See the attached comments.

Review of Focal-TSMP: Deep learning for vegetation health prediction and agricultural drought assessment from a regional climate simulation by Mohamad Hakam Shams Eddin and Juergen Gall

This paper describes a new deep learning model to generate pseudo observations of normalized differetial vegetation index (NDVI) and brightness temperature (BT) from numerical simulations with a terrestrial model.

General:

Overall, this is a convincing deep learning study with robust and relevant results. Nevertheless, the paper requires some major editing to make it accessible to readers and put it into context. In particular the first two sections lack information and it is then difficult to understand why the authors performed this study and why they set it up in the way they did. A lot of the missing information is provided later, for example in the discussion. Therefore, substantial re-ordering of content is required.

More specifcally, the abstract lacks some context. Not many readers will know TSMP, so it needs to be pointed out more clearly (and earlier) what the applications of this model system are and, then, why it is important to generate pseudo observations of NDVI and BT from simulation output, when real Earth observations are available for these quantities. There are some hints at the end of the abstract (climate simulations), but this remains vague and doesn't help to understand why this study was performed. A side aspect of this is that this lack of claritiy makes it difficult to evaluate the stated model errors. Are MAEs of 0.027 for NDVI and 1.9 K for BT good or better than SOTA? What would be the reference here? Retrieval errors?

Section 1 takes a couple of short-cuts and doesn't always provide good explanations to motivate this study. Please see detailed comments below. Related to this, in section 2 it is unclear why, for example, radiative transfer models are discussed here, and some of the content of this section would better belong elsewhere.

Sections 3-6 are largely OK, except for minor comments listed below.

Section 7 variable importance: I like this analysis very much. However, I think one could point out that channel importance does not "explain" everything. First: if two variables are correlated, the network may decide to focus on one of them and the other one would seem unimportant, while it could provide almost the same information if it were alone. Second: seemingly "unimportant" variables may play an important role to get the final few percent accuracy out of the models. This could of course be tested by training a model on only the N most relevant variables and compare the results.

Section 8: this discussion comes as a surprise as it goes much deeper into remote sensing and modelling issues than any of the other parts of the paper. As indicated above, there is lack of information in the Introduction and related work sections. I therefore suggest to re-arrange some of the text and use some of the material of the discussion in these earlier sections. The discussion could then be shortened and focus more on the applicability and prospects of the new method.

Even though new experiments may not be necessary, I suggest to accept this manuscript after major revisions, because substantial rewriting will be needed.

Minor comments:

Abstract l.3: why "intermediate step"? The image synthesis is the main product of the DL model, not an intermediate step in the modelling itself. The derivation of various indices is post-processing. Suggest to remove "in an intermediate step".

l.7: suggest rewording "... to assess the model's applicability to different seasons and regions..."

l.12 the unit of temperature is K, not K\degree

Introduction l.20: Suggest to remove the first sentence (motherhood statement) and integrate "under a changing climate" in the following sentence, which provides a more concise and precise start of the text. Not all droughts are extreme, and while extreme events are a good motivation, this study does not focus on extreme events, but rather tries to provide information to assess droughts or the risk of droughts.

l.32 delete "in the future"

l.33 ff. The link made here between climate models and the water cycle is a bit too direct. It is a known weakness of general circulation models (aka "climate models") that convection and rainfall are not well captured over many world regions. This is why there is a need for more specialized hydrology models, which are frequently used for regional instead of global simulations. Also, the introduction of drought indices comes somewhat unmotivated. The rationale behind these is usually to convert information from some instrument (or model) into a meaningful quantity that can be used to assess the state of some ecosystem or the climate system. Why focus on agricultural indices here? If this is intended, then this should be stated in the first motivation sentences for this study.

l.50-55 The "discussion" about retrievals is good and can be used for the motivation of this study, but it is missing an explicit reference to retrieval errors. The problems with current retrievals (or perhaps even fundamental problems = theoretiscal limitations of physics-based retrievals?) should be described more precisely and with some more detail.

l.66 I would avoid the word downstream-application here (even though it is technically correct) and rather formulate "To showcase the value (or potential) or our approach, we calculate (or derive) ..."

l.74 This sentence is confusing in the context. Before, you give the impression that you rely on the model (implicitly assuming TSMP is perfect), whereas you now state that you can use the derived products to "examine the predicitve capability" of the model. As stated in the major comment above, the Introduction needs to be rewritten with a clearer explanation what this study is based on, how it is mootivated (what doesn't work well at present?) and whta are its primary objectives. Certainly, the aspect of model errors and their impact - or the potential of the method to quantify them - are one very relevant aspect that should at least shine through in the Introduction.

Section 2.1: the review of radiative transfer models is OK, but the reader doesn't understand why there is half a page or more on cloud retrievals when this paper is about vegetation indices. It would be helpful to add an introductional sentence or two explaining why section 2 is structured in the way it is and what content is expected. The discussion in 2.1 is perhaps a little too detailed.

l.118: The paragraph introducing the work of this study does not seem to connect to the general radiative transfer discussion above. The connection appears to be only methodologically (use of AI). 

l.132 grammar "the interaction ... exhibits ... behavior"

l.162 awkward phrasing "a single indicator like NDVI excluding BT" - do you mean "... either NDVI or BT"? Or simply cut after "indicator".

l.162 ff. After reading section 2, it becomes clearer what this paper aims to do. Some of the text here should be moved to the Introduction, and section 2 should no longer explain what is done in this study, but concentrate on discussing what has been made available so far.

l.178 delete "at" before "IBG-3 institute"

l.180 "near nature realization" - what do you mean by this? Every model is an abstraction of some sort, and many models aim to produce realistic results. However, this expression is not scientific.

l.182 ff. please harmonize grammar in the bullet list - some bullets have verbs others don't

l.190 "a dynamic equilibrium" - this is not unambigious and depends on the choice of start and end date, for example.

l.195 comment, related to l.180: a free running model without DA will always be further away from "nature" than a model with DA. 

l.198 extending

l.199 wh ymention "with various vegetation types and climate conditions"? If you refer to Europe as a region, then this is kind o fobvious and doesn't add information to this sentence descrbing the model set-up. If this refers instead to a property of the model or model output, then it doesn't belong here, but in a section where you describe the data and data distributions.

l.201 and *the* model set-up

l.204 I think you could add an extra sentence to say that DL has already been applied to TSMP simulations, instead of just referring to the papers.

l.221 what are upper and lower bounds of an ecosystem? Do you mean bounds of NDVI and BT for a specific ecosystem class? Also, the sentence "Consequently, ..." doesn't fit well. Better to write "Hence, ..." or "Thus, ..." or "Therefore, ..."

l.230 Is alpha a fixed coefficient or does it vary with ecosystem class or other parameters?

l.234 delete "Moreover,"

h (l.248-254 grammar (OK, but clearly non-native English)

l.271 Please describe the data cube dimensions. Is there one datacube with (time, lat, lon) for each variable? Remove "observed"

l.275 zero

l.282 the theta doesn't belong in eq 7, which describes the mapping objective. It only comes in when you in fact use a model, i.e. when you describe the U-net.

l.300 number of channel*s*

l.325 period missing. And: pixel representations

l.328 Please be more specific: you refer to the quadratic scaling, which primarily limits the attention span, but not "applications" per se

l.337 input channel*s*

l.347 remove "a" before "one"

l.350 reduce the number of model parameters

eq.11 I suggest to replace the somewhat clumsy expressions FocalModulationBlock etc. by shorter variable names which then need to be defined in the text, of course. This would improve readibility of the equation.

l.379 "less" compared to what? I assume you mean MSE loss.

l.392 play *a* more important role

l.425 I think it would be easier to describe the U-net baseline model by simply stating what it consists of instead of "reverse engineering" it by abstracting the focal attention blocks away.

l.428 ff Please provide a few more details on the competitor models, such as number of layers, size of attention matrix etc. This could also be summarized in an Appendix, which should then be referenced here.

l.438 Do you mean "Apart from"?

l.442 Why is the second climatology computing the future? 2016 is in the past. Also, grammar: future should be singular

l.450 randomly perturbing

l.451 this seems to contradict the preprocessing description in section 3. There you wrote that samples were averaged over a week, which implies that *all* samples are used in thre average. What is written here is a random estimator of the weekly average based on two days. 

l.460 typo finally

l.466 remove comma

l.475 I don't understand the reference to radiative transfer models here and suggest it be removed.

l.484 As shown

l.486 shown *in* Figs...

l.493 weakness*es*

l.505 This discussion on errors could go one step further. While I agree that a full error attribution may be beyond the scope of this paper, you could at least give some indications, which errors come from the TSMP input data, which are from the observations and which may be DL model errors. This would be especially relevant for the TSMP data. For example, by showing the variability in different 2-day "weekly" samples and how this translates into different DL model outputs. Or you could apply some systematic perturbations to the TSMP output, thereby roughly correcting known model biases, and then see how this change the results. (OK, l.537 provides at least some indication already, and you could also refer to section 7 for additional insights)

l.516 typo "for"

l.540 , who showed

l.541 affecting

l.723 Klaus Goergen