This manuscript uses global vegetation models (GVMs) to project the likelihood of biome shifts in different climate change scenarios. The study shows that the used gvms are capable of reproducing current biome distribution as biome maps derived from model outputs agree with observation-based biome maps. The gvm biome projections show increasing risks of biomes shifts in higher emission scenarios and that higher latitudes were more likely to have biome shifts than lower latitudes, which the authors attribute primarily to temperature effects.

The manuscript is well written and generally clear, despite the methodology requiring some clarifications and the discussion being excessively long and descriptive in my opinion. I found the study results interesting and useful for scientists interested in climate change. My main criticism is related with how some hypotheses are tested. The authors use a classification based on expert knowledge of temperature limited and not-limited biomes and compare how these groups shift. The problem with this approach is that the actual temperature change in the grid cell is not considered, then it is impossible to claim if the biome really shifted because of temperature effects or some other reason (that is, even a very temperature limited biome still couldn’t have shifted because of temperature if there was no temperature changes in that particular grid cell to being with). I believe a more rigorous test should involve assessing the relationship between biome shifts and the actual temperature change in the grid cell.

Specific comments:

L24: Provide references for the claim that biomes are often used to assess vegetation change.

L35: I think this statement needs to be clarified. Model projections of precipitation are more uncertain than temperature projections but that doesn’t mean precipitation regimes won’t change in the future, only that model capacity to project it is currently more limited/varied between models. Therefore, I think it needs to be clarified that temperature limited biomes could respond more to current climate change projections, which does not necessarily mean they will be the most affected by climate change. In addition, it is important to consider that increased temperatures could also contribute to drought, through temperature effects on atmospheric vapor pressure deficit.

L95: I don’t fully understand the justification for using only increased CO2 scenarios. This was the combination that had the highest number of models available?

L98: Please clarify if all these models have a dynamic vegetation component (e.g. PFTs competing for grid-cell space, Prentice et al 2007)? I think it should be better explained how each model handles vegetation dynamics and responses to climate (especially temperature)  in supplementary material S1.

L115: Is it really sensible to discuss biome shifts in anthropogenic areas (for example Agricultural lands)? You are assuming agricultural lands are currently a natural biome based on potentially minuscule PFT fractions on it? Wouldn’t be more sensible to exclude these areas?

L145: Why 2 PFT products? One current and one future? Please clarify

L187: That doesn’t seem like a very rigorous test of the temperature effect.

L240: Please show graphically or at least describe these effects better. From the info in the previous paragraph, the difference between scenarios does not appear very large. P-values from traditional statistics are not always very informative when using very large samples that can be obtained from model outputs and will be highly “significant” even when the actual effect is very small. 

L270: Where do this p value comes from and what it refers to? The difference between coefficients of variation?

L279: I do not think this claim was rigorously tested in your results. I suggest investigating the temperature changes in grid cells with different likelihood of biome shifts.

L308: If fire is included in the GVMs you use and they still have low-agreement with data it shows that either the fire model is not working very well or that other factors not explicitly included in the models (such as herbivores) or even other aspects of the model structure or parametrization do not work well for the biomes here.

L333: Couldn’t these acclimation effects also affect the vulnerability of higher altitude biomes? As far as I know they aren’t included in GVMs.

L385: Do any of the used GVMs use these empirical limits? That would make the high agreement with observation maps less impressive than if were the result of bottom-up vegetation dynamic models as suggested in the beginning of the discussion.

L394: It would be important to clarify which processes are affected by temperature in each model in the methods section.

Fig. 4. Please define “Man” in the legend.

The authors performed a study to determine potential biome shifts under climate change. They make various projections, using 3 RCP scenarios, 5 GVMs and 31 biome maps. The novelty of this study is that they use random forest classification models to translate GVMs’ LAI results to determine biome distributions.

Overall, the paper is well written. My main concern is regarding the focus of this study in relation to its novelty. Why do you focus so much on the future predictions, while the novelty of this study focuses on how to use GVM output in a more meaningful or coherent way. I would much rather see that the study focusses on that, and describes the differences between biome maps (e.g. why do maps differ so much, and maybe if the future scenarios remain, what causes the larger biome shifts at the borders of biomes – is that really just temperature change?). There are already quite some future biome map studies, and while a more coherent one would be nice, this study doesn’t specifically focus on why the maps are different, so this study simply adds to the uncertainty in future biome mapping. I am not saying the study is not relevant, I am merely pointing out that, as I see it, the strong suit of this study is away from RCP scenarios and zooming in on how to make the best of GVM outputs and use them more directly, more consistently. The discussion now reads like a comparison of different papers on what biome changes where, which is not the main focus of this study and therefor perhaps not necessary (and not what you want to present as your main take away for this modelling effort). 

Abstract

There are a few questions that come to mind. First, why do we need biome projections if we have 31 observation-based maps already and we have 5 GVMs that project future vegetation. What is the benefit of the biome maps created in this study? Second, the 4-56% projected change is huge, and as I read it, it is a certainty check of the outcomes: depending on the GVM and observation based biome map the random forest is build on, you get these large variations in projected biome change. The RCP’s are simply scenarios you put in, so this is a simulation and not a model check. Thus, this sentence needs to be rephrased, as I probably do not understand what message you are actually trying to convey here. Third, what are your novel results? There are other studies that have projected biome changes under climate change, so what is new? I do read that Equatorial rainforest remains stable while other studies find forest dieback (yet I also know a study that projects increasing area for this biome). But what else? What is your method or result different from published literature?

Introduction

It doesn’t really become clear why this approach is necessary. Sure, many different models with different outcomes currently exist, but the introduction reads like a summation of what goes wrong instead of a delineation of what these maps are actually used for and how they are created. The observation based biome maps is still an unclear aspect to me, how do you have a global (?right – not mentioned until now) observational map of biomes? Especially since it exists in 31 different versions? Additionally, the uncertainties mentioned in the introduction around GVMs seep through in the future predictions. Not only the aftermath (going from model results on biomes) is what causes uncertainty, but also various aspects within the GVMs. How do you propose that is handled by your random forest? I don’t understand how a more direct link between GVM output and biomes helps reduce most of the uncertainty. Relatively speaking, how much of the uncertainty in GVM results comes from the final aggregation of data into biomes?

Methods

Is there a way to indicate how LAI and PFT are linked?

The biome classification was done for each combination of GVM result and biome map. Why not make one model, to enhance comparison? You mention that the aftermath of GVM results into biomes creates much uncertainty, but now you do the same but with a random forest instead of the method GVM people decided to use. What is the difference?

I wonder how the random forest models are dependent on the prevalence of the different biomes. Are biomes that are more frequent on a global map not also predicted more often, simply because the change that a ‘large’ biome occurs is bigger? Random forests can only predict ‘small’ biomes in the case that the predictors there are extremely specific.

Results

The results presented in Figure S2 are the basis of the rest of the results, right? I am not sure it is best placed in the SI in that case. Also, the fact that some GVMs have such large disagreements, I find it tricky to understand what this means and why all GVMs are all still used for the rest of the analyses. Additionally, you mention that biome shifts are mostly present at borders of biomes. That is quite logical, but what I would like to see (perhaps in the discussion you mention it already) is that the tropical regions don’t change. Great impact of climate change in e.g. the Amazon has been reported, but this is not reflected. How do GVMs deal with novel climate combinations? Is there an upper limit in climate suitability for the warmest biomes or lower limit for the driest biomes? Or are all warmer or drier conditions simply outputted as the biome with the warmest or driest climate? This links to the statement in lines 322-325 in the discussion.

In line 219, you mention susceptibility category and link this to consensus between models. I don’t understand that. Biome change and model agreement is something else, and the link you refer to is not made specific.

I am wondering about the comparison with the Olsen map you make in section 3.3. I know that somewhere in the methods you describe that nothing changes in biome structure between early 2000 and 2020, but it is strange that you look into the future until 75 years from now (2099) and ignore the first 20 years. Coming back to section 3.3, you describe that the olson map is reproduced with high agreement. But that map was made before 2001, so should the GVM result from the current climate really match ‘such an old map’ that well?

I do not understand the difference between figure 2 and 5.