The present study focuses on two sources of uncertainty: (i) algorithm (Defourny et al. 2014) of converting surface reflectance into LC classes (LC mapping), and (ii) CW procedure into PFT. The range of uncertainties in PFTs map derived from ESA-CCI-LC map is quantified and compared with the range of uncertainty in forest cover observations and recent LULC change (Section 3.1). Furthermore, the impact of PFT uncertainty on MPI-ESM simulated land surface fluxes and near-surface climate is calculated (Section 3.2).
The range of PFT uncertainty
Table 2 sums the amount of the area (in Mha) covered with major vegetation types, global, and for four latitudinal zones (40 N–70 N, 10 N–40 N, 20 S–10 N, 60 S–20 S) classified by Hartley et al. (2017) as the latitudinal zones with distinctive vegetation uncertainties. Comparing these areas (Table 2 for the globe) with the area of historical LC change estimated in the literature, it turns out that the range of uncertainty is about the same order of magnitude as the historical LC change. Ramankutty and Foley (1999) estimated that approximately 1200 Mha of trees have been removed globally since 1700 up to 1992. In the simulations that minimize vegetation cover, there is 721 Mha (minLC_refCW) and 1740 Mha (minLC_minCW) less trees than in the reference experiment (refLC_refCW). In the simulations that maximize vegetation, there is 633 Mha (maxLC_refCW) and 1229 Mha (maxLC_maxCW) more trees than in the reference simulation. According to ESA-CCI-LC data for epoch 2010 used as reference LC, the area currently under farming is 2365 Mha, while Ramankutty and Foley (1999) estimated that there was 1800 Mha under farming in the year 1992. Here, it is also interesting to note the non-uniform distribution of farming area in our experiments, so that minLC_refCW simulation has the largest farming area of 2635 Mha, while maxLC_minCW has the smallest farming area (1709 Mha).
The range of uncertainty for forest distribution can be compared with other datasets. According to Forest Resources Assessment (FRA) reports (Keenan et al. 2015, Table 11), global forest area, for example, in the year 2000 ranges from 3870 (FRA 2000) to 4056 Mha (FRA 2015). Therefore, the quality of data used in various FRA surveys resulted with 186-Mha area of global forest uncertainty. Based on satellite data, Hansen et al. (2010) estimated global forest area of 3269 Mha, while in a later study Hansen et al. (2013) have estimated 4145 Mha. The difference between these two studies is 876 Mha, which is about the same order of magnitude as the uncertainty due to LC mapping or CW procedure in our study. In addition, the most recent studies seem to be more accurate than earlier studies, i.e., forest cover estimates from FRA 2015 are considered to be more accurate than estimates from FRA 2000 and estimates from Hansen et al. (2013) are considered to be more accurate than estimates from Hansen et al. (2010). Especially for the latter, this is due to the use of better input data (e.g., finer resolution imagery), improved methods, and different definition. Also Gross et al. (2017) report more accurate results for more recent estimates from finer resolution imagery. But comparing estimates from different studies does not necessarily provide reliable information about the reduced uncertainty of the most recent estimates. Another example to illustrate this point is the “discovery” of 400 Mha of forests in the drylands (Bastin et al. 2017). These “missing” forests are mainly (or largely) open forest (i.e., between 10 and 50% tree cover) which are considered as forests by Bastin et al. (2017) (FAO definition) but should not be considered as forests for climate simulations (predominance of shrub and grass cover).
Table 2 also shows vegetation variations due to uncertainty across the latitudinal zones. For example, the largest relative increase of forest area (81%) appears in 10 N–40 N zone for maxLC_maxCW, while the largest relative decrease (69%) appears for the minLC_minCW case in the same zone. However, the largest absolute variation occurs in the 40 N–70 N zone.
Figure 2 identifies regions where the variations in JSBACH PFT distribution occur due to the uncertainty in LC mapping algorithm (minLC_refCW and maxLC_refCW) and CW procedure (minLC_minCW and maxLC_maxCW). As already noted by Hartley et al. (2017), these variations are more pronounced for CW uncertainty than for LC mapping uncertainty. For extra-tropical evergreen trees, the largest variation in geospatial distribution of PFTs occur in Northern North America and Canada, Scandinavia, Northern Russia (from Baltic to Ural), and Southeastern China. For the extra-tropical deciduous trees, the largest variation due to uncertainty is located in Northern Russia ranging from the West Siberian Plain to the Bering Strait, in Zambezi river basin and in the South American Pampas. The most notable variation in the distribution of tropical trees is in Amazon and Congo River basins.
Other notable variation appears for shrubs and herbaceous types. For example, maxLC_maxCW is characterized by the decrease in shrubs from approximately 40 N to 70 N latitude (Table 2 and Fig. 2), as well as along the northwestern border of Parana River basin in South America and in the area between the Indochina peninsula and the Yangtze River basin. This experiment is also characterized by the increase of shrubs especially along the southern and eastern coast of Australia and in some parts of sub-Saharan Africa (see green line on the shrubland panel in Fig. 2).
The global increase in the grassland area in comparison to reference case (refLC_refCW), in particular of the C3 type, characterizes all experiments. However, the most notable increase is for minLC_minCW that minimizes vegetation due to CW uncertainty (see Table 2 and red line on the grassland panel in Fig. 2).
The largest variations in croplands are in the sub-Saharan area, between 10 N and 50 N over the Eurasian continent, along the eastern coast of South America, in Central America and to the north of the Gulf of Mexico. Crops have a productivity comparable to trees, but albedo and transpiration properties are similar like grasses. Thus, variations in crops are expected to have a nonlinear feedback across the five experiments. Therefore, note that the largest increase in crops appears in minLC_refCW (cf. Cropland panels in Figs. 1 and 2 and Table 2), in particular on the southern hemisphere (SH).
MPI-ESM response to the PFT uncertainty
The impact of the LC uncertainty and the range of the MPI-ESM response of annual mean climate are summarized in Fig. 3 and Tables 3 and 4. Table 3 shows comparison of JSBACH offline and MPI-ESM data with observations. Though, albedo in JSBACH and MPI-ESM shows some differences in interannual variability (Fig. 3), the range of uncertainty is the same for both of them and amounts 0.024 (ranging from 0.304 to 0.280 in JSBACH and from 0.298 to 0.274 in MPI-ESM, see Table 3). GPP shows larger interannual variability in MPI-ESM simulation (Fig. 3), but uncertainty is larger for JSBACH simulations ranging from 135.917 to 173.253 Pg C year− 1, while for MPI-ESM, GPP ranges from 134.990 to 167.492 Pg C year− 1. ET shows larger uncertainty and interannual variability for MPI-ESM simulation. However, this is not due to coupling of surface and atmospheric processes, but due to model deficiency in the JSBACH offline version used in previous study that is resolved in a coupled setup within MPI-ESM used in the present study.
Comparing the values for evapotranspiration in MPI-ESM simulation (Table 3, ranging from 72748 to 77017 km3, i.e., ∼± 2000 km3 from the reference) with the estimated decrease in terrestrial evapotranspiration due to deforestation (Sterling et al. 2012, ∼3500 km3), it turns out that the range of uncertainty for certain variables is about the same order of magnitude as the estimated LULC climate change. Figure 3 shows normalized scores for annual means of various climate variables for the five experiments conducted with MPI-ESM and JSBACH offline, where the latter was taken from Hartley et al. (2017). Similar as Figure 6 in Hartley et al. (2017), the normalized scores for MPI-ESM simulations convey the same message as offline simulations. Albedo is the most impacted, equally affected by LC and CW uncertainty. It decreases with an increase in vegetation. The differences in albedo between the JSBACH offline and MPI-ESM simulations are due to differences in the prescribed WFDEI precipitation and MPI-ESM simulated precipitation which result in different snow cover in both types of simulations.
The response of GPP to the PFT uncertainty is similar for both setups, except that MPI-ESM simulations show stronger interannual variability. In both cases, GPP is much strongly affected by the CW uncertainty than by the LC mapping uncertainty. This is probably because the biggest variation in tree cover occurs for this uncertainty and trees are the largest primary producers. However, minimizing vegetation due to LC mapping uncertainty (minLC_refCW) shows a similar anomaly as in Hartley et al. (2017), i.e., it shows an increase of GPP with a reduction of vegetation. This is probably due to the largest area (2635 Mha) covered by crops in this experiment and crops have larger productivity than grasses.
In the previous version of JSBACH-offline used in Hartley et al. (2017), ET did not show much variation due to PFT uncertainty. This bug was specific to the offline version of JSBACH, but is not included in coupled setup so that the ET behavior is improved in the MPI-ESM simulations presented in this study, where ET linearly increases with increasing vegetation. As total precipitation (TP) over land and 2m air temperature (T2M) over land are prescribed in the offline simulations, they are not analyzed for JSBACH simulations but only for the MPI-ESM simulations. Here, TP increases linearly with the increase of vegetation, while T2M does not show a systematic dependence on vegetation globally, but rather regionally.
Table 4 shows global and regional (four latitudinal zones) uncertainty of five key surface climate variables in MPI-ESM. For example, global uncertainty in GPP is estimated to be ∼± 16 Pg C year− 1. The largest zonal uncertainty from − 5 to 6.6 Pg C year− 1 occurs in the 40 N–70 N Table (4) zone. This is also the zone featuring the largest variation in the tree distribution (from 573 to 2326 Mha, Table 2), also affecting albedo uncertainty to range from − 0.035 to 0.026. Evapotranspiration uncertainty (in 40 N–70 N zone) ranges from ∼− 926 to 995 km3 year− 1 or from ∼− 20 to 21 mm year− 1. All these parameters depend largely on the uncertainty in vegetation distribution, i.e., their stomatal conductance and reflective properties. Largest uncertainty in precipitation (ranging from − 1473 to 1302 km3 year− 1 or from ∼− 51 to 45 mm year− 1) and 2m temperature (ranging from ∼− 0.1 to 0.2 K ) are estimated in the 20 S–10 N zone.
The box plots on Fig. 3, overlaid over annual mean scores, provide an interesting insight in the distribution of frequencies and how the simulated climate is affected by uncertainty in vegetation. For example, the median amount of precipitation for minLC_minCW lies just below the 5th percentile for maxLC_maxCW, i.e., the median amount of precipitation for the minLC_minCW experiment equals the precipitation of a very dry year in the maxLC_maxCW experiment. For ET, this difference is even more pronounced, leading to the conclusion that a median year for the minLC_minCW experiment would be a dry year in the reference (refLC_refCW) simulation, and an extremely dry in maxLC_maxCW. Positive extremes show a similar behavior. The median amount of precipitation for maxLC_maxCW lies above the 95th percentile of minLC_minCW, i.e., it has a similar amount of precipitation as the wettest year in minLC_minCW.
This redistribution of precipitation pattern indicates that PFT uncertainty has a considerable impact on large-scale phenomena, such as NAO and ENSO, and their regional implications such as monsoons and weather regimes simulated by an ESM. While studying offline LSMs, Hartley et al. (2017) could only consider land surface variables. In the present study, with MPI-ESM, we can also investigate the impact of LC uncertainties on atmospheric variables.
Figure 4 shows boreal winter (December, January, February—DJF) deviations of mean sea level pressure and 10-m winds from the reference experiment for the uncertainty experiments. Those experiments that either minimize or maximize vegetation due to CW uncertainty (minLC_minCW, and maxLC_maxCW) show a clear impact on the mid-latitude westerlies in northern hemisphere (NH) during DJF. In the minLC_minCW experiment, westerlies are strengthening while in the maxLC_maxCW experiment they are weakening. Experiments that either minimize (minLC_refCW) or maximize (maxLC_refCW) vegetation due to LC uncertainty both contribute to the formation of blocking like features, the former above the Atlantic Ocean to the north of the Great Britain, and the latter above the Central Europe. Both of them seem to have impact on increasing the Azores and Siberian high and deepening the Icelandic depression during the boreal winter. This results in intensified westerlies over the Atlantic Ocean. It is more pronounced for maxLC_refCW. These deviations in circulation pattern can be explained by an increase of surface roughness with an increase of vegetation. In addition, there are variations in the atmospheric water vapor distribution that impact the atmospheric pressure patterns and, hence, the circulation.
Figure 5 shows the related deviations in circulation during the boreal summer (June, July, August—JJA). On the NH, only minLC_minCW shows some amplification of westerlies over the Eurasian mid-latitudes while the other simulations show negligible variation in wind speed for that area. The SH also features perturbations in circulation pattern during both seasons (Figs. 4 and 5). It is interesting to note the strengthening of the high-pressure field to the south of the African continent during the JJA season, in particular for minLC_minCW and minLC_refCW. This high-pressure field brings moist oceanic air to the Indian subcontinent and it may intensify the Indian monsoon. Hence, this demonstrates that vegetation uncertainties have a noticeable impact on the large-scale atmospheric circulation.
The complex pattern of seasonal (DJF and JJA) variations in 2m temperature due to vegetation uncertainty as well as variations in albedo and evapotranspiration are shown in Figs. 6 and 7. Variations in temperature depend on several factors such as vegetation type, snow cover, and solar insolation related to geographic latitude. During the winter (DJF, Fig. 6), variations in NH temperature are controlled by albedo feedback and advection (Figs. 4 and 6). Variations in SH 2m temperature during winter, in particular for cases with increased vegetation (maxLC_maxCW and maxLC_refCW), are predominantly controlled by evaporative cooling. Experiments that decrease vegetation (minLC_minCW and minLC_refCW) show impact of albedo feedback and evaporative cooling on temperature. During the summer (JJA, Fig. 7), evaporative cooling takes a predominant control over 2m temperature changes, especially over North America. The albedo feedback seems to be more important for the cases that minimize vegetation (minLC_minCW and minLC_refCW).
The impact of vegetation uncertainty on annual mean T2M and TP over land is shown in Figs. 8 and 9, respectively. Statistical significance of the annual mean (T2M and TP) deviations from refLC_refCW has been tested with a T test and with a simple standard deviation test. The latter is performed as following. Model internal variability is defined as the standard deviation of five-member ensemble performed by de Vrese and Hagemann (2016). The model setup is identical, but the simulations were started using slightly differing initial conditions. In that way, defined model internal variability is compared with the deviations of uncertainty experiments from the reference simulation. Grid points where the deviations of uncertainty experiments are larger than two standard deviations (internal variability) of the ensemble roughly coincide with the grid points showing 95% significance level according to T test. Therefore, only the former are indicated on Figs. 8 and 9. Figure 8 shows the net annual impact of seasonal variations in albedo feedback, evaporative cooling, and other factors related to PFT uncertainty, on the 2m temperature. Boreal latitudes of North America and in particular Canada exhibit cooling with decrease of vegetation (minLC_minCW and minLC_refCW) and warming with increase of vegetation (maxLC_maxCW and maxLC_refCW) indicating albedo feedback control over temperature. On the other hand, South America and sub-Saharan Africa exhibit the opposite signal, i.e., warming with decrease of vegetation and cooling with increase of vegetation in particular related to CW uncertainty (minLC_minCW and maxLC_maxCW). Therefore, this indicate evaporative cooling as dominant control over the 2m temperature for South America and sub-Saharan Africa. The Eurasian continent shows interference of both effects and the largest changes in temperature due to PFT uncertainty. The most significant warming with increasing vegetation occurs along the northeastern coast of the Eurasian continent. Seasonal variations (Figs. 6 and 7) can be even stronger. During the boreal spring (March, April, May—MAM), maxLC_maxCW shows a local warming up to 3 K along the northeastern coast of the Eurasian continent and the northwestern part of North America.
The most significant impact on precipitation (Fig. 9) appears due to CW (minLC_minCW and maxLC_maxCW) uncertainty. The major variations in precipitation occur in the Amazon, Congo, and Indonesian rainforest, but also in North America and Central Eurasia. The feedback is positive, i.e., less vegetation–less precipitation and vice versa. Figure 9 also shows monsoon rain domains as defined by Devaraju et al. (2015) following Wang and Ding (2006). Seasonal and annual mean variations of terrestrial precipitation in monsoon regions are shown in Fig. 10 The largest relative deviation of precipitation (∼± 14%) is about the same order of magnitude as in Devaraju et al. (2015), though they do not occur in the same regions. The South African domain shows ∼14% increase during the JJA season and the Australian region exhibit ∼14% decrease of precipitation during MAM season. Except for the East Asian boreal winter (DJF) monsoon, all other NH DJF monsoons (North American, North African, and South Asian) intensify with decrease of vegetation, similarly as the SH DJF monsoons (Australian, South American, and South African). During austral winter (JJA), the weakening of precipitation with decrease in vegetation is a dominant feature. However, NH monsoons are not so strongly affected as the SH monsoons by the decrease of vegetation. Experiments that maximize vegetation predominantly show intensification of DJF monsoonal precipitation, while impact on JJA monsoons is negligible.