Excessive positive response of model‐simulated land net primary production to climate changes over circumboreal forests

Abstract Land carbon cycle components in an Earth system model (ESM) play a crucial role in the projections of forest ecosystem responses to climate/environmental changes. Evaluating models from the viewpoint of observations is essential for an improved understanding of model performance and for identifying uncertainties in their outputs. Herein, we evaluated the land net primary production (NPP) for circumboreal forests simulated with 10 ESMs in Phase 5 of the Coupled Model Intercomparison Project by comparisons with observation‐based indexes for forest productivity, namely, the composite version 3G of the normalized difference vegetation index (NDVI3g) and tree‐ring width index (RWI). These indexes show similar patterns in response to past climate change over the forests, i.e., a one‐year time lag response and smaller positive responses to past climate changes in comparison with the land NPP simulated by the ESMs. The latter showed overly positive responses to past temperature and/or precipitation changes in comparison with the NDVI3g and RWI. These results indicate that ESMs may overestimate the future forest NPP of circumboreal forests (particularly for inland dry regions, such as inner Alaska and Canada, and eastern Siberia, and for hotter, southern regions, such as central Europe) under the expected increases in both average global temperature and precipitation, which are common to all current ESMs.

. Tei and Sugimoto (2018) determined that this is a common phenomenon observable by satellite imagery. However, such a time lag response could not be reproduced by the DGVM used by , Tei, Sugimoto, Liang, et al. (2017).
It remains unclear whether the above discrepancies in forest responses to climate change between the observation-based indexes (e.g. composite version 3G of the NDVI (NDVI3g) and RWI) and DGVM output is a feature of the DGVM used by , Tei, Sugimoto, Liang, et al. (2017) or a more generalized feature for all models. Klesse et al. (2018) reported that most of the current DGVM generations strongly overestimate the temperature sensitivity of cold environments over European forests. In addition, mismatches in correlations with summer climate variables between DGVMs and RWI (Babst et al., 2013) and those of lag effects of previous-year summer climates (Zhang et al., 2018) have also been reported for European forests. However, it is still unclear whether these discrepancies may be observed for other ecosystems beyond European forests. Therefore, we compared observation-based indexes for forest productivity, namely, the NDVI3g and RWI, with the land NPP simulated by land carbon cycle components of 10 ESMs involved in Phase 5 of the Coupled Model Intercomparison Project (CMIP5) (as opposed to a single vegetation model). We performed this comparison to gain a better understanding of model performance and identify uncertainties in their output over circumboreal forests, similar to the aforementioned discrepancies in previous studies Tei & Sugimoto, 2018;Tei, Sugimoto, Liang, et al., 2017;.
Circumboreal forest ecosystems cover 22% of the Earth's terrestrial surface and account for 12% of the global NPP, thereby playing an important role in attenuating recent global warming through photosynthesis (Chapin et al., 2005;Kimball et al., 2006;McGuire et al., 2009). Under the strong warming trend in high latitudinal areas, i.e., Arctic amplification (Serreze & Barry, 2011), drastic changes, such as permafrost degradation (Lemke et al., 2007), changes in the tree line location (Frost & Epstein, 2014;Tchebakova, Parfenova, & Soja, 2009), shifts in tree growth patterns (Tei, Sugimoto, Liang, et al., 2017;, and forest/tree decay (Allen, Breshears, & McDowell, 2015;Tei, Sugimoto, Yonenobu, Kotani, & Maximov, 2019), have all been observed and are processes that would be expected to continue in circumboreal forest ecosystems. However, significant uncertainty remains about the magnitude and location of these spatial variations, which serve to alter the ecosystems' ability to behave as a carbon sink (i.e., the amount of carbon dioxide absorbed from the atmosphere would be altered) (Ciais et al., 2010;Goodale et al., 2002).
Therefore, understanding spatial variations in the response of forest ecosystems to climate change in the northern high-latitude regions is crucial for accurate projection of the terrestrial carbon cycle and global climate.

| Satellite vegetation greenness datasets
We employed the bimonthly maximum NDVI3g dataset, which is the latest dataset released by the Global Inventory Modeling and Mapping Studies group (Pinzon & Tucker, 2014;Tucker et al., 2005), and covers the period of 1981-2015 with a native spatial resolution of 0.083°. The NDVI product is an indicator of photosynthetic activity ( Running, 1997) and interannual variability in tree growth over circumboreal ecosystems (Berner, Beck, Bunn, Lloyd, & Goetz, 2011;Bunn et al., 2013), computed as the difference between the nearinfrared and red reflectance of the land surface normalized by the sum of the reflectances. NDVI3g comprises recently revised data of its previous version, NDVI version G (NDVIg), which specifically aims to improve the data quality for high-latitude regions (Zhu et al., 2013). Explicit corrections of the effects of orbital drift and stratospheric aerosols from volcanic eruptions have also been applied (Tucker et al., 2005). Therefore, it is suitable for studying changes in vegetation activities in northern high-latitude regions (Guay et al., 2014;Xu et al., 2013).
NDVI3g average summer values were used for the analyses. In consideration of the different lengths of the plant-growing season (e.g., Lund et al., 2010), the definition of the summer season varied depending on the latitude. Specifically, only one month (July) was defined as summer for the sites to the north of the Arctic circle (~67° N), whereas three months (June, July, and August) were considered as summer for the sites located between 50° N and 67° N.
The following criteria were employed for sample site selection: (a) sites were located farther north than 50°N to represent the southern limit of boreal forests; (b) the sites had an elevation lower than 2000 m above sea level to avoid excessive sampling from temperature-sensitive trees due to the cold environment in high altitude areas; and (c) sites had a chronology ending after 1990 to better represent the effects of recent global warming (Table S1). In addition, we incorporated six published chronologies from our previous works (Tei, Sugimoto, Liang, et al., 2017;Tei, Sugimoto, Yonenobu, Ohta, & Maximov, 2014) (Table S2).
We generated a standard RWI for each site using the ARSTAN software (Cook, 1985). The standard RWI was developed by first detrending the growth trends associated with tree age and natural disturbances in a raw time series, which were derived from the single dominant tree species in each site by cubic smoothing splines and subsequently averaging the standardized ring widths among the samples (Cook, 1985). Therefore, we applied spline fitting to standardize the chronologies at all sites, which is regarded as being a relatively site-insensitive method (Cook, 1985;Cook & Kairiukstis, 1990); thus, this is in line with our approach involving various forest ecosystem types (  (Tables S1 and S2). Therefore, for subsequent analyses, we used the RWI of these 554 sites.
Where multiple ensemble members were available from a single historical simulation, we calculated the ensemble mean for these members. Finally, we calculated the annual sum of the monthly land NPP.
These parameters were all used for further analyses.

| Data re-sampling for a consistent grid
There was a substantial mismatch in spatial scale between the RWI, NDVI3g, and ESM-simulated land NPP. The majority of the tree cores were collected over an area around one hectare, whereas each NDVI3g grid cell covered ~64 km 2 and each ESM grid cell covered 10,000-30,000 km 2 . Therefore, in this study, all ESMs were re-sampled into a 2.81° × 2.81° consistent grid, which was the largest grid used in the model. NDVI3g grid cells were then averaged to the consistent grid. RWIs from multiple sites were averaged if they occurred within the consistent grid. In this way, all ESMs and NDVI3g represented a spatial scale of the same size, and the RWI was as close to that spatial scale as possible, allowing for better aggregation.
Aggregating the RWI and NDVI3g over this broader area ensured a better representation of the conditions simulated by the ESMs. higher latitudes (July for Arctic circle (~67°N) and June-July-August for the sites located between 50°N and 67°N).

| Correlation function analyses
Correlation function analyses were performed using the R software  where the percentage of grids with significant positive correlations was 10.5%-40.5% (Table 3). The correlations between the ESM MPI and current-year summer precipitation were more negative than those for the RWI, NDVI3g, and the other ESMs, in which 48.7% of the total grids had significant negative correlations (Table 3). The plot shapes were very similar among the ESMs. In contrast, for the RWI, the peak of the negative correlation was biased toward being more significant than that of the ESMs and NDVI3g. The density plots for current-year data are shown in Figure 7b. Similar to those for previous-year data, the shapes of the plots were similar among the ESMs; however, the RWI and NEVI3g showed larger peaks for the negative correlation side, and peaks on the positive correlation side were biased toward being less significant than those for the ESMs. Figure 7c shows the density plots for both previousand current-year data. The correlation coefficients were calculated based on meteorological data for temperature and precipitation in both the previous and current years for each grid, and the highest correlation coefficient was then used for the density plot. The shapes were similar to those for the current-year data, showing that the RWI and NDVI3g had larger peaks for the negative correlation side, and peaks on the positive correlation side were biased toward being less significant than those for the ESMs.

| Responses of the RWI and NDVI3g to past climate change
The RWI and NDVI3g had clear opposite responses to summer temperature and precipitation between the previous-and current-year data, which were never observed in the ESMs (Figures 5-6). Both these indexes showed positive correlations with current summer temperatures and negative correlations with previous summer temperatures, although the RWI covered large areas with significant correlations (Figures 1 and 3). An opposite correlation pattern was observed for the summer precipitation (Figures 2 and 4). The observed negative response to summer temperatures and the positive response to precipitation likely indicate a reduction in the stomatal conductance of plants, which, under water-limited conditions, would have resulted in lower plant carbon uptakes (Tei et al., 2014). Water availability is also closely coupled with cambial activity and the wood formation of trees (Balducci, Deslauriers, Giovannelli, Rossi, & Rathgeber, 2013), and radial growth requires the maintenance of high cell turgor pressure, which has an irreversible influence on cell extension and wall polymer deposition (Proseus & Boyer, 2005). In addition, warming-induced carbon losses through plant respiration might have been the cause for the negative response.
The negative response was clearer for the RWI-temperature relationship than for the NDVI3g-temperature relationship (Figure 5a,b and Table 2) and was dominant within inner Alaska and Canada, central Europe, and eastern Siberia (Figure 1). This negative response of the RWI to warming corresponded with previous studies on local/regional tree-ring and continental dry climate regions (Andreu-Hayles et al., 2011;Barber et al., 2000;Silva, Anand, Oliveira, & Pillar, 2009;Tei, Sugimoto, Liang, et al., 2017;. A higher summer temperature could have caused a high evapotranspiration from surfaces, resulting in drier conditions that would have reduced the stomatal conductance in dry regions, which TA B L E 4 Intensity of effects for summer temperature and precipitation for the RWI, NDVI3g, and ESMs Note: Positive_t and Negative_t indicate the number of grids with more significant positive and negative correlations for summer temperature, respectively. Positive_p and Negative_p indicate the number of grids with more significant positive and negative correlations for summer precipitation, respectively. could have saved water losses by plant transpiration, but would also have reduced plant carbon uptakes (Tei et al., 2014). (2018)  The observed time lag responses of the RWI and NDVI3g to climate change ( Figures 5-6), i.e., negative and positive responses to summer temperature and precipitation, respectively, were also important features in the observation-based indexes for forest response to climate change (Barber et al., 2000;Girardin et al., 2016;Tei & Sugimoto, 2018;Tei et al., 2014;Walker et al., 2015). Large differences in the RWI and NDVI3g responses to climate variables between the previous-and current-year data were observed at lower (<500 mm) annual total precipitation sites ( Figures S7-S8), suggesting that water-stressed environments may be related to the observed time lag responses.

Tei and Sugimoto
The adaptation of forest ecosystems to severely water-stressed environments, i.e., carrying the carbon fixed during one particular year into the following year (Kagawa, Sugimoto, & Maximov, 2006;McDowell et al., 2008), could be the cause for this forest/tree response to climate change. Plants store fixed carbon as nonstructural carbon compounds, which are energy sources for biosynthesis in the following growing season (Chapin et al., 1990;Wiley & Helliker, 2012). Several studies have suggested that carbon storage could be an active process, occurring at the expense of superior growth during a particular year (Chapin et al., 1990;Genet et al., 2010) and could be a beneficial adaptive measure to severely water-stressed environments. Increased carbon storage could reduce the risk of carbon starvation under severe drought conditions with reduced growth as a trade-off (McDowell et al., 2008).

| Different forest responses to past climate change in observation-and model-based estimates
The simulated land NPP from most ESMs showed more significant correlations with climate variables in the current-year data than that in the previous-year data, with mostly positive correlations ( Figures 5 and 6). The ESMs covered larger geographical areas (areal extent) and had more significant correlations with current summer F I G U R E 7 Density plots for correlations of the RWI, NDVI3g, and land NPP by ESMs with summer temperature or precipitation for (a) previous; (b) current; and (c) both years. The correlation coefficients were calculated based on meteorological data from both temperature and precipitation in the previous, current, and both years, respectively, for each grid with the highest correlation coefficient used for the density plot climate variables (0.6%~86.7% for p < .05) (Table 3) than with previous summer climate variables (0.0%~13.8% for p < .05) ( Table 2).
These responses of the ESMs had a clearer relationship with annual average temperature than the RWI or NDVI3g (Figures S9 and S10).
The simulated land NPP responses to the current-year summer temperature changed from positive to negative as the annual average temperature increased ( Figure S9), and the opposite change was observed for responses to the current-year summer precipitation ( Figure S10), suggesting that the land NPP responded negatively and positively to summer temperature and precipitation, respectively, in relatively high temperature regions. These responses were also observed in the RWI and NDVI3g; however, they were much weaker.
The observed overestimation of the simulated land NPP sensitivity to precipitation was applicable to all ESMs used in this study, except for ESM MPI, (Figure 6, Tables 2 and 3) as reported in previous RWI-model studies (Rammig et al., 2015;Tei, Sugimoto, Liang, et al., 2017;. Some ESMs showed significant negative correlations with summer temperature within inner Canada, central Europe, and/or eastern Siberia, similar to the RWI, although this involved the current year and not previous year summer data (Figures 1 and 3, Tables 2 and 3). However, we clarified that the negative effect of summer temperature was masked by a much more critical contribution of summer precipitation for most of the simulated land NPP (Table 4). Therefore, the land NPP appeared to respond extremely positively to past temperature and precipitation changes compared with the RWI and NDVI3g (Figure 7b,c), both of which have been reported as RWI-DGVMs climate sensitivity discrepancies in European forests .
Thus, projections of the land NPP of circumboreal forests might be overestimated under the expected increases in both average global temperature and precipitation (IPCC, 2013). It would be difficult to estimate the extent of overestimation accurately; however, according to the observed dominated and significant negative response of the RWI to summer temperatures (Figure 1a), an overestimation is likely to be more pronounced in boreal forests than in the Arctic tundra and boundary ecosystem, especially for inland dry regions, such as inner Alaska and Canada, and eastern Siberia, and for hotter, southern regions, such as central Europe. In addition, the ESMs did not reproduce the time lag and the negative response to increased temperatures, which were observed in the RWI and NDVI3g ( Figure 5). Especially in Alaska, no ESM reproduced the time lag and negative response (Figure 1).
These features for responses of the simulated land NPP to climate changes have been reported for SEIB-DGVM (Sato et al., 2016) by    Girardin et al., 2016;Tei & Sugimoto, 2018). However, our results indicated a different climate sensitivity for the RWI and NDVI3g, e.g., a weaker sensitivity of the NDVI3g to changes in summer temperature and precipitation in both previous and current years (Tables 2   and 3). The most likely reason for this discrepancy is that NDVI3g approaches could not reflect respiratory plant carbon losses because they only detect changes in leaf/needle greenness. Moreover, the observed large interannual variability in plant carbon allocation among organs and in the production of secondary plant compounds might also be a contributing factor (Litton, Raich, & Ryan, 2007).
Although our results indicated the differences between the RWI and NDVI3g, these differences were smaller than those between observation-and model-based estimates.

| Future perspectives
We evaluated the land NPP simulated by the 10 ESMs involved in CMIP5 by comparing them with observation-based indexes for forest productivity, namely NDVI3g and RWI, over circumboreal forests.
We focused on forest responses to summer temperature and precipitation (Figures 1-4 and Tables S1 and S2). However, climate variables other than the season have been found to affect vegetation activities over the northern high-latitude regions (Tei & Sugimoto, 2018;Wang et al., 2011). Tei and Sugimoto (2018) reported that in the Arctic tundra and boundary ecosystems, the climate variables of the current summer were the main climatic drivers for the positive response to the increase in temperatures shown by both the NDVI3g and the RWI indices. In contrast, in boreal forest ecosystem, climate variables of the previous year (from summer to winter) were also important climatic drivers for both the NDVI3g and RWI. Importantly, both indices indicated that the temperatures in the previous year negatively affected the ecosystem. Such forest responses to climate variables other than the summer season were probably derived from changes in the timing of snowmelt in spring and/or the availability of nitrogen in the following summer (Kirdyanov, Hughes, Vaganov, Schweingruber, & Silkin, 2003;Sidorova et al., 2009;Tei, Sugimoto, Liang, et al., 2017;Vaganov, Hughes, Kirdyanov, Schweingruber, & Silkin, 1999). Tei and Sugimoto (2018) also reported that large geographical areas (areal extent) over circumboreal forests were significantly affected by temperature and precipitation in summer for both the RWI and NDVI3g. Nonetheless, the extent to which the spatial variability of seasonal climatic variables, controlling temporal variations in forest greenness (NDVI3g) and tree growth (RWI), over the northern high-latitude regions could be reproduced by current ESMs is an important issue for further development of land carbon cycle components in ESMs.
In addition, the distribution of RWI sites in the International Tree-Ring Data Bank is biased toward Europe and North America (Figures 1-4 and Tables S1 and S2). Therefore, all datasets herein may have some regional bias, especially because we used only the NDVI3g and land NPP from grids over the RWI chronology sites. Further efforts to fill the gaps in RWI data, e.g., in central and western Canada and Siberia, are required. This study contributes to the understanding of RWI-based forest responses to climate changes with fewer regional biases, resulting in a more useful validation dataset for simulations.