Journal of Geophysical Research: Biogeosciences

Upscaling terrestrial carbon dioxide fluxes in Alaska with satellite remote sensing and support vector regression

Authors


Corresponding author: M. Ueyama, Graduate School of Life and Environmental Sciences, Osaka Prefecture University, 1-1 Gakuen-cho, Sakai, Osaka 599-8531, Japan. (miyabi-flux@muh.biglobe.ne.jp)

Abstract

[1] Carbon dioxide (CO2) fluxes from a network of 21 eddy covariance towers were upscaled to estimate the Alaskan CO2 budget from 2000 to 2011 by combining satellite remote sensing data, disturbance information, and a support vector regression model. Data were compared with the CO2 budget from an inverse model (CarbonTracker). Observed gross primary productivity (GPP), ecosystem respiration (RE), and net ecosystem exchange (NEE) were each well reproduced by the model on the site scale; root-mean-square errors (RMSEs) for GPP, RE, and NEE were 0.52, 0.23, and 0.48 g C m−2 d−1, respectively. Landcover classification was the most important input for predicting GPP, whereas visible reflectance index of green ratio was the most important input for predicting RE. During the period of 2000–2011, predicted GPP and RE were 369 ± 22 and 362 ± 12 Tg C yr−1 (mean ± interannual variability) for Alaska, respectively, indicating an approximately neutral CO2 budget for the decade. CarbonTracker also showed an approximately neutral CO2 budget during 2000–2011 (growing season RMSE = 14 g C m−2 season−1; annual RMSE = 13 g C m−2 yr−1). Interannual CO2 flux variability was positively correlated with air temperature anomalies from June to August, with Alaska acting as a greater CO2 sink in warmer years. CO2 flux trends for the decade were clear in disturbed ecosystems; positive trends in GPP and CO2 sink were observed in areas where vegetation recovered for about 20 years after fire.

1 Introduction

[2] Ongoing climate change over northern high latitudes has detectable impacts on terrestrial ecosystems [Chapin et al., 2005; Hinzman et al., 2005], where summers have warmed by up to 0.4°C decade−1 in Arctic Alaska and western Canada in recent decades [Chapin et al., 2005]. While the warming climate has increased vegetation productivity due to an increase in the length of the growing season in some areas [Euskirchen et al., 2006, 2009], warming has also stimulated respiration in other areas [Oechel et al., 1993, 2000; Ueyama et al., 2009; Hayes et al., 2011]. Moreover, warming has also intensified drought, impacted the fire regime, increased permafrost degradation [Hinzman et al., 2005; Kasischke and Turetsky, 2006], and decreased vegetation productivity [Barber et al., 2000; Zhang et al., 2007]. A major concern is that associated change in high-latitude ecosystems may amplify or dampen climate change through land and atmospheric feedbacks [Chapin et al., 2005; Randerson et al., 2006].

[3] Eddy covariance measurements have been conducted in various ecosystems in the Arctic and subarctic regions in order to understand high-latitude terrestrial carbon cycling [Oechel et al., 2000; Harazono et al., 2003; Kwon et al., 2006; Ueyama et al., 2006]. These measurements have provided important information on the impact of climate variability in CO2 exchange [Harazono et al., 2003; Ueyama et al., 2006; Welp et al., 2007], the effects of soil moisture on Arctic tundra respiration [Oechel et al., 1993; Kwon et al., 2006; Huemmrich et al., 2010], estimates of winter soil CO2 emissions [Oechel et al., 1997], the role of fire disturbance on CO2 fluxes [Welp et al., 2007; Iwata et al., 2011; Rocha and Shaver, 2011a, 2011b], and estimates of the spatial distributions of CO2 fluxes [McFadden et al., 2003; Beringer et al., 2005; M. Ueyama et al., Growing season and spatial variations of carbon fluxes of arctic and boreal ecosystems in Alaska, Ecol., Appl., in press, 2013]. Although site-scale studies clarified important processes controlling the high-latitude terrestrial carbon cycle, the importance of some processes, such as fire disturbance, on the regional-scale exchange is still unclear [McGuire et al., 2012].

[4] Recent data have allowed for the estimation of regional CO2 fluxes using models that combine remote sensing data with eddy covariance-derived CO2 fluxes [e.g., Papale and Valentini, 2003; Yang et al., 2007; Date et al., 2009; Ueyama et al., 2010; Jung et al., 2009]. These methods include light-use efficiency models (the concept that photosynthetic productivity is proportional to light absorption) [e.g., Heinsch et al., 2006], empirical regressions (e.g., multiple linear and stepwise regressions) [Vourlitis et al., 2003; Date et al., 2009; Huemmrich et al., 2009; Ueyama et al., 2010], and machine-learning-based regressions [e.g., Yang et al., 2007; Jung et al., 2009; Xiao et al., 2010].

[5] Machine-learning-based regressions are becoming widely used because the models are driven by observed data without complicated assumptions and a large number of parameters. These methods have been used to upscale gross primary productivity (GPP) [Yang et al., 2007; Xiao et al., 2010] and evapotranspiration [Yang et al., 2006] by combining eddy covariance CO2 and H2O fluxes to satellite remote sensing data (e.g., vegetation indices, land surface temperature, and landcover data).

[6] Support vector regression (SVR) is a kind of machine-learning-based regression technique. Previous studies have applied SVR to upscale evapotranspiration in the U.S. [Yang et al., 2006; Ichii et al., 2009] and GPP in the U.S. [Yang et al., 2007] and East Asia [Saigusa et al., 2010]. Yang et al. [2006] showed better performance of SVR than multiple linear regressions and an artificial neural network to predict continental-scale evapotranspiration by comparison of the obtained root-mean-square errors (RMSEs) in a cross-validation experiment of AmeriFlux data. In addition, a better performance of SVR than artificial neural network methods was reported in retrieval of oceanic chlorophyll concentration [Zhan et al., 2003].

[7] Top-down approaches based on atmospheric CO2 concentration measurements with inversion models provide independent measures of regional CO2 balance [Hayes et al., 2011; McGuire et al., 2012]. These approaches better estimate CO2 balance than bottom-up approaches at large spatial scales over long time periods, when specific regions are well constrained by measurements. The CarbonTracker is a state-of-the-art data assimilation system to estimate CO2 balance using the top-down approach [Peters et al., 2007]. Previously, the CarbonTracker estimated CO2 balance of North American continent with 1° spatial resolutions [Peters et al., 2007], but has not been well evaluated for high-latitude regions, such as Alaska. Comparisons between bottom-up and top-down approaches could clarify robustness and limitation of regional CO2 balance.

[8] Among high-latitude regions, Alaska supports a relatively dense network of long-term eddy covariance measurements, across various boreal and tundra ecosystems [M. Ueyama et al., in press, 2013]. Based on the previous synthesis of this eddy covariance network, satellite-derived leaf area index (LAI) combined with growing degree days and growing season length, explained much spatial variation in the growing season CO2 budget for nondisturbed ecosystems [M. Ueyama et al., in press, 2013]. By combining satellite-derived data, climate, and disturbance information, this network can be useful in creating detailed maps of CO2 fluxes using a machine-learning technique.

[9] In this study, the regional CO2 budget of Alaska was estimated based on machine-learning techniques using SVR methods with observed eddy covariance CO2 fluxes and satellite remote sensing data. Sufficient CO2 flux data from northern high latitudes have only recently become available, making possible the application of machine-learning techniques for this region. We incorporated fire history and observed CO2 fluxes at fire-disturbed sites into the SVR-based modeling in order to understand changes in CO2 flux associated with fires. Our objectives were to (1) accurately estimate regional fluxes including GPP, ecosystem respiration (RE), and net ecosystem exchange (NEE), and (2) examine interannual variations in regional fluxes from 2000 to 2011. The applicability and limitations of the upscaling technique were examined by comparing fluxes estimated by bottom-up (our SVR) and top-down (CarbonTracker) approaches.

2 Method

2.1 Study Region

[10] Our study focused on boreal forests and Arctic tundra of Alaska, covering the region of 72°N–52°N, 140°W–170°W (Figure 1). Based on vegetation type, climate, and soil characteristics, the study region can be divided into two major regions—the Arctic and interior Alaska. The Arctic is located north of the Brooks Range and is covered by tundra vegetation on continuous permafrost. Mean annual air temperature is −12°C, with the warmest monthly temperatures between 4 and 10°C occurring in July, and annual precipitation ranging from 100 to 130 mm yr−1 [Shulski and Wendler, 2007]. Interior Alaska lies north of the Alaska Range and south of the Brooks Range and is mostly covered by boreal forest over discontinuous permafrost. Mean annual air temperature is a few degrees below freezing, the warmest monthly temperature is 20°C in July, and annual precipitation ranges from 300 to 460 mm yr−1 [Shulski and Wendler, 2007].

Figure 1.

The locations of the eddy flux towers used in this study. The landcover is created by merging a state wide vegetation map of Alaska (Fleming, 1999) and fire location map by Alaska Fire Service. Fire scars in the figure are for the year of 2004, as an example. ENF and DBF represent evergreen needleleaf forest and deciduous broadleaf forest, respectively.

2.2 Observed CO2 Fluxes

[11] We used CO2 fluxes of NEE, GPP, and RE based on the eddy covariance method from 21 flux tower sites in Alaska (Figure 1 and Table 1) for the model development and validation. Tower sites are located across various representative ecosystems in Alaska, including wet sedge tundra (BRW, CMS, BEC, and IMW), moist tussock tundra (ATQ, IVO, IMT, and ARU), moist tundra (COT), heath tundra (IMH), burned tundra (ARS and ARM), shrub (COS), black spruce forest (PFA, FAI, and DLS), white spruce forest (COF), deciduous aspen forest (DLA), and burned black spruce forest (CRF, PFR, and DLB). For 13 of these sites, data for both winter and summer growing season were available, while for 7 of these sites, only summer growing season data were available (Table 1). Data for the COS site were only available for 1 month from mid-May to late June, but the site was nevertheless included in this study as the shrub tundra ecosystem type is currently underrepresented in the network of Alaska eddy covariance towers. All measurements were conducted with the eddy covariance method by using open-path gas analyzers at 17 sites (BRW, CMS, BEC, ATQ, IVO, IMH, IMT, IMW, ARS, ARM, ARU, CRF, PFR, FAI, DLS, DLA, and DLB), closed-path gas analyzers at 3 sites (COS, COF, and COT), and an enclosed-path gas analyzer at 1 site (PFA). The details of the measurement systems are shown in Table 1 and M. Ueyama et al., in press, 2013.

Table 1. Site Characteristics, Including Dominant Vegetation Type, Presence of Permafrost, Years Used in the Analysis, Instruments Used in the Eddy Covariance Measurements, and References
Site Name (Code)LocationElevation (m)Vegetation TypePermafrostYearAvailabilitybSonic anemometerInfrared gas analyzerReferences
  1. aElevation was derived from Google Earth©.
  2. bFY and GS represent full year and growing season, respectively.
  3. cData for COS are only available from DOY 131 to 171 in 2000.
Barrow (BRW)71.32°N, 156.61°W1Wet Sedge TundraPresence00–02FYR3 (Gill)ATDD (NOAA), LI-7500 (Li-Cor)Kwon et al. [2006]
Central Marsh (CMS)71.31°N, 156.62°W1Wet Sedge TundraPresence00–05FYDA600 (Kaijo)E009a (Advanet), LI-7500 (Li-Cor)Harazono et al. [2003]
Barrow Environmental Observatory (BEC)71.28°N, 156.60°W5aWet Sedge TundraPresence05–08GSWindmaster Pro (Gill)LI-7500 (Li-Cor)Zona et al. [2009]; Zona et al. [2011]; Zona et al. [2012]
Atqasuk (ATQ)70.47°N, 157.41°W23aMoist Tussock TundraPresence01,04–06FYR3 (Gill)ATDD (NOAA), LI-7500 (Li-Cor)Kwon et al. [2006]
Ivotuk (IVO)68.49°N, 155.75°W570aMoist Tussock TundraPresence03–07FYR3 (Gill)LI-7500 (Li-Cor)Epstein et al. [2004]
Imnavait Creek (IMH)68.61°N, 149.30°W940Heath TundraPresence08–10FYCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Euskirchen et al. [2012]
Imnavait Creek (IMT)68.61°N, 149.30°W930Moist Tussock TundraPresence08–10FYCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Euskirchen et al. [2012]
Imnavait Creek (IMW)68.61vN, 149.31°W920Wet Sedge TundraPresence08–10FYCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Euskirchen et al. [2012]
Anaktuvuk River (ARS)68.98°N, 150.28°W379aSeverely Burned TundraPresence08–09GSCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Rocha and Shaver [2011a, 2011b]
Anaktuvuk River (ARM)68.95°N, 150.21°W402aModerately Burned TundraPresence08–09GSCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Rocha and Shaver [2011a, 2011b]
Anaktuvuk River (ARU)68.93°N, 150.27°W428aTussock TundraPresence08–09GSCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Rocha and Shaver [2011a, 2011b]
Cascaden Ridge (CRF)65.39°N, 148.94°W314aBurned ForestNo Presence11GS81000 (Young)LI-7500 (Li-Cor) 
Poker Flat Research Range (PFA)65.12°N, 147.49°W211aBlack Spruce ForestPresence11FYWindmaster Pro (Gill)LI-7200 (Li-Cor)Nakai et al. [2013]
Poker Flat Research Range (PFR)65.12°N, 147.43°W491Burned ForestNo Presence08–11FYWindmaster (Gill)LI-7500 (Li-Cor)Iwata et al. [2011]
Fairbanks (FAI)64.87°N, 147.86°W157Black Spruce ForestPresence03–10FYCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Ueyama et al. [2006]; Iwata et al. [2010]
Council (COS)64.94°N, 164.74°W137ShrubNo Presence003 weekscmodel HS (Gill)LI-6262 (Li-Cor)Beringer et al. [2003, 2005]
Council (COF)64.91°N, 163.67°W83White Spruce ForestNo Presence00GSmodel HS (Gill)LI-6262 (Li-Cor)Beringer et al. [2003, 2005]
Council (COT)64.84°N, 163.69°W49Moist TundraPresence00GSmodel HS (Gill)LI-6262 (Li-Cor)Beringer et al. [2003, 2005]
Delta Junction (DLS)63.92°N, 145.75°W492aBlack Spruce ForestNo Presence02–04FYCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Randerson et al. [2006]; Welp et al. [2007]; Liu and Randerson [2008]
Delta Junction (DLA)63.92°N, 145.37°W384aAspen ForestNo Presence02–04FYCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Randerson et al. [2006]; Welp et al. [2007]; Liu and Randerson [2008]
Delta Junction (DLB)63.89°N, 145.74°W455aBurned ForestNo Presence02–04FYCSAT3 (Campbell Scientific, Inc.)LI-7500 (Li-Cor)Randerson et al. [2006]; Liu and Randerson [2008]

[12] Continuous estimates of NEE, GPP, and RE were generated by applying standardized quality control, gap-filling, and flux-partitioning procedures to the measured data. Quality control of eddy covariance data was primarily conducted (e.g., through applying stationary and high-order moment tests) by site managers, and remaining outliners and spike-like data were carefully removed manually [M. Ueyama et al., in press, 2013]. Data with strong, stable stratification during nighttime were removed using friction velocity thresholds examined at each site [M. Ueyama et al., in press, 2013].

[13] We applied consistent gap-filling and flux-partitioning methods [M. Ueyama et al., in press, 2013]. Data gaps were filled using the combination of a lookup table and nonlinear regression methods. A lookup table was created each day for 15 day moving window if photosynthetic photon flux density, air temperature, and vapor pressure deficit were available. If the lookup table method was not applicable, then the nonlinear regression method was applied; the nonrectangular hyperbolic functions were determined everyday using 15 day moving window for daytime fluxes, whereas the exponential relationships between air temperature and nighttime NEE (the so-called “Q10 model”) were determined everyday using a 29 day moving window for nighttime. Daytime RE was estimated by extrapolating the Q10 relationships into daytime. GPP was determined as the difference between RE and NEE. One limitation of this partitioning approach is that GPP and RE are derived from nighttime data. Because of limited availability of nighttime data in the Arctic, it should be cautioned that performance of the respiration models might be less accurate than that in other regions [M. Ueyama et al., in press, 2013] and might contain specious correlation among GPP and RE. In this study, negative and positive NEE indicate a CO2 sink and source, respectively.

2.3 Data

2.3.1 Satellite-Based Data

[14] We used four vegetation indices: normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), green ratio (GR) [Harazono et al., 2009], and leaf area index (LAI) and land surface temperature (LST) from the Moderate Resolution Imaging Spectroradiometer (MODIS) on board the Terra satellite for model development (collection 5; https://lpdaac.usgs.gov/). NDVI, EVI, and GR were created from MODIS surface reflectance products (MOD09A1) as an 8 day composite. Although an official MODIS vegetation index product is available with a 16 day composite, we preferred to be consistent with the composite period of other MODIS products. The MOD11A2 daytime LST [Wan et al., 2002] and MOD15A2 LAI [Myneni et al., 2002] were used as inputs. The spatial resolution of the original data is 1 km for the MOD15A2 and MOD11A2 and 500 m for the MOD09A1. For site-scale analyses, we selected the pixels centered on each site from the satellite data set. Poor quality data were removed using the MODIS state and quality flags attached in each MODIS product.

[15] For the spatial analyses, we created 1/30° spatial resolution data (NDVI, EVI, GR, LAI, and daytime LST) from the MOD09A1, MOD15A2, and MOD11A2 products. First, spatial averaging was conducted using the original spatial resolution by selecting good quality pixels (pixels with highest quality and no cloud and no cloud shadow conditions) in each 8 day period and at each 1/30° grid. Second, if an insufficient number of pixels were available (less than 25% pixels in a 1/30° grid), the corresponding pixel value was replaced by a mean seasonal variation created using good quality data from 2000 to 2011.

2.3.2 Gridded Climate Data

[16] We used the gridded solar radiation, air temperature, and precipitation climate data from the Japanese Re-Analysis 25 years (JRA25) reanalysis data set [Onogi et al., 2007]. This global reanalysis data set from JRA25 was a product of an assimilation of station- and satellite-based observations and presented a temporal resolution of 6 h and a spatial resolution of 1.1°. The original 6-hourly data were averaged in 8 day periods. For spatial analyses, we converted the original gridded data to 1/30° spatial resolution (approximately 2.5 km resolution) using the nearest neighbor method. Air temperature and precipitation of the JRA25 were only used for interpreting the regional upscaled fluxes, though solar radiation was used for model input for the regional-scale application, as observed photosynthetically active radiation (PAR) cannot be extended to the regional-scale application. For model tuning for the regional application, the solar radiation pixels centered on each station were used to input the model tuning, though observed PAR at the tower sites was used for site-scale analysis. We confirmed that 8 day JRA25 solar radiation was linearly correlated with observed solar radiation at the tower sites (R2 = 0.92, RMSE = 31 W m−2, n = 2186, slope = 1.05, and offset = 1.62 W m−2), indicating good correspondence between ground observations and the JRA25 solar radiation product.

2.3.3 Landcover and Fire Data

[17] We used landcover data (1.1 km grid spatial resolution) [Fleming, 1997] provided by the Alaska Geospatial Data Clearinghouse. Landcover data were regrouped to represent tundra, evergreen needleleaf forest, deciduous broadleaf forest, and fire scars (Figure 1). The spatial resolution of the landcover was adjusted to 1/30° spatial resolution using the nearest neighbor method.

[18] We added a new classification of fire scar in the landcover data because the relationship between CO2 fluxes and environmental and satellite observations on recently burned ecosystems differs from later successional ecosystems. In this study, fire scars were treated as any location where fires occurred within the last 10 years, based on fire location data from the Alaska Fire Service (http://agdc.usgs.gov/data/blm/fire/). Fire scars were annually updated, and after 10 years since fire, it was assumed that vegetation had recovered from fire [Fleming, 1997]. This recovery does not mean that vegetation always fully recovered within 10 years—only that the relationship between CO2 fluxes, environment, and satellite observations was different between recently burned ecosystems (≤10 years) and unburned ecosystems [M. Ueyama et al., in press, 2013]. The assignment of the 10 year fire scar was based on results showing that at sites approximately 10 years postfire, EVI and CO2 uptake during midday became comparable to those at unburned sites [Goulden et al., 2006].

2.3.4 Top-Down Estimate of CO2 Budget

[19] For the purposes of comparing estimates of the CO2 budget upscaled with the SVR model to independent data, we used a net atmosphere-biosphere flux from the CarbonTracker 2011 [Peters et al., 2007]. The CarbonTracker is a global inverse model with a nested atmospheric transport structure, and estimates daily CO2 flux with 1° × 1° spatial resolution. Biosphere and ocean fluxes are adjusted to match atmospheric CO2 observations. Biosphere CO2 fluxes are separated from the influences of the ocean uptake and fires and anthropogenic emissions. CO2 fluxes from fossil fuel burning and fire are prescribed from the existing database. A priori biosphere fluxes are from simulations by the Carnegie Ames Stanford Approach (CASA) model [van der Werf et al., 2006; Giglio et al., 2006]. Due to limited atmospheric observations, short-term variations in CarbonTracker primarily represent processes in the CASA model, though interannual variability is reflective of the atmospheric CO2 constraint [Desai et al., 2010]. Data from the CarbonTracker 2011 were available for the 2000 and 2010 period. We used the biosphere CO2 flux of the CarbonTracker 2011 by averaging land area of our study region at an 8 day time scale.

2.4 Empirical Upscaling Method: Support Vector Regression

[20] SVR is a machine-learning technique-based regression that transforms nonlinear regression into linear regressions by mapping the original low-dimensional input space to a higher-dimensional feature space with nonlinear mapping (kernel functions) [e.g., Cristianini and Shawe-Taylor, 2000]. The training process determines the parameter set of SVR by minimizing the generalization error (a combination of model complexity and training errors). Major advantages of the SVR are the training always converges to the global optimal solution, there are few free parameters to adjust, the architecture of the SVR does not need to be determined by experimentation, and SVR is robust to small errors in the training data. More details on SVR algorithm are described in Cristianini and Shawe-Taylor [2000], Zhan et al. [2003], and Yang et al. [2006].

[21] SVR analysis consists of three main steps for model tuning and testing. First, the SVR model parameters (constant determining the trade-off between the model complexity and the training error, width of an insensitive error band, and kernel parameter) were obtained from a training set using a grid search with a threefold cross-validation training process. In this process, training data are randomly divided into three nonoverlapping subsets. Training is performed 3 times on two of the subsets, with the remaining subset reserved for testing. Parameters yielding the lowest cross-validation errors are selected. Second, with the obtained parameters for the model structure, we trained the model. Finally, we evaluated the model based on a test set, using independent test data. As a kernel function of SVR, we chose the radial basic function kernel based on previous SVR studies [Yang et al., 2006, 2007]. These procedures are implemented by the LIBSVM software, a library for support vector machines [Chang and Lin, 2011].

[22] The SVR model has an advantage in unambiguously treating a nonlinear problem compared to multiple linear regression and artificial neural networks. Multiple linear regression is conceptually simple, but this method is not suited for nonlinear problems, such as estimating ecosystem CO2 fluxes. The performance of an artificial neural network is highly determined by the structure, such as the number of hidden layers and neurons in each hidden layer, and it is difficult to determine the most suitable structures from various choices. Detailed comparison among SVR, artificial neural networks, and multiple linear regressions had been conducted by Yang et al. [2006] for continental evapotranspiration estimation and Zhan et al. [2003] for ocean chlorophyll concentration estimation.

2.5 Analysis

2.5.1 Site-Scale Analysis

[23] As a first step toward regional estimates, we trained and tested this model on a site scale, as described above in section 2.4. Observed flux data were separated into two groups of training and test data. Test data were selected as 1 year data from sites that contained more than 2 years of data, amounting for 16 site years of data—2001 at BRW, CMS, and ATQ; 2006 at BEC and IVO; 2009 at IMS, IMH, IMT, ARS, ARM, and ARU; 2003 at FAI, DLS, DLA, and DLB; and 2010 at PFR. Training data were selected as all data that were not used for test data, amounting to 47 site years from all sites, and were used for model tuning. By using these flux data, a parameter set was determined for predicting the spatiotemporal dynamics of CO2 fluxes of all ecosystem types during the 12 years. For the model input, we used observed photosynthetically active radiation (PAR), landcover, and satellite-derived vegetation indices and LST.

[24] In order to select the best combination of input variables, we examined input variables of vegetation indices. First, we trained and tested the SVR model by changing input of vegetation indices (NDVI, EVI, GR, and LAI) of MODIS products and then examined the best input of vegetation indices using RMSE and the determinant coefficient (R2) for the test data.

[25] The importance of each input variable for GPP and RE estimation was also examined. The model was tuned and tested without one input among the vegetation indices, LST, PAR, and landcover. For example, the importance of LST was identified when the SVR model was tuned by inputs of vegetation indices, PAR, and landcover. RMSE and R2 were used to examine the importance of the inputs.

2.5.2 Regional-Scale Analyses

[26] GPP and RE were calculated throughout Alaska from 2000 to 2011 using the parameterized SVR model and spatially explicit climate and remote sensing products. These 12 year outputs were analyzed to clarify the spatial patterns, interannual variations, and trends during this decade of the terrestrial CO2 fluxes in the study region. Since observed PAR was not available at the regional scale, SVR was tuned with JRA25 solar radiation instead of observed PAR.

[27] For evaluating uncertainties associated with available data, sensitivity to the selection of eddy covariance data was examined with three different input data sets by eliminating data from original data set. First, a site-reducing experiment was conducted, in which input data were reduced to sites where more than 4 years of data were available for tundra (CMS, BEC, ATQ, and IVO) and more than 3 years were available for boreal (PFR, FAI, DLS, DLA, and DLB) ecosystems. Second, a year-reducing experiment was conducted in which only 2 years of data for each site were used for model tuning (including 1 year of data for training and 1 year for testing, though if only 1 year of data were available for a particular site, they were used as training data). A third experiment was conducted that consisted of excluding disturbed sites (ARS, ARM, CRF, PFR, and DLB) in the model tuning. Regional GPP, RE, and NEE were evaluated by the four SVR models (one original model and three models from the sensitivity analysis), after which uncertainties were estimated. Regional flux uncertainties were calculated with confidence intervals based on a t test (p = 0.01) among fluxes from the four models.

[28] Lastly, as an independent evaluation, the bottom-up (this study) and top-down (CarbonTracker) approaches were compared. The seasonal and interannual variations of biosphere flux (NEE) by the CarbonTracker were compared with NEE by the SVR model for the entire study region. Since variations in biosphere flux on a small spatial scale are not well constrained by atmospheric inversion [Peters et al., 2007], we only compared NEE results aggregated over the entire study region.

3 Results

3.1 Site-Scale Analyses

[29] SVR performance was best when inputting the visible band vegetation index of GR among vegetation indices in terms of R2 (Table 2). RMSEs of CO2 fluxes were also smaller when inputting GR than when inputting other vegetation indices, although inputting LAI resulted in similar performance for GPP (Table 2). Hereafter, GR was used for the vegetation index when calculating CO2 fluxes.

Table 2. Root-Mean-Square Error (RMSE) (g C m−2 d−1) and R2 for Various Vegetation Indices as Input
  EVINDVIGRLAI
RMSEGPP0.560.580.520.51
 RE0.420.330.270.36
 NEE0.480.540.440.55
R2GPP0.880.900.920.88
RE0.850.880.900.84
NEE0.670.640.750.64

[30] The SVR model successfully predicted the CO2 fluxes of ecosystems at 8 day time scale in Alaska (Figures 2 and 3). The predicted fluxes at the 8 day time scale were highly correlated to the observed fluxes for both the training and test data (RMSE = 0.52 g C m−2 d−1 and R2 = 0.92 for GPP; RMSE = 0.27 g C m−2 d−1 and R2 = 0.90 for RE; Figure 2). Model performance (R2 = 0.75) for NEE was lower than for GPP and RE (Figure 2). Seasonality of the observed CO2 fluxes was also well reproduced by the model (Figure 3).

Figure 2.

Comparison between the observed and predicted 8 day (a) GPP, (b) RE, and (c) NEE. The predicted NEE was calculated as RE minus GPP. The tuning and test data consisted of 21 and 16 sites, respectively.

Figure 3.

Comparison in seasonal variations of the observed and predicted 8 day average CO2 fluxes and the predicted fluxes for years of the model testing data. The SVR model was only applied when satellite-derived GR was available.

[31] The mean of predicted GPP and RE at each site was also correlated with that of the observations but that of NEE was slightly underestimated (Figure 4). The slope for the linear regression did not significantly differ from 1.0, nor did the intercept differ significantly from 0.0 g C m−2 d−1 (p < 0.01; p value was determined by a Student's t test), for both GPP (RMSE = 0.17 g C m−2 d−1, R2 = 0.94, n = 16, and p < 0.01) and RE (RMSE = 0.18 g C m−2 d−1, R2 = 0.81, n = 16, and p < 0.01), showing that spatial variations of GPP and RE were well captured by the model. Predicted NEE underestimated the spatial variations of observed NEE, even though the model explained 76% of this variation (Figure 4c). Based on a prediction interval (p < 0.01), the slope (1.18) differed significantly from a slope of 1.

Figure 4.

Comparison between the observed and predicted mean (a) GPP, (b) RE, and (c) NEE in the test data. The mean fluxes were calculated at each site for the period when both observed and predicted fluxes were available. Bold black and gray lines represent linear regression lines and their prediction interval (p < 0.05), respectively.

[32] SVR predicted the fluxes for the forest sites better than for tundra and burned sites (Table 3). R2 of GPP and RE were greater than 0.9 for all the forest sites. In contrast, R2 for the tundra sites were generally small and slope values were deviated from 1.0. Specifically, slope of RE for the BEC site showed negative value, highlighting the difficulty predicting RE of tundra. In terms of R2 and slope, SVR tended to better predict GPP than RE in each site.

Table 3. Root-Mean-Square Error (RMSE) (g C m−2 d−1), R2, and Slope of Linear Regression for the Test Dataa
Site GPPRENEE
RMSER2SlopeRMSER2SlopeRMSER2Slope
  1. aSince the slopes were calculated for SVR output as independent variable and for observations as dependent variable, slopes greater than 1.0 mean underestimates of SVR.
ForestFAI0.410.971.060.200.981.110.270.850.88
DLS0.360.950.970.250.961.160.390.780.85
DLA0.820.931.000.270.981.060.700.800.90
Average0.530.951.010.240.971.110.450.810.88
TundraCMS1.030.821.970.220.971.290.860.982.22
ATQ0.530.580.420.200.751.230.530.200.19
BRW0.290.951.130.230.691.220.160.930.87
BEC0.450.450.930.150.79−0.720.330.690.92
IVO0.440.871.060.230.831.370.330.640.67
IMH0.880.760.710.300.310.530.310.710.85
IMW0.920.811.110.450.811.370.390.741.02
IMT0.780.800.850.270.600.700.480.770.82
ARU0.550.400.610.200.220.980.660.010.07
Average0.650.720.980.250.660.890.450.630.85
BurnARS0.260.810.760.210.610.730.160.890.74
ARM0.170.981.040.320.821.280.310.840.90
PFR0.190.971.010.230.811.140.100.940.91
DLB0.260.880.990.240.901.250.130.640.69
Average0.220.910.950.250.791.100.180.830.81

[33] Among input variables, removal of landcover for GPP and GR for RE caused the largest reduction in performance of SVR (Table 4), indicating that these parameters are most important in predicting GPP and RE. GR was the second most important input for predicting GPP, and LST was the second most important input for predicting RE. Eliminating the input of PAR only slightly increased RMSE (Table 4), indicating that PAR was less important than other input variables for predicting seasonal and spatial variations in GPP and RE.

Table 4. Root Mean Square Error (RMSE) (g C m-2 d-1) and R2 for the Model Testing Without One Input Variable
  Discarded input 
  LSTPARGRLandcoverControl
RMSEGPP0.560.520.690.720.52
 RE0.430.330.570.340.27
 NEE0.800.560.530.740.44
R2GPP0.880.910.880.850.92
 RE0.850.860.770.870.90
 NEE0.890.730.790.540.75

[34] Using PAR or JRA25 solar radiation in the models had a small impact on the model performance. RMSE showed similar values when inputting JRA25 solar radiation (0.37 g C m−2 d−1 for GPP and 0.31 g C m−2 d−1 for RE) compared with when inputting observed PAR (0.52 g C m−2 d−1 for GPP and 0.27 g C m−2 d−1 for RE). The resulting NEE (RE–GPP) had a RMSE of 0.39 g C m−2 d−1.

3.2 Regional-Scale Analyses

3.2.1 Regional CO2 Fluxes

[35] Predicted total GPP averaged 284 g C m−2 yr−1 (369 Tg C yr−1) from 2000 to 2011, while that of RE was 282 g C m−2 yr−1 (362 Tg C yr−1). This suggests that terrestrial ecosystems acted approximately CO2 neutral (−2 g C m−2 yr−1 or −7 Tg C yr−1) for the study period. GPP and RE were high in the boreal forest and low in the Arctic tundra (Figure 5). GPP and RE from the tundra in the Arctic Coastal Plain were smaller than those of the high alpine tundra at the foothills of the Brooks Range (Figures 5a and 5b). Even though tundra type was not explicitly included in our model, a change in GPP and RE across the soil pH and moist/wet tundra types [Walker et al., 1998] was clearly shown (clear boundary around 70°N in Figures 5a and 5b). Tundra ecosystems of the Arctic Coastal Plain represented a larger CO2 source with smaller GPP but similar RE, compared to the tundra areas farther south in the foothills of the Brooks Range (Figures 5a and 5c). Consequently, spatial variation in the CO2 balance of the Arctic tundra was generally determined by GPP rather than RE. South of the Brooks Range, GPP and RE showed higher values that those north of the range (Figures 5a and 5b), due to a change in ecosystem type from arctic tundra to boreal forest (Figure 1). Boreal forest in interior Alaska acted as a CO2 sink, while the Arctic tundra acted as a CO2 source (Figure 5c). From 2000 to 2011, average GPP, RE, and NEE were 383, 335, and −48 g C m−2 yr−1, respectively, for the boreal forest region of Alaska.

Figure 5.

Spatial distributions of mean annual (a) GPP, (b) RE, and (c) NEE between 2000 and 2011 estimated from the SVR (kg C m−2 yr−1), and linear trends (g C m−2 yr−2) in (d) GPP, (e) RE, and (f) NEE during the period. Only the grids with statistically significant trends (p < 0.05) are only colored in the map.

[36] Predicted CO2 fluxes reproduced the observed variations associated with ecosystem types in the boreal region [Welp et al., 2007; Iwata et al., 2011]. Averaging regional fluxes for each ecosystem type, GPP of deciduous broadleaf forests (418 g C m−2 yr−1) was larger than that of evergreen needleleaf forests (409 g C m−2 yr−1) and fire scars (211 g C m−2 yr−1). In addition, CO2 sink strength (negative NEE) was greater in deciduous broadleaf forests (103 g C m−2 yr−1) than that in evergreen needleleaf forests (59 g C m−2 yr−1), and fire scars acted as a CO2 source (76 g C m−2 yr−1).

[37] Based on our sensitivity analyses, uncertainties related to selection of observed data were 6 g C m−2 yr−1 for GPP, 56 g C m−2 yr−1 for RE, and 54 g C m−2 yr−1 for NEE. We found that GPP did not differ substantially among the four models, indicating that eliminating sites or the use of short-term data did not substantially influence regional GPP. In contrast to GPP, the uncertainties in RE were large, where eliminating sites for short measurement periods overestimated RE by 59 g C m−2 yr−1 and decreasing years at each site for tuning underestimated RE by 48 g C m−2 yr−1. The large uncertainty in RE was the main reason for the large uncertainty in NEE. Eliminating the fire-disturbed sites did not substantially influence the regional fluxes, with an increase of the total regional CO2 sink of only 3 g C m−2 yr−1, as annually burned areas were smaller than the extent of nonburned ecosystems (Figure 1). The fire scars occupied total land ranging from 3.3% to 6.7%.

3.2.2 Interannual Variations in Regional CO2 Fluxes

[38] The predicted CO2 fluxes of the entire Alaska region showed interannual variability (Figure 6a) but with no statistically significant trends in CO2 fluxes from 2000 to 2011. GPP and RE were positively correlated (R2 = 0.87; p < 0.01), and both fluxes showed peaks in the warm years of 2004 and 2007. Annual GPP and RE were positively correlated with air temperature between June and August with R2 = 0.90 (p < 0.01) for GPP and R2 = 0.93 (p < 0.01) for RE (Figure 6d), indicating more GPP and RE in warmer years and less GPP and RE in cooler years. The higher values of GPP and RE in the warmer years were associated with higher regional mean GR (Figure 6f), where growing-season mean regional GR and air temperature tended to show positive correlation (R2 = 0.28; p = 0.08). Annual precipitation was not correlated with GPP and RE (R2 = 0.00), although precipitation between June and August showed a weak negative correlation with GPP (R2 = 0.37; p = 0.04) and RE (R2 = 0.24; p = 0.11; Figure 6e). Regional NEE was negatively correlated with GPP (R2 = 0.85; p < 0.01) and represented a greater CO2 sink in warmer years, and vice versa (Figure 6b). Interannual variations were evident, even after considering uncertainties associated with the sensitivity studies for site selection (vertical bars in Figures 6a and 6b). This indicates that the examined interannual variations were robust, even after considering uncertainties associated with the choice of available data.

Figure 6.

Interannual variation of (a) annual GPP and RE predicted by the SVR, (b) annual and (c) growing season (DOY 145–281) NEE predicted by the SVR and CarbonTracker 2011, (d) air temperature between June and August, (e) precipitation, and (f) GR between June and August for the entire Alaska region. The air temperature and precipitation were derived from the JRA25 reanalysis data for the study region. The vertical bars in GPP, RE, and NEE are uncertainties in interannual variation, estimated as the confidence interval (p = 0.05) of anomalies in the fluxes from the sensitivity analysis (shown in section 2.5.2).

3.2.3 Trends in a Decade of CO2 Fluxes

[39] The total CO2 flux for the entire region did not present any significant trends over the 12 years (2000–2011), but considerable spatial variation in the temporal trend of annual CO2 fluxes was estimated (Figure 5). In part of the undisturbed Arctic tundra area near the northern foothills of the Brooks Range, annual GPP showed positive trend from 2000 to 2011 (shown in Figures 5d–5f, with an aggregated trend for tundra in Figure 7). The aggregated trend was defined as the mean of trends at grids that had statistically significant trends (p < 0.05). In contrast, most of the undisturbed boreal ecosystem had no significant (p > 0.05) trend during the study period (Figures 5d–5f).

Figure 7.

Relationship between stand age at 2011 and aggregated trend in (a) GPP, (b) RE, and (c) NEE between 2000 and 2011. Points and vertical bars represent mean and standard deviation of trends. Cumulative ratio (%) of data point at each stand age to total disturbed area was also shown as a line. Grids with statistically insignificant trends (p > 0.05) were excluded in this analysis, and average and standard deviation were only calculated if sampling size for averaging was greater than 10. The average and standard deviation of trends in nondisturbed arctic tundra and boreal forest were also calculated.

[40] In contrast to nondisturbed ecosystems, disturbed ecosystems showed clear positive or negative trends in annual fluxes (Figures 5d–5f). The direction and magnitude of the aggregated trends were dependent on the number of years postfire (Figure 7). Based on our assumption of a 10 year recovery, drastic changes in aggregated trends were shown during the first 10 years after fires. During this initial period (for grids where fire occurred from 2003 to 2011), disturbed ecosystems presented significant (p < 0.05) negative aggregated trends in GPP (−20 g C m−2 yr−2) and RE (−6 g C m−2 yr−2) (Figures 7a and 7b), because fires at this period caused decrease in the fluxes. Following the decrease in GPP associated with recent disturbances, NEE showed an increasing aggregated trend (15 g C m−2 yr−2) (Figure 7c). For ecosystems burned between 1982 and 2002 (age 10–20 years), GPP showed a positive aggregated trend (20 g C m−2 yr−2), due to estimated recovery (Figure 7a), and resulted in an increasing aggregated trend in the CO2 sink strength (−15 g C m−2 yr−2) (Figure 7c). For ecosystems burned more than 20 years ago, positive aggregated trends in GPP and RE tended to gradually decrease, indicating that recovery of the fluxes tended to decrease as the forests matured. Even though recently burned ecosystems showed significant trends in CO2 fluxes, most of ecosystems were not affected by fire. As a result, there were no significant trends (p > 0.05) in the regional mean of the CO2 balance during the last 12 years (Figure 6).

3.2.4 Comparison Between Bottom-Up and Top-Down Approaches

[41] NEE by bottom-up (our SVR model) and top-down (CarbonTracker) approaches showed similar magnitude of the annual (Figure 6b) and growing season (Figure 6c) CO2 sink. The total CO2 sink between day of year (DOY) 145 and 281 was 86 Tg C (64 g C m−2 season−1) according to the SVR model and 87 Tg C (65 g C m−2 season−1) from the CarbonTracker during the 2000 and 2010 period. Seasonal variation from both approaches showed similar patterns (Figure 8), though the SVR summer peak was smaller than that of the CarbonTracker.

Figure 8.

Seasonal variations in 8 day NEE by the SVR and the CarbonTracker 2011 for the entire Alaska region.

[42] The interannual variations of annual and growing season CO2 fluxes were inconsistent among those approaches. The interannual variation (standard deviation of annual CO2 flux) was underestimated by the SVR and/or overestimated by the CarbonTracker. The 2009 variations were anomalous, due to a large uptake estimated by the CarbonTracker during mid-July (Figure 8). The annual NEE by both methods showed inconsistent interannual variation, even excluding the anomalous 2009 (RMSE = 12 g C m−2 season−1, R2 = 0.20, and p = 0.19) (Figure 6b). The cumulative fluxes between DOY 145 and 281 also showed inconsistent interannual variation (RMSE = 14 g C m−2 season−1, R2 = 0.27, and p = 0.11; Figure 6c). Even though we could not confirm either estimate as the more accurate, observations in 2009 did not support the anomalous uptake estimated by the CarbonTracker for both tundra (IMH, IMT, IMW, and ARU) and boreal (FAI and PFR) regions.

4 Discussion

4.1 Upscaling CO2 Fluxes in the High-Latitude Region

[43] The upscaling technique was evaluated at the site and regional scales. Upscaled eddy covariance-derived CO2 fluxes from 21 towers within Alaska were fairly consistent with regional CO2 fluxes from CarbonTracker in terms of magnitude of the annual and growing season NEE. The consistency between the bottom-up and top-down approaches indicated that our upscaling technique was valid at this regional scale. In contrast, interannual variations from both methods were inconsistent. One possibility of this discrepancy was that inversion models was not well constrained in northern high-latitude regions; different inversion models estimated totally different interannual variation of high-latitude CO2 balance [McGuire et al., 2012]. Further comparison to results at large spatial scales using multiple inversion models would be helpful.

[44] According to a sensitivity analysis that reduced the number of input sites or site years, the choice of data within the current eddy covariance network in Alaska did not substantially change the upscaled GPP, indicating that the regional estimate of GPP and its interannual variation could be robust in the context of even further availability of long-term data measured at similar ecosystem types. However, uncertainties associated with the choice of available data were relatively large in RE estimation, which propagated to uncertainties in NEE. The large uncertainties in RE indicate that RE was more heterogeneous than GPP and possibly was not well constrained by the observations and/or input variables (discussed in later in this section).

[45] The satellite-derived visible reflectance index of GR was the most important input to predict CO2 fluxes in Alaska. Since productivities of Arctic ecosystems are heavily regulated by LAI [Williams and Rastetter, 1999; Street et al., 2007], GR could be a good index to estimate LAI. NDVI uses red and infrared reflectance, which could change with background conditions, such as soil; consequently, LAI-NDVI relationships differed in each vegetation type [Street et al., 2007]. In contrast, GR is only sensitive to green reflectance whose changes reflect change in productive portion within an ecosystem. The high sensitivity to GR was probably due to small LAI of Arctic ecosystems [Williams and Rastetter, 1999; McFadden et al., 2003; Lund et al., 2010]. Due to the small LAI, GR could reflect development of productive portion of Arctic ecosystems, even though GR could be saturated for vegetations that have complex structures with high LAI. Changes in surface conditions that control CO2 fluxes can be easily detected by a change in the green portion. For example, many land surfaces (e.g., snow, soil, autumn color of leaves, and defoliated trees) have a similar minimum value of GR, which helped to distinguish the vegetation growing period from the dormant period [Ide et al., 2011]. Recently, the advantage of GR in predicting seasonal and interannual variations of GPP was also reported in several ecosystems [Ahrends et al., 2009; Harazono et al., 2009; Ide et al., 2011]. Further site-level studies using GR could improve CO2 flux estimation in Arctic ecosystems.

[46] Landcover classification was the most important input for predicting GPP in Alaska. This was because characteristics of CO2 fluxes were different among tundra and boreal ecosystems under given environmental variables, such as NDVI, EVI, LAI, and air temperature [M. Ueyama et al., in press, 2013]. In contrast to GPP, GR and LST were quite important in predicting RE, indicating that variations in RE were mostly associated with temperature-related conditions, including phenology that was controlled by temperature dynamics. This was consistent with a previous synthesis that growing degree days could explain spatial variation of RE, rather than GPP, among Alaska's ecosystems [M. Ueyama et al., in press, 2013].

[47] Modeling CO2 fluxes in heterogeneous landscapes is challenging. Stoy et al. [2009] examined coarse spatial resolution of NDVI that may bias the results of upscaled CO2 fluxes in an Arctic ecosystem. Geographic distortion of high-latitude MODIS data [Ji et al., 2010] could limit accurate SVR learning at the site scale. Since small-scale microtopography is a major characteristic of the Arctic landscape [Goswami et al., 2011; Zona et al., 2011], the role of heterogeneous landscape to CO2 fluxes must be considered in future.

[48] Including explicit water status will be required to accurately estimate of CO2 fluxes in the Arctic. Although vegetation indices could capture the drought stress of vegetation [Verbyla, 2008; Zhao and Running, 2010; Gamon et al., 2013], RE in the Arctic has been shown to be strongly dependent on water table depth [Vourlitis et al., 2000; Kwon et al., 2006; Huemmrich et al., 2010]. Spatially explicit water table depth is currently not available in the Arctic, because water table depth is highly heterogeneous across the landscape and dependent on microtopography [Goswami et al., 2011; Gamon et al., 2013].

[49] It should be cautioned that the current network of eddy covariance towers in Alaska could underrepresent fluxes of certain ecosystem types, because of missing data on important ecosystem types in Alaska (e.g., white spruce and birch forest, various aged ecosystems after fire, wetlands, bogs, and various types of shrubland). Since only one white spruce forest, a drought-susceptible ecosystem [Barber et al., 2000], was included in this analysis, drought sensitivity may have been underestimated. According to the site-scale analysis, SVR could poorly reproduce fluxes, especially RE, in the tundra sites. Moreover, further observations will be required at tundra. Further data availability of various aged ecosystems from chronosequence studies [Amiro et al., 2010; Goulden et al., 2010] could improve the accuracy of age-dependent change in fluxes.

[50] Intercomparison among various empirical upscaling methods is needed in the future to quantify uncertainties in the estimation caused by different algorithms and input variables. Various machine-learning algorithms have been applied to estimate continental or global carbon and water fluxes. For example, Jung et al. [2009] applied a model tree ensemble algorithm with more than 20 input parameters to upscale observed CO2 fluxes in FLUXNET. Xiao et al. [2010] applied a regression tree approach with several satellite-based input parameters. Papale and Valentini [2003] applied an artificial neural network algorithm with 12 input variables to estimate CO2 budget in Europe.

4.2 Mechanisms of Year-to-Year Variations in CO2 Fluxes

[51] Even though regional total fluxes showed no significant trends, they presented considerable spatial variation. A slight but significantly increasing trend in GPP in the tundra near the foothills of the Brookes Range was observed, which is consistent with previous studies using satellite remote sensing data of the past 20–30 years [e.g., Goetz et al., 2006; Verbyla, 2008]. This increasing trend in GPP has been linked to enhanced productivity in tundra ecosystems, which may be associated with the change in species composition in the mountainous shrub tundra [Hinzman et al., 2005]. Positive GPP trends in coastal tundra were rather sporadic because vegetation changes in coastal tundra associated with warming were different for different plant communities [Villarreal et al., 2012]. Since year-to-year variations were larger than trends in the decades at nondisturbed boreal forests, trends were only obvious in recently burned areas (discussed in section 4.3).

[52] Air temperature was a dominant factor in the control of the interannual variation of regional-scale CO2 fluxes. This was consistent with the field observations in Alaska [Harazono et al., 2003; Welp et al., 2007]. High sensitivity of CO2 flux to air temperature in a high-latitude ecosystem is consistent with previous satellite-based estimates using a light-use efficiency model input with MODIS data [Zhao and Running, 2010].

[53] Upscaled CO2 fluxes did not detect decreases in vegetation productivity associated with warming-induced drought [e.g., Goetz et al., 2006]. During the study period, annual precipitation by JRA25 significantly decreased at a rate of −7.9 mm yr−1 (p = 0.02) and this rate was greater than that between 1981 and 2011 (−1.7 mm yr−1; p = 0.10; Figure 6e). Despite the drying conditions, drought-related changes in CO2 flux were not apparent, even in the warm years. Satellite-derived GR showed the highest values in 2004, when warm and dry conditions were pronounced [Welp et al., 2007], and resulted further in the highest estimated GPP. Low sensitivity to water conditions was consistent with field observations, where CO2 fluxes and evapotranspiration in black spruce forests were not strongly affected by drought [Welp et al., 2007; Iwata et al., 2012].

[54] The contrasting results for drought sensitivity for the previous satellite analyses [e.g., Goetz et al., 2006; Verbyla, 2008] were caused by several mechanisms. Previous satellite analyses were conducted for the period between 1981 and 2003 using the advance high-resolution radiometer (AVHRR) data, whereas our study was conducted for period between 2000 and 2011 using the MODIS data. Drought sensitivity might be masked for year-to-year variations over the short-term period. The insufficient atmospheric correction of the AVHRR sensor was known to induce artificial negative trends in NDVI at boreal North America [Alcaraz-Segura et al., 2010], which might cause part of previously examined negative trends. Analyses using multiple satellite data will help to understand how drought influenced the regional productivity of boreal forests.

4.3 Role of Disturbance on Decadal CO2 Flux

[55] Past disturbance history greatly influenced trends in interior Alaska boreal forest CO2 fluxes in our 2000–2011 study period, although most nondisturbed forests presented no significant trend in CO2 fluxes during the study period. In half of the disturbed ecosystems, GPP, RE, and CO2 uptake showed decreasing trends associated with initial fire disturbance, while the rest showed an increasing trend due to recovery. Negative trends in CO2 fluxes associated with fire were alleviated 5 to 10 years after fire (Figure 7), perhaps caused by dominating grasses and shrubs at initial stages of recovery [e.g., Chapin et al., 2006]. The gradual increase in GPP from 5 to 10 years and from 10 to 15 years indicated that actual initial vegetation recovery could be completed over this period. The age-dependent change in the aggregated trend also indicated that a second stage of recovery in CO2 fluxes would be complete 20 years postfire (Figure 7), and then the GPP and CO2 uptake could be stable afterward from 20 to 60 years after fire (Figure 7). This is consistent with previous chronosequence studies [Amiro et al., 2010; Goulden et al., 2010]. Since the highest average area (767,000 ha yr−1 or approximately 0.5% of total land in Alaska) occurred in the 2000s [Kasischke et al., 2010], CO2 fluxes and sequestration may increase as a result of recovery in the coming decades. However, the long-term trends for boreal forest CO2 fluxes will strongly depend on changes in the fire regime [Kasischke et al., 2000].

[56] New fire scar landcover was introduced in order to characterize CO2 flux at disturbed ecosystems with the satellite-derived products. This classification was based on a simple assumption that relationships between CO2 fluxes and environmental variables showed a step-like change after a decade. This assumption generates major uncertainties in the prediction of fluxes over disturbed ecosystems, because the relationship between CO2 fluxes and environmental variables could gradually change as an ecosystem recovers. At present, the data availability from fire-disturbed ecosystems are limited in Alaska, and thus, a sufficient number of data points for incorporating stand age into the SVR model are not available.

[57] Interannual variations of the regional CO2 fluxes were less influenced by fire than meteorology (Figure 6). This was consistent with a previous estimate of fire impact at a Canadian boreal forest [Bond-Lamberty et al., 2007]. They pointed out that the fire effect, which included associated changes in ecosystem composition, was the major driver of the carbon balance at a timescale greater than several decades, whereas the interannual variations were mostly caused by meteorology [Bond-Lamberty et al., 2007, Figure 1]. Since such long-term effects were shown in our study (Figure 7), the importance of fire to the regional CO2 balance will appear at longer time scales (more than several decades).

5 Conclusion

[58] In this study, we upscaled GPP, RE, and NEE in Alaska with the SVR model by inputting multisite data from a network of 21 eddy covariance towers, satellite remote sensing data, and disturbance information. Based on our model predictions, the satellite-derived visible reflectance index of GR was the most important variable in predicting CO2 fluxes, indicating that accurate measurements of vegetation phenology are important to calculate spatially explicit regional estimates of high-latitude CO2 fluxes based on the methodology presented here. The upscaled regional CO2 sink was greater during warmer growing seasons, and changes in CO2 fluxes due to water stress were not detected, indicating that expected high-latitude warming might increase the CO2 sink of Alaska ecosystems. During the study period between 2000 and 2011, the regional totals of GPP, RE, and NEE had no statistically significant trend, despite considerable spatial variation within the trends. Disturbance history was the most important determinant of trends in the CO2 fluxes at each location during the study period.

[59] Predicted regional CO2 fluxes were consistent with field observations, and its magnitude was also consistent with a top-down approach from the CarbonTracker. The good correspondence between these two methods indicates that our method was valid for estimating regional fluxes. Since the interannual variations between the two approaches were inconsistent, further validation data including CO2 fluxes from multiple inversion models will be required. The predicted distributions of fluxes could be useful for validating, parameterizing, and optimizing process-based models at a regional scale. The upscaled fluxes provide independent validation data for top-down approaches, as well as data for downscaling the results of top-down approaches.

Acknowledgments

[60] We are grateful to James T. Randerson of University of California, Irvine, for providing us the eddy covariance data measured at three sites in Delta Junction, Alaska. The eddy flux data for the three sites at Council on the Seward Peninsula were recorded by J. Beringer of Monash University, and F. S. Chapin III and C. Copass of University of Alaska Fairbanks, and provided through the NCAR Earth Observing Laboratory. We thank all individuals who contributed field observations and data analysis, and N. Bauer of University of Alaska Fairbanks for editing the manuscript. We appreciate comments by D. Baldocchi of University of California, Berkeley; J. Gamon of University of Alberta; and an anonymous reviewer. CarbonTracker 2011 results were provided by NOAA ESRL, Boulder, Colorado, USA, from the website at http://carbontracker.noaa.gov. This study was partly supported by the Environment Research and Technology Development Fund (RF-1007 and RFa-1201) of the Ministry of the Environment, Japan. Data provided by NCAR/EOL are under sponsorship of the National Science Foundation (http://data.eol.ucar.edu/). D. Zona was funded by the National Science Foundation (ARC, 1204263). A. Rocha was funded by the National Science Foundation (EF, 1065587). Funding for the Imnavait Creek eddy flux towers was provided by the NSF Office of Polar Programs.

Ancillary