4.1. Ring Width and NDVI Correlations
 Recent efforts to link NDVI with ground-based measurements of vegetation productivity have met with mixed success. In Scandinavia, mean monthly MODIS NDVI and flux tower GPP showed moderate correlations (r = 0.7–0.79) at seven forested sites; however, NDVI saturation during periods of high productivity (NDVI > 0.9) was a noticeable issue [Olofsson et al., 2007]. At three flux tower sites located in Southeast Asian tropical forests, Huete et al.  found that the relationship between NDVI and gross ecosystem production varied considerably with forest type (r2 = 0.00–0.53). Satellite vegetation indices, both NDVI and the enhanced vegetation index (EVI), show stronger associations with tower measurements in forests with seasonal, rather than evergreen, canopy cover [Olofsson et al., 2007; Huete et al., 2008]. NDVI and EVI are related to productivity via light absorption, though differences in canopy phenology will affect the degree to which light absorption is biochemically decoupled from utilization for carbon assimilation [Goetz and Prince, 1996]. Additionally, decoupling of light absorption and utilization occurs as a result of differences in response time to short-term weather events, with GPP fluctuating rapidly and NDVI responding much more slowly [La Puma et al., 2007]. Ephemeral cloud cover not adequately removed during image processing can also depress the relationship between satellite and ground-based vegetation measurements [Nagai et al., 2010]. Recently, considerable strides have been made in validating satellite productivity measurements using flux tower data; however, validation requires a multifaceted approach to overcome the limitations of each technique [Running et al., 2004].
 Past research has demonstrated that for areas above 40°N, ring width and mean monthly NDVI are positively related from May through July [Kaufmann et al., 2008]; however, the strongest correlations occur during June and July when NDVI is at its highest [Kaufmann et al., 2004]. Our use of a pixel-based, growing season NDVI capitalized on these previous findings and yielded stronger correlations between NDVI and RWI than did an analysis using NDVI during June or July, or when NDVI was summed over the June through August period (data not shown). This suggests that our definition of the NDVI growing season helped alleviate the issue of latitudinal and topographical differences among sites in the timing of maximum summer canopy development, which can be quite high in boreal regions [Beck et al., 2007]. However, as the growing season length was derived from post-2000 MODIS data and held invariant between years, relationships with NDVI driven by annual changes in growing season length will not be captured.
 While NDVI and ring width are positively related during the growing season, the Granger causality analysis provided little evidence of G-causal relationships between the two variables at a one year lag. The lack of G-causality implies that neither metric of tree productivity can be used to forecast the other. This finding is consistent with that of Kaufmann et al. , who used Granger causality tests to investigate the relationship between NDVI and tree rings for 48 sites located at middle to high latitudes in North America and Eurasia. The researchers found no G-causal relationship between NDVI (June and July) and ring width over the 1981 to 1999 period, concluding that the two variables instead share a common causal source, potentially NPP. The lack of G-causality suggests that there is a mediating factor between the various elements of the organism that fix and store C.
 The magnitude of the RWI-NDVI correlations we present are similar to those reported by Lopatin et al. , who examined the growth of Picea abies subsp. obovata and Pinus sylvestris at five mixed-species sites in northwestern Russia. Both studies, however, report much lower correlations than those reported by Wang et al. , who found very strong relationships between NDVI and ring width, seed production, litter fall and foliar biomass in three oak forests (Quercus spp.) in Kansas, U.S.A. Differences in land cover, landscape topography, and canopy architecture may partially account for this discrepancy. Conifers, larch in particular, tend to have lower crown width:height ratios than broadleaf deciduous trees [Gower and Richards, 1990], thus reducing their radiometric influence on satellite imagery acquired from an overhead perspective. Additionally, Wang et al.  used the 1.1 km AVHRR Local Area Coverage data whereas both our work and that of Lopatin et al.  used 8 km GIMMS data which were derived from the ∼4 km Global Area Coverage data. One P. banksiana chronology from a mixed-species site displayed a negative correlation with NDVI, which may have been due to interspecific competition within the stands, or an exogenous factor (e.g., insects). These studies suggest that time-integrated NDVI can reflect interannual variability in forest productivity, although they highlight that there are technical and physiological limitations to such an approach.
 Many of the technical limitations of relating tree growth and satellite-derived productivity estimates are related to differences in the scale at which each technique senses growth. Tree growth can vary significantly among individuals within a small area as a result of microhabitat variability and species-specific differences. Tree ring data are highly sensitive to this among-tree variability and it is possible that the extent of sampling at each site did not adequately capture the range in tree growth variability that occurred within the NDVI window. Imagery with coarse spatial resolution will thus represent a blend of potentially contrasting tree growth patterns. In mixed-species stands, this might obscure relations that would otherwise be apparent in single species stands [Lopatin et al., 2006].
 Coarsely resolved satellite imagery necessarily blends different plant functional types into a single productivity value, potentially obscuring the effects of any one functional type (e.g., trees). This is likely to be particularly important in sparsely forested areas, where the radiometric influence of understory plants is more pronounced [Goetz and Prince, 1996; Rees et al., 2002]. Interestingly, we found no relationship between the extent of forest cover represented by the tree ring measurements (or forest cover in general) and the strength of the NDVI and ring width correlation. This finding might reflect difficulties in satellite mapping of tree cover near the tundra-taiga transition, where shrubs and trees can be difficult to distinguish [Montesano et al., 2009]. Alternatively, D'Arrigo et al.  suggested that the relationship between NDVI and tree rings might hold when the sampled species covers only a small proportion of the landscape if tree growth and regional productivity are limited by the same environmental influences. The heterogeneous landscape, a complex mosaic of lakes, bogs, forest fire remnants, shrubs and mixed-species forests of varying tree density and age structure, hinders the ability of NDVI at a coarse spatial resolution to precisely reflect growth patterns of individual trees or even the spatially averaged growth of individuals within a stand. The influence of land cover on efforts to link satellite and ground-based measurements of productivity warrants further attention.
 In addition to technical complications, there are a number of physiological factors that contribute to the uncertainty in relating C accumulation, as measured by tree ring widths, to satellite-derived estimates of productivity, which generally measure C uptake (GPP). Trees experience interannual variability in C allocation among organs and in production of secondary plant compounds. The proportion of GPP allocated to above ground wood production varies depending on climatic conditions, nutrient availability, and stand age [Gower et al., 1994; Litton et al., 2007], and the relative allocation to foliage and wood has been shown to vary over time as a function of climatic variability [Lapenis et al., 2005]. Trees throughout the northern high latitudes also experience trade-offs between reproduction and biomass accumulation, with years of high seed production generally characterized by reduced ring widths [Koenig and Knops, 1998; Selas et al., 2002]. Interannual variability in C allocation thus complicates the use of tree rings in assessing past C uptake.
 While C allocation pathways change over time, in boreal forest ecosystems annual wood growth is closely related to NPP [Gower et al., 2001] and there is considerable evidence that across forest ecosystems annual wood growth strongly reflects GPP [Litton et al., 2007]. Nonetheless, the relationship between ring width and ecosystem C exchange is both complex and poorly understood. Rocha et al.  found that, over a ten year period in Manitoba, Canada, the ring widths of P. mariana were uncorrelated with estimates of ecosystem C uptake (gross ecosystem productivity or net ecosystem exchange) derived from flux towers. However, Rocha et al.  found that ring widths were positively correlated with net ecosystem productivity (NEP; r = 0.854) and negatively correlated with two estimates of respiration (r = −0.733 and −0.779). The researchers proposed two possible explanations for the lack of correlation between C uptake and ring width. First, they propose that ring width may be controlled by a factor other than C uptake. Second, they suggest that tree ring growth may rely heavily upon carbohydrates stored during previous years, and that these may effectively buffer ring width against interannual variability in C uptake. The very strong relationship between annual wood production and GPP shown by Litton et al.  argues against ring width being controlled by a factor other than carbon uptake. Our results, including the low to moderate NDVI-RWI correlations, as well as the similarity in autocorrelation patterns between NDVI and ring width, support the idea that ring width is controlled by a variety of factors, potentially spanning many years, and not just by rates of photosynthesis or canopy development during a single season.
4.2. Ring Width and NDVI Autocorrelation
 Tree growth during a growing season is affected by sugars, buds, leaves, hormones and other plant compounds synthesized during previous years [Fritts, 2001]. This year to year carryover imparts a degree of persistence in tree growth patterns, with potential manifestation being statistical autocorrelation in ring width and canopy measurements [Fritts, 2001]. While autocorrelation in ring width measurements has been relatively well studied [Fritts, 2001; Monserud and Marshall, 2001; Kagawa et al., 2006a, 2006b], autocorrelation in NDVI measurements from forested regions has not been widely examined. Over the 27 year study period, we observed distinct differences in both NDVI and ring width autocorrelation patterns among larch, on one hand, and spruce and pine on the other. Larch, a deciduous needleleaf conifer, displayed little autocorrelation in either measurement at a lag of one year, while spruce and pine, both evergreen conifers, exhibited significant autocorrelation at lags of one and two years. The differences among these species in both NDVI and ring width autocorrelation may relate to needle retention length. Monserud and Marshall  observed differences in ring width autocorrelation associated with needle retention length and hypothesized that evergreens would exhibit higher autocorrelation than deciduous species, while evergreen species with longer needle retention would exhibit the highest autocorrelation.
 In spruce and pine, phenotypic plasticity permits intraspecific needle longevity to vary widely across latitudinal and elevational gradients, though retention lengths are generally three to ten years [Reich et al., 1996]. Larch, on the other hand, tend to replace foliage yearly [Gower et al., 1993]. Studies using stable C isotopes show that both functional groups store photoassimilates and use them in following years for xylem and needle production [Jäggi et al., 2002; Monserud and Marshall, 2001; Kagawa et al., 2006a, 2006b]. Annual replacements of larch foliage draws upon C stored during previous years and reduces the amount that can be allocated to structural growth [Kagawa et al., 2006a]. Evergreen conifers, on the other hand, have a lower foliage replacement demand and can use stored C to buffer xylem production to a greater extent during difficult growing conditions. In spruce and pine, lower interannual variability in xylem production may lead to higher ring width autocorrelation in comparison to larch, which exhibit higher interannual variability in ring growth and correspondingly lower autocorrelation strength.
 We propose that interannual persistence in canopy extent, leaf structure and chemical quality leads to autocorrelation in NDVI measured over spruce and pine stands, while for larch, annual foliage replacements leads to higher interannual variability canopy and leaf characteristics, thus limiting the NDVI autocorrelation strength. Across sites, we observed a significant, positive association between the strength of NDVI and ring width autocorrelation at a one year lag and believe that this reflects the important role of needle retention in buffering year to year xylem growth. Due to relatively high shrub cover at many of the larch sites, a large unknown is the degree to which the NDVI measurements from those stands reflected the growth dynamics of nonlarch plants. This is an area that should be investigated further.
4.3. Ring Width and NDVI Trends
 Although the annual correlations between NDVI and RWI were relatively weak, we believe this largely reflects strong autocorrelation, land cover heterogeneity, interannual variability in carbon allocation, and that NDVI and RWI measure different aspects of carbon exchange (uptake and accumulation). We view comparison of time series trends as a means of examining the relationship over a decadal period and not just on an annual basis. Negative trends in both RWI and NDVI were primarily concentrated in Canada at sites with tree cover >30%, while positive trends in NDVI occurred almost exclusively in sparse larch forests of northeast Siberia. The spatiality of NDVI trends is similar to those reported by Bunn and Goetz , who found that across the northern high latitudes browning was most prevalent in areas of dense tree cover, while greening was common in sparsely forested areas near the tundra. There was a high consistency between NDVI and RWI trends for sites with either no trend or a negative trend in NDVI: 5 of 8 (62.5%) of sites with no trend in NDVI also showed no trend in RWI, while 5 of 7 (71%) of sites with a negative trend in NDVI also exhibited a negative trend in ring width. The biggest discrepancy between NDVI and RWI data occurred in those sites with a positive trend in NDVI (n = 7). None of those sites exhibited a positive trend in RWI: 6 of 7 had no significant trend in RWI, while one experienced a positive trend. The positively trending NDVI sites were predominantly in areas with sparse tree cover, where we would expect correlations with RWI to be weakest. Greening observed in these stands may not reflect enhanced tree growth so much as it reflects enhanced understory shrub growth, as has been observed throughout the pan-Arctic [Bunn and Goetz, 2006; Tape et al., 2006; Verbyla, 2008; Forbes et al., 2010].
 Most sites (n = 13, 59%) showed no trend in ring width over the study period. Nontrending tree ring chronologies were most abundant in larch (n = 7) and mixed-species stands (n = 4). Spruce exhibited the highest frequency of negative trends in both NDVI and ring width, though negative trends in both measurements were also common in pine. In a circumpolar analysis of tree ring data, inverse relationships between ring width and temperature were most prevalent in these two taxa [Lloyd and Bunn, 2007]. Negative trends in NDVI at some sites may reflect inverse relationships between temperature and tree growth. Ring width at one spruce (P. abies) stand exhibited a positive trend since 1981 and was a located along the Lena River in central Russia (∼68°N). Growth at this site was positively correlated with growing season temperature and negatively correlated with precipitation over the 1902 and 2002 period [Lloyd et al., 2010]. Examining within population responses of tree ring growth to climatic conditions at seven sites along the Lena River (includes sites presented in this study), Lloyd et al.  found a relationship between NDVI trend and the response of trees to temperature: trees that were positively correlated with temperature were surrounded by pixels with more positive NDVI trends than others. This suggests that landscape-level NDVI trends may reflect within-population responses of trees to climate forcing.
4.4. NDVI Threshold of Detectability and Saturation
 The GIMMS NDVI data set has undergone extensive processing to remove exogenous sources of error and, when evaluated over time, shows consistent nontrending values in desert regions, which implies that the remaining influences of atmospheric conditions, bidirectional reflectance, calibration, and orbital drift are very small [Zhou et al., 2001; Tucker et al., 2005].Given that atmospheric conditions and sun-surface-sensor geometry in low to midlatitude deserts might not be representative of conditions in northern high-latitude forests, Zhou et al.  assessed the relationship between land surface temperatures and GIMMS NDVI and found a strong relationship, suggesting that interannual variation in GIMMS NDVI is measuring vegetation responses to temperature. Calibration error in the 1 km AVHRR NDVI data has been reported as 0.02–0.04 NDVI units [Vermote and Kaufman, 1995]. Chilar et al.  found that beyond calibration error, interannual variability exceeding 0.02–0.04 NDVI units could be detected in these data. However, the detection threshold in the GIMMS NDVI data set is considerably lower than the 0.04–0.08 thresholds suggested by the two prior studies [Tucker et al., 2005], owing to the superior calibration and solar zenith angel corrections in the GIMMS data processing. The site-level changes in GIMMS NDVI reported here, (about 0.08 NDVI units at greening sites and −0.12 NDVI units at browning sites) exceed the aforementioned detection thresholds and are therefore not attributable to measurement error in the data.
 When used as a proxy for leaf area, NDVI saturates once the leaf area index (LAI) exceeds ∼4; however, saturation is generally less outspoken when NDVI is used as a proxy for maximum photosynthetic rate [Sellers, 1985, 1987], although still hinders the use of NDVI to model productivity in areas of high biomass [e.g., see Olofsson et al., 2007]. Given the variability in tree cover across the coarse spatial resolution of GIMMS-NDVI data, in situ leaf area measurements were not taken at our study sites. In Saskatchewan and Manitoba, however, roughly 5°S of our study sites in the Northwest Territories, Chen et al.  reported the LAI ranging from one to four (mean ∼2) in P. banksiana stands and from one to six (mean ∼3) in P. mariana stands. It should be noted, that in the present study, vegetation cover at the Russian sites (66°–69°N) tended to be sparser than at the Canadian sites. In theory, a reduced sensitivity of NDVI to productivity changes with increasing LAI could thus reduce our ability to detect temporal trends in NDVI in the most densely forested (highest LAI) areas of our study domain, as it would reduce the signal-to-noise ratio. However, our results, as well as earlier findings by Bunn and Goetz , show that, contrary to this expectation, negative trends in NDVI are most prominent in areas of high tree cover implying that NDVI saturation effects have minimal bearing on our analysis.
 Our findings demonstrate a loose coupling between space-based measurements of canopy greenness and xylem production for three common tree taxa in northern Canada and Russia. The correlation magnitude is not entirely surprising given the complexity of trying to monitor tree ring growth from space, particularly since NDVI and ring width represent different metrics of carbon exchange. While neither is a perfect proxy, we know that NDVI is related to carbon uptake (GPP) via light harvesting, while ring width is more closely related to carbon assimilation (NPP). Starting from this point of divergence, there are many places where the relationship can further break down (sensor to canopy, canopy to needle, needle to plant, and plant to cambium). Landscape heterogeneity, decoupling of light absorption and utilization for carbon fixation, interannual variability in carbon allocation, and technical limitations (e.g., spatial and temporal resolution) all complicate relating NDVI and ring width on an annual basis. In spite of these limitations, we still observe a weak, though consistently positive and frequently statistically significant, association between the two measurements of carbon exchange.
 Across the 22 sampling sites, greening, browning and nontrending sites occurred with roughly equal frequency, though browning was concentrated in spruce and pine stands, while greening was most prevalent in sparse larch stands near tree line. Though the majority of sites that browned also experienced negative trends in ring width, greening was not accompanied by positive trends in ring width and may have largely resulted from increased shrub growth. Satellite-observed browning at some sites thus appears to relate to reduced forest C uptake and sequestration.
 We observed distinct autocorrelation patterns between evergreen (spruce and pine) and deciduous (larch) conifers in both NDVI and ring width measurements. This suggests that needle retention length influences interannual C storage and utilization for canopy and structural growth. This work illustrates that multiyear, lagged affects need to be considered when modeling forest productivity. Global carbon models are partially driven by spectral vegetation indices using space-based records that cover wide geographic areas but are only a few decades long. Tree ring records are spatially limited but span centennial to millennial timescales and show high decadal to centennial variability. It may be possible to incorporate some of the long-term (low frequency) variability into carbon models to develop scenarios for future changes to carbon fluxes that more realistically follow long-term tree growth dynamics.