmost latitudes, the land cover consists primarily of tundra vegetation, which changes naturally from Arctic tundra to shrubland and forest tundra. South of the tundra zone, the dominant forest species is larch. Towards the south, conifers (mostly pine, spruce, fir, and Siberian cedar) represent the mature stage of forests, whereas young forests are mostly characterized by deciduous pioneer species such as birch and aspen. This is a result of natural regeneration following forest disturbance or natural succession, which has a rotation of about 70–100 years. 2.3. Québec The province of Québec in eastern Canada covers 1,667,441 km2 between 45°N and 63°N, and −80°E and −57°E. Nearly half of the area is covered by forests representing about 20% of all forested lands in Canada (Figure 1, [42]). The topography is mostly flat with some gently rolling hills and steep-sided low altitude mountains in localized regions. Numerous water bodies occupy more than 20% of the territory. The forests of Québec belong to three major vegetation zones that fall predominantly into the Boreal Shield and Taiga Shield described by the terrestrial ecozones of Canada [43]. The southern region is characterized by deciduous and mixed temperate forests dominated by maple, yellow birch, and balsam fir with the largest GSV found in Québec. The large central region in the southern half includes dense boreal forests composed mainly of black spruce, balsam fir, and white birch, whereas Remote Sens. 2013, 5 4507 the northern half corresponds to the large boreal/taiga/tundra south/north transition zone dominated by increasingly sparse and fragmented black spruce forests. The northern region includes tundra dominated by shrubland. Figure 1. Land cover patterns according to the GLC2000 land cover product [42] for the study regions. Brown corresponds to shrubland. Olive corresponds to larch forest. Light and dark green tones correspond to broadleaved/mixed and coniferous species, respectively. Yellow corresponds to cropland. 3. Material and Methods This Section provides an overview of (i) the SAR datasets, (ii) the BIOMASAR algorithm used to estimate GSV from the SAR data, (iii) the datasets used for validating or comparing the SAR-based GSV estimates and assessing the reliability of SAR-based GSV, and (iv) the approach followed for the inter-comparison. 3.1. SAR Datasets The SAR datasets consisted of images of the radar backscatter acquired by Envisat ASAR in ScanSAR mode. This imaging mode covered a swath of approximately 400 km, which in turn implied a high frequency of observations because of the large overlap of adjacent swaths. At 60°N, measurements were possible on a daily basis. Depending on the sampling bit rate, the image products were characterized by a spatial resolution of approximately 150 m (Wide Swath Mode, WSM) or 1,000 m (Global Monitoring Mode, GMM) [32]. Data are provided to the user by the European Space Agency (ESA), slightly oversampled (75 m for WSM and 500 m for GMM). GMM data were acquired regularly, whereas WSM data were acquired either upon request or according to strategic planning by ESA. In ScanSAR mode, ASAR acquired either (transmit-receive) Horizontal-Horizontal, HH, or Vertical-Vertical, VV, polarized data. For each study region, all GMM images acquired between December 2004 and February 2006, were used. For South Sweden, all WSM images acquired during the same time period were also used because Remote Sens. 2013, 5 4508 of relatively poor coverage in GMM. The time interval maximized the number of winter-time images in the SAR backscatter time series, which were found to be most suitable for retrieving GSV [36]. The average and the standard deviation of the number of backscatter measurements are listed in Table 1 for each study region. The slightly larger standard deviation in Québec is a result of substantially less acquisitions south of 50°N. At least 50 backscatter measurements per pixel were available over forested areas which was found to be sufficient to obtain a reliable spatial representation of the GSV patterns [36]. To the best of our knowledge, no significant disturbance occurred within the time span of the SAR datasets. The dataset consisted of images acquired in HH- and VV-polarizations. Table 1. Mean value and standard deviation of the number of synthetic aperture radar (SAR) backscatter observations. Region Mean Value Standard Deviation Sweden 76 16 Central Siberia 93 26 Québec 140 38 3.2. BIOMASAR Algorithm The BIOMASAR algorithm [36] consists of (i) a SAR processing block to obtain calibrated, geocoded and co-registered images of the radar backscatter, and (ii) a GSV retrieval block where each of the radar backscatter measurements is inverted to GSV by means of a semi-empirical forest backscatter model and individual GSV estimates are blended with a weighted linear combination to form the final estimate of GSV. A summary of the main components of the algorithm and details about the implementation for this study are given below. Further information regarding the development and validation of the BIOMASAR algorithm has been reported in [36]. 3.2.1. SAR Processing The SAR data were obtained in the form of image strips of backscattered intensity in the geometry of acquisition of the radar sensor. Each data strip was multi-looked (i.e., boxcar averaged) to reduce speckle noise. As a trade-off between spatial resolution and speckle noise reduction, the data were multi-looked to form pixels with a size of 1,000 m. Each strip was then terrain geocoded [44] to the equiangular projection and to the pixel size of 0.01°. Terrain information was available from Digital Elevation Models (DEMs) [45–47]. The resulting geocoding accuracy was on the order of 1/3 of the pixel size. The SAR backscatter was then corrected for the pixel area determined from the DEM [48]. The SAR images were subsequently tiled according to a regular 2° × 2° grid for optimal management of computing resources. To further mitigate speckle noise, each set of images within a tile was filtered with a multi-channel approach [49]. The estimated Equivalent Number of Looks (ENL) [50] was close to 60 in the case of a GMM dataset with approximately 100 measurements or a WSM dataset. For GMM datasets with less than 50 observations, the ENL was close to 40. The corresponding uncertainty of the measured backscatter was 0.4 dB and 0.5 dB, respectively. Although such an uncertainty can play a significant role on the single image retrieval because of the exponential type of model used for the retrieval (see Section 3.2.2), the impact on the multi-temporal retrieval is much weaker because of Remote Sens. 2013, 5 4509 the large number of images that are combined. In [36], the average uncertainty related to the retrieval approach was on average 10%. 3.2.2. Forest Backscatter Model and Estimation of Model Parameters The retrieval was based on a Water Cloud type of model linking the forest GSV, V, in m3/ha to the forest backscatter coefficient, σ0 , in linear scale [33,36,51,52]. This modeling solution allows for a for complete characterization of the backscattered intensity from a forest at C-band in terms of its main contributions with a simple formulation. ( ) σo =σo e−βV +σo 1−e−βV (1) for gr veg In Equation (1), the forest backscatter is expressed as a contribution from the ground with a backscatter coefficient of σ0 , and from the vegetation layer with a backscatter coefficient of σ0 , both gr veg expressed in the linear scale. The backscatter coefficients are influenced by the environmental conditions at the time of image acquisition as well as by look geometry and polarization. Nonetheless, at coarse spatial resolution, the environmental conditions are mostly affecting the backscatter coefficients [38]. Both contributions are weighted by the forest transmissivity expressed as e−βV, where β is an empirically defined coefficient (unit: ha/m3). Higher order scattering contributions (double and multiple bounces) are not included as shown to be of minor importance at C-band in boreal forest [51,53]. The two backscatter coefficients σ0 and σ0 and the coefficient β are unknown. In [36], we gr veg concluded that using a constant value for the latter (0.006 ha/m3) affected the retrieval only marginally when compared to a more rigorous estimate, which would require knowledge of local environmental conditions (i.e., weather data) as the transmissivity depends on the dielectric properties of the canopy and the amount of gaps within it. As the SAR backscatter at C-band is characterized by significant spatial variability as a consequence of the strong effect of environmental conditions, the two remaining unknown model parameters were estimated on a pixel-by-pixel basis. The backscatter coefficient σ0 was estimated by means of an automated procedure involving the gr computation of the mean value of the measured backscatter for pixels that can be labeled as “unvegetated” within a window centered on the pixel of interest [36]. The mask supporting the selection of backscatter measurements corresponding to unvegetated pixels was based on the MODIS Vegetation Continuous Field (VCF) tree cover product [54]. An adaptive threshold for the upper tree cover percentage and an adaptive window size were used [36]. The backscatter coefficient σ0 was obtained in a similar manner using measurements of the veg backscatter coefficient corresponding to pixels labeled as “dense forest”. Dense forests were identified using the MODIS VCF product based on a minimum threshold on tree cover percentage [36]. Threshold and window size were adaptive. The mean value of the backscatter coefficient of “dense forest”, referred to as σ0 , was then compensated for the proportion of ground contribution to obtain df the effective backscatter coefficient of the vegetation layer σ0 . The compensation term was derived veg from Equation (1) by expressing the backscatter coefficient σ0 as a function of the measured forest veg backscatter coefficient of dense forest and the modeled ground contribution. σo −σo e−βVdf σo = df gr (2) veg 1−e−βVdf Remote Sens. 2013, 5 4510 In Equation (2), V represents the GSV value assumed to be representative for “dense forest”. By df definition, V corresponds to the 90th percentile of the GSV distribution in the area of interest [36]. In df support of the definition, knowledge of the cumulative distribution function (cdf) of GSV within the sampling unit of the parameter is required. In the case where the cdf is unknown, the value is estimated depending on which parameters of the cdf are available or can be derived, for example, from the literature. It was here assumed that a single value is representative for a 2° × 2° tile. To obtain a smooth variation of the parameter between adjacent tiles, the individual values were interpolated using a bilinear function. The parameter V was also used to set the maximum GSV that can be df retrieved [36]. As a consequence, erroneous values of V could bias the retrieved GSV. In the case of df biases greater than 20% over several adjacent tiles, fine-tuning of V was applied. The presence of df biases was assessed by comparing aggregated ASAR GSV and GSV of a reference dataset (see Table 2) at the very coarse resolution of 2°. 3.2.3. GSV Retrieval and Multi-Temporal Combination of GSV Estimates For a measurement of the forest backscatter coefficient σ0 , the forest backscatter model in for Equation (1) is inverted to retrieve GSV using the corresponding estimates of σ0 and σ0 . gr veg 1 σo −σo  V = − ln for veg  β σo −σo  (3)  gr veg  The individual GSV estimates, V, are finally used in a weighted linear combination to form the i multi-temporal estimate of GSV [35]. N w  i V w i V = i=1 max (4) mt N w  i w i=1 max In Equation (4), the multi-temporal estimate of GSV, V , expresses the sum of the individual GSV mt estimates V weighted by the backscatter difference between vegetation and ground (w = σ0 − σ0 , i i veg gr in the dB scale). The coefficient w is equal to the largest of the weights w. Such weighting approach max i maximizes the contribution of GSV estimates from images characterized by a large contrast between forest and non-forest backscatter and avoids that images with little backscatter variability between low and high GSV can distort the final GSV estimate [36]. N represents the number of estimates being combined. Estimates corresponding to w below 0.5 were discarded because likely to be uncertain [36]. i In this study, GSV estimates were obtained at 0.01°. In addition, spatial averaging was applied to derive estimates at coarser spatial resolution. Our interest was primarily to derive estimates for spatial resolutions commonly used in carbon and biosphere models (0.1° and 0.5°). 3.3. GSV Datasets Three types of GSV datasets were used in this study to assess the accuracy and the reliability of GSV estimated with the BIOMASAR algorithm from the ASAR datasets: forest field inventory Remote Sens. 2013, 5 4511 datasets, spatially explicit estimates obtained from Earth Observation (EO) data, and global estimates of GSV or carbon stocks obtained from multiple input sources. Forest field inventory datasets qualify as validation datasets. Nonetheless, if the scale of the forest field inventory data does not match the scale of the EO data, the characterization of GSV in each of the datasets is different. Therefore, they are of limited usefulness to derive correct information on the retrieval accuracy of GSV from the remote sensing data. In this study, we considered several datasets based on forest field inventory measurements at multiple scales to finally assess the impact of scales on the GSV retrieval. Estimates of GSV from forest field measurements are often used in combination with EO satellite imagery to provide spatially explicit estimates of GSV. This procedure is implemented by several national forest inventories, for example in Sweden and Canada. The accuracy of the estimates of the forest variable of interest can be low at the pixel level whereas with aggregation of adjacent pixels, increased accuracy can be achieved [10,55]. In this study, we considered EO-based GSV raster datasets that are produced by, or in support of, national forest inventories and therefore assumed to be sufficiently reliable to identify inconsistencies in the ASAR GSV estimates. Global maps of GSV (or alternatively carbon stocks) derived from a combination of multiple data sources (in situ measurements, administrative statistics, model predictions, and remote sensing datasets) were considered as they cover all regions and, thus, allow comparison of results among them with a common basis. Although such datasets can only be considered for comparison purposes as they do not qualify as reference sets for validation, they were of interest to benchmark the ASAR GSV and assess its overall reliability towards quantifying forest resources and carbon stocks in the boreal zone and to pinpoint areas of discrepancy. Table 2 provides an overview of the three different types of GSV datasets for the three study regions ([6,7,10,29,39,55–57]). If necessary, datasets were re-projected and resampled into the geographic projection used for the ASAR data. Further details about each dataset are provided in related Sections below. 3.3.1. Forest Field Inventory Datasets Plot-wise measurements of GSV from the Swedish National Forest Inventory (NFI) were available for the entire country (Table 2). The Swedish NFI data consisted of GSV measured within plots of 7 m or 10 m radius, located on a pre-defined grid throughout the country [7]. For this study, the GSV from NFI plots field inventoried between 2004 and 2006 were averaged using 1 km × 1 km grid cells to form estimates at the same spatial scale as the estimates of ASAR GSV. In total, 11,425 GSV estimates were obtained. The number of recruited inventory plots within an ASAR pixel ranged from 1 to 13 with 90% of the 1 km estimates based on less than 4 plots. For Central Siberia, a first dataset consisted of forest field inventory data arranged to form polygon- wise measurements of GSV. A polygon corresponds to an area with similar forest properties in terms of tree species, age, productivity, GSV and local homogeneity. Such data were available from ten forest enterprises located in Irkutsk Oblast [29] (Tables 2 and 3) and originated from forest surveys carried out according to the Russian Forest Inventory Manual [58]. Forest stand boundaries were based on manual interpretation of aerial photographs. The dataset was last inventoried in 1998, with updates to 2003 to Remote Sens. 2013, 5 4512 reflect forest cover changes. While the error of the field inventory data was within requirements to not exceed 15% of GSV of the Forest Inventory, measurements in dense forest were systematically underestimated by approximately 10% [59]. Accordingly, a correction factor was applied. To take into account the 7-year time difference between the field inventory and the acquisition of the ASAR data, yearly growth factors were applied [36]. This corresponded to a yearly increase of productivity in the taiga forest over the recent decades of 0.3–0.5% per year [60]. For the inter-comparison, the data were rasterized to 100 m, averaged to 1 km and resampled to 0.01° pixel size. Table 2. Description of growing stock volume (GSV) and carbon stocks (last two) datasets in terms of input spatial scale, output spatial scale and processing applied. Change of map projection refers to transformation from the original map projection to the map projection of the Advanced Synthetic Aperture Radar (ASAR) dataset and its pixel size. Pixel Pixel Size/Scale Dataset/Site/Origin Size Processing (Native) (Output) Average of measurements within NFI plots/Sweden/ 7–10 m 0.01° 1 km grid cells, change of map forest field inventory [7] diameter projection kNN Sweden 2005/Sweden/ Spatial average to 1 km pixel size, 25 m 0.01° EO raster [10] change of map projection Forest polygons/Central Siberia/ Rasterize to 1 km pixel size, ~1 km2 area 0.01° forest field inventory [29] change of map projection Vegetation database/Central Siberia/ Rasterize to 1 km pixel size, ~100 km2 area 0.01° forest field inventory, aerial photography [39] change of map projection Spatial average to 1 km pixel size, EOSD/Québec/EO raster [55] 30 m 0.01° change of map projection IIASA GSV/ global/multiple sources [6] 0.5° 0.5° – Olson database of carbon stocks/global/ 0.089° 0.1° Resampling to output pixel size multiple sources [56] Ruesch & Gibbs dataset of carbon stocks/global/ 0.0083° 0.01° Resampling to output pixel size multiple sources [57] A second dataset was a regional vegetation database covering the entire study region [39] (Tables 2 and 3). The polygons of the vegetation database were delineated by Russian regional forest inventory and vegetation experts using aerial photographs. Although the polygons were defined in terms of broad homogeneous forest cover, occasionally they could include very different forest cover conditions on relatively small areas (e.g., harvested forest and dense mature forest), which are not accounted for by field measurements. Due to the vastness of the area and the remoteness of large parts of the region, the field inventory data had different levels of accuracy and currency. The majority of the data in the vegetation database stemmed from inventories undertaken at the end of the 1990s. For some areas in the north of the study region, the initial inventory data were 20–30 years old. For the southernmost regions, availability of recent aerial photography and satellite images allowed for correction of obsolete values. Exact information on the year of inventory and accuracy of the data were not available