In this manuscript, the authors applied a data-driven or machine learning approach, the conventional neural network (CCN), to estimate global distribution of the potential vegetation types. They evaluated accuracy of retrieving the present state, and then estimated future shifts under projected climate change. Finally, they discussed the merits and limitations of the empirical approach.

General comments

My first impression of the manuscript is that this is a mixture of old problem and new technique. Such revisiting is sometimes effective, but only if the new technique provides deeper insights and/or apparently higher comprehensiveness than those in precedented studies. In my view, regretfully, I could not find enough advancements in this study; it looks like an exercise of the CCN. In other words, I am unsure whether this manuscript falls within the scope of Geoscientific Model Development.

              The manuscript is short and well-focused but need more methodological descriptions and insightful discussion. The manuscript starts from several statements about the Holdridge Life Zone, but I think this part is unnecessary. On the other hand, the authors gave few words on remote sensing of vegetation, even for validation of the estimation result.

              As the authors discussed, the data-driven approach has limitations. The model may not be applicable to the states outside the range of trained data, and the present CCN model used only temperature and precipitation as input data. Namely, it did not account for the effects of atmospheric CO2, nutrient, and disturbance, each of which is hot issues in the study area and so needs further discussions.

              I agree with the meaning of examining the potential vegetation, because natural disturbances and human impacts (e.g., land-use) are too complicated to discuss climatic impacts on global-scale vegetation. In this regard, the study is one of a few attempts to apply the machine-learning method to capture the potential vegetation. However, because of critical limitations and deficiencies, I cannot recommend accepting the manuscript for publication.

Minor points

1. Introduction: As mentioned in my general comments, Introduction starts from classic studies. I recommend putting more focus on modern and recent studies.
2. Line 72: I am quite unsure why the ISLSCP2 data were selected as benchmark data of potential vegetation and why any remote sensing data were not used.
3. Line 84: Please give references to NCEP/NCAR, HadGEM2-ES, and MIROC-ESM.
4. Line 102: Did you used daily temperature? Or, monthly?
5. Line 129: The computational times should depend on machine ability.
6. Line 172: I could not find explanation about the ‘certainty’ of the CCN output in the method sections.
7. Line 190: ‘quantity’ may be removed.
8. Line 195: Table S9 should be moved to main body. Otherwise, you may rewrite this sentence.
9. Line 203: Did you mean stand-replacing disturbances such as wildfire and wind throw? It may be better to provide several examples.
10. Line 213: I could not understand the sentence. Why the model should have better performance than you showed, if the CRU dataset had high efficiency irrespective of grain size?
11. Line 249–253: Here, you mentioned about the limitation associated with the atmospheric CO2 concentration. Indeed, atmospheric CO2 levels in 2100 are largely different between RCP2.6 and RCP8.5. So, you should make more discussions about this limitation.
12. Line 275: At this very last part of the manuscript, you first mentioned about the hardware issue (NVIDIA DIGITS 6.0).

This paper presents research into predicting global terrestrial biomes with a CNN using a correlative climate-vegetation approach. The manuscript is of very high standard w.r.t. introducing the problem and motivation, describing the approach, discussing the results and pointing out the limitations. The authors also correctly state that the presented study is not a ground breaking new innovation, but a demonstration of how existing tools can be used in the context of predicting the future of complex systems.

Besides the limitations discussed in section 3.3, a few additional aspects come to mind. Firstly, most climate models have no dynamic vegetation models built in. In addition to what the authors stated regarding the lack of feedback between vegetation and climate, it is also known that large ecosystems create their own climate and therefore changes to the ecosystems - due to whatever factor - may affect the future climate as well. It is also not clear to me how to separate the human effects that are partly, implicitly included in the model (e.g. human-made landuse changes in the training period) and, more important, the ones that are not included. Recent and future rapid development, sealing of surfaces, large-scale deforestation and irrigation, large-scale relocation of humans due to rising sea levels and temperatures, the development and use of genetic manipulated crops etc. are all factors that may influence future terrestrial biomes. It would be nice to see section 3.3 expanded to include some of this in the discussion and, if possible, to include some suggestions on how to incorporate these complex interactions in a next step.

Lastly, there seem to be a bit of a mixup of present and past tense in section 3.1 that should be made consistent. For example, Line 189-190, "The probability of the most plausible biome tend to be ..." (where it should be tends if it is present tense) versus line 184, "... the allocation disagreement was much larger ...".

The authors present a concise study about the application of the LeNET convoluted neural network (CNN) to predict global terrestrial biome distribution from climate data and then apply the CNN to predict hypothetical future biomes from climate model output. 

General comments

1. Scope: The manuscript applies and tests the LeNET CNN, which is a published and widely applied CNN to predicting biomes, which is a new application to for this particular CNN. I am not sure to what extent such an experiment falls under model development and therefore the scope of GMD. It is my understanding that validation/ model evaluation manuscripts are also permissible in GMD, but I find the model validation part somewhat lacking (see next point)

2. Validation/ comparison to other methods

The authors test the CNN predictions against true biomes, which produces an approximate 50% success rate. The authors also address the limitations of the model, such as the fact that biome change are transient and that real world biomes are much more fractured compared to modeled biomes due to human managements and climate conditions suitable to several biomes/ or plant functional types. Apart from this 1-1 comparison, there is no additional validation against other methods. The authors outline in the introduction several methods for predicting biomes (either empirically or based on pyhsiological limits of vegetation), but never address how their method compares to methods of less, similar or higher complexity. For example, does their method outperform the HLZ scheme or what else is being gained by throwing machine learning at this problem? I want to be clear here: I am not saying that this is not a valid and useful approach, but I don't think that the authors provide sufficient discussion to establish this.

3. Implementation of CNN

Based on the supplementary information, the CNN (LeNET) is run with default parameters and input parameters are air temperature and precipitation visualized as RGB images, which each image encoding a log transformed and normalized value for the two variables as color. The authors also conduct several experiments (see supplementary tables), but overall all of these have almost equal performance. I am wondering in this context, why this is the case. Is this something that has to do with the CNN that could be overcome by changes to model training/ model architecture changes or has this to do with the fact that the CNN is already extracting all the information that is extractable from the training data. I feel that this may be the case, considering that CNNs are conventionally used to classify images/photos that are very complex (such as is this a dog or a cat), while the images fed into the CNN are very simple monocolor images. Once again this is an open question that could be addressed in additional discussion.

Specific comments

Introduction: I am missing some information about what motivates this model application and why predicting future biomes using AI may be useful.

L72: ISLSCP2: Given that ISLSCP2 is potential land cover for the training, it would be good to discuss any potential issues with this dataset. Is this an unbiased representation of the true potential land cover.

L103: "mean of positive air temperature" > I am a bit confused about the positive. how are negative air temperatures treated? I would also encourage to replace positive with 'above freezing' for clarity.

L113: "he model with monthly mean air temperature and monthly precipitation had the highest test accuracy" > given that biomes are most often visualized along air temperature and precipitation axes, this does not seem to be surprising.  Humidity and SW radiation may somewhat covary with T and P. I am wondering given that the CNN allows for 3 channels, whether there is some other variable (either climate or altitude) that may be useful to add.

Section 2.3: Training of the monthly CNN. The authors should elaborate here on the procedure for using monthly data.

L190-194: I am not fully following this reasoning which seems to completly discout allocation disagreement. What the authors say may be true, but I don't think this is proven based on the information provided in the manuscript. One problem with this may be the map representation of results, which makes in depth comparisons and deep dive into potential reasons for model misses difficult.

L195: "Table S9 compared the dependence of reconstruction accuracy on combinations of climate datasets for training and test195

climate datasets" > I am a bit confused by this, given that the authors reasoned that using the same dataset for train and test could lead to overfitting and then argue here that using the same dataset for train and fit leads to higher accuracies which show robustness of the approach.

L282: "Since this method is simply an application of image classification AI, it demands much less technical skill and computer resources compared to other modern techniques such as those evaluated by Levavasseur et al. (2012), Levavasseur et al. (2013), and Hengl et al.

(2018), for example." > I am not sure that this is a fair comparison. One could similarly run a versy simple logistic regression or ANN from a standard package such as scikit-learn, which can easily be executed on a standard desktop PC.