1. General Comments [an initial paragraph or section evaluating the

overall quality of the preprint]

This work presents a novel an exciting application of the random

forest (RF) reconstruction technique to the estimation of summer vapor

pressure (VPD) deficit from an expanded network of European and

Eurasian oxygen isotopic data obtained from tree rings.  The goal is

to place recent observations of temporal changes in summer VPD into a

longer term context.  I suggest revision of introduction and data and

methods sections to provide important missing information on the

training dataset, the RF algorithm and its modification and use here,

why and how the results are produced, their uncertainty, how they

validate against candidate semi-independent reconstructions.

Reordering of some elements of Section 2 will improve clarity and

logical sequential progression of the work.  In the results and

discussion section, I suggest revisions to clarify the features of the

results that pass validation testing, clearly illustrate the primary

features that are interpreted, and integrate discussion of uncertainty

into figures and interpretation.  In the conclusion, I suggest

revisions to more closely align these points with revised results, and

suggest that the abstract also be reconsidered once these revisions

have been made.

2. Specific Comments [individual scientific questions/issues ("specific

comments")]

2.0. Title and Abstract

2.0.1. Define "summer" at first use, and clarify everywhere that VPD

refers to summer VPD throughout the paper.  Elsewhere clarify the

season of other targeted and reconstructed variables.

2.0.2. Consider revising the title to be more specific to

the key motivation "The response of evapotranspiration to

anthropogenic warming is of critical importance for the water and

carbon cycle." and associated conclusion: "Based on our

reconstruction, we show that from the mid-1700s, a trend towards

higher [summer] VPD occurred in Central Europe and the Mediterranean

region which is related to a simultaneous increase in [summer]

temperature and decrease in precipitation."  However, also see notes

on Data and Methods, Results, and Discussion below, as this statement

is not clear from the results and analysis presented, in particular,

Fig 6.

2.0.3. For widespread further analysis and impact of the contribution,

please also deposit reconstructions in a public long-term archive such

as PANGAEA or the NCEI/Paleoclimatology repository, and provide the

URL.

2.0.4. 26 observational sites is a sparse network for the development

of a gridded reconstruction over Europe.  Add a sentence to the

abstract stating the level and the nature of the independently

validated skill of the results, specifically summarizing what is shown

in Fig 4 and revisions to Figs 5-8.  See also notes below on Data and

Methods and Results to further refine the content of Figs 4-8.

2.0.5. With respect to the conclusion described in item 2.0.1, add a

sentence to the abstract to explain why the authors think the summer

VPD over the region is increasing since the mid-1700s.  What changed

after this time to cause the observed result?  What is the forcing,

and is this the expected sign and amplitude in the response?  I am not

sure I find the development of these results in the manuscript in its

present state, but this would be a welcome expansion.

2.0.6. The results are based on o18 but a growing season VPD effect

ought to also be observed in c13 (e.g. Farquhar et al, 1989,

https://www.annualreviews.org/doi/abs/10.1146/annurev.pp.40.060189.002443

; see also Siegwolf et al 2022, previously cited).  If the records

exist, do they corroborate the result?  If not, why not?

2.1. Introduction

2.1.1. l. 53-55: "For instance, studies have shown that the [summer?]

VPD has been increasing sharply at a global scale since the year 2000

(Simmons et al., 2010; Willett et al., 2014; Yuan et al.,

2019). Spatially explicit [summer?] VPD records derived from remote

sensing data cover only 55 the last ~50 years and vary in quality, so

long-term perspectives of VPD variability are lacking."  Revise to

cite the remote-sensing and reanalysis-based results more specifically

(e.g. Yuan et al 2019 is a remote-sensing based study) so the reader

can easily learn more about each of these sources and corresponding

results.  Plot these records in Fig 6 for comparison with the

reconstructions.

2.1.2. l. 55-57: can summer VPD may be estimated from available

historical reanalyses and other gridded products?  Yes, as described

in section 2.4.  It would be good to revise this sentence to indicate

the potential to do so, and possibly to evaluate the results from

publicly available products. Plot the VPD estimate from the 20CRv3 in

Fig 6 for comparison with the reconstruction.  Do they agree or

disagree with results from the more recent reanalyses and

remotely-sensed estimates?  This would help better establish the

validity of the o18 training target.

2.1.3. l. 62-63: This sentence needs support and development, perhaps

its own paragraph: why recontruct European summer regional VPD?  For

instance, is it because there is a significant modern regional summer

VPD trend that is projected to continue and endanger the health of

European forests?  In what season is the trend observed?  This would

also support the argument that we need high resolution records

(l. 62-63), at what resolution, and for what season.  Consider adding

a figure that shows this motivating problem, including the

uncertainties in the different sources of information

(remotely-sensed, modern reanalysis, historical reanalysis) - this

will support the development of the present study in the present

region at the reconstructed resolution.

2.1.4. l. 64-65: pursuant to item 2.1.3, tree-ring records generally

record environmental conditions with a specific growing season and

possibly antecedent seasons.  Revise the sentence to add this point,

as it is pertinent to item 2.1.3.

2.1.5. l. 67-70: You might also cite the recent book on stable

isotopes in tree rings, [Siegwolf et al, eds, 2022,

https://link.springer.com/book/10.1007/978-3-030-92698-4] For this

point, I think the same could be said of carbon isotopic composition,

based o the same cited references.  Explain the model for c13.  Could

c13 data be used to reconstruct growing season VPD?  Why/why not?

This helps make the case for use of the o18 dataset, and/or it might

provide independent corroborating evidence in support of the results

presented.

2.1.6. l. 73-76: Explain why variations in the oxygen isotopic

composition of atmospheric water are not a confounding factor for

summer VPD influences on o18 of tree rings in the European region.

For example, a change in subtropical vs temperate or subpolar airmass

might masquerade as a change in local evapotranspirative demand.  See

also item 2.1.5: c13 of tree rings might enable you to distinguish the

latter from the former effect.

2.1.7. l. 78-83: Revise to precisely define the novelty of this

contribution: expansion of the ISONET dataset (l. 111-115; Figs 1,2)

and its use to reconstruct VPD using the random forest reconstruction

algorithm.  To clarify this, expand the sentence here to specifically

describe the novelty of this work, and in Figs 1 and 2, indicate the

new data series vs the ISONET series with differences in symbols,

colors or otherwise as best clearly shown.  Or you may choose to cite

Table 1 which provides this information.

2.1.8. l. 85-88: Expand the introduction of the RF reconstruction

method [but check if this happens sufficiently in Section 2].  Is it

particularly skillful for reconstruction of large-scale spatial

patterns from sparse observational networks?  If so, cite examples of

this, as it will help you defend the technique and its application

here.

2.1.9. l. 99-101: I think the sentence is saying that the regions of

analysis are defined in the IPCC AR6 WG1 report.  If this is correct,

then please add the particular WG1 Chapter and ideally the figure in

which these regions are defined within, and refine the cited reference

accordingly.

2.1.10. l. 104: Since there is validation for extreme year events,

potentially add summary of these results and conclusions to the

abstract.  This could be an important result, speaking to the

motivation for the proposed study, especially if it means that

European forests might have survived past extreme VPD deficits, but

that conditions are different now with the anthropogenic climate

change effects superimposed? [check: do not get ahead of authors on

this]

2.2. Data and Methods

2.2.1. l. 127-132: The different sampling and pooling methods for

developing sitewise values implies differences in the observational

uncertainties assumed for the RF reconstruction.  Also see items 2.1.3

and 2.1.4 regarding the season for which information is encoded in the

tree rings.  Do you take these differences into account or do you

assume the same observational uncertainty in the various series?  Are

the results sensitive to these assumptions?  Demonstrate if and

how/why or how not/why not.

2.2.2. l. 132-133: I believe ISONET had a ring comparison of precision

and accuracy of o18 analyses across the contributing labs [Boettger et

al 2007, https://pubs.acs.org/doi/full/10.1021/ac0700023]; did the

labs producing the expanded dataset also participate in this study?

This might be important if there are differences in the amplitude of

variation across labs for working standards across the range of

observed o18 values.  This is likely to be small compared to intersite

differences in VPD amplitude of variation (e.g. Table 1) and

observational uncertainty in the wood o18 records themselves, but it

might be acknowledged that these interlab differences are assumed

negligible.

2.2.3. l. 150-154: Clarify and expand: For what purpose have you

extracted gridded values from 20CR product?  For what region?  If for

the summer season (which needs to be defined at first use of the term

summer), then why are the other seasons of interest?  (see also items

2.1.3, 2.1.4). Also please introduce the calculation of VPD from the

reanalysis variables (see also items 2.1.3, 2.1.4). --> This item can

be resolved by moving section 2.4 here (l. 158-159; section 2.4).  Use

the 20CRv3 ensemble statistics to estimate the uncertainty in the

products and estimations based thereupon (see items 2.1.3 and the

potential to add uncertainty into the RF training, item 2.2.6 below).

2.2.4. l. 159-160: Clarify: Missing data are infilled for which

dataset?  Or both?  if this sentence refers to missing data in the o18

dataset, then it belongs at the end of section 2.1.  If in the 20CR

data, it belongs in section 2.2.  If both, then it might belong where

it is, but clarify the infilling procedure is performed on both

datasets, and the percentage missing data in each dataset separately.

2.2.5. l. 165-166: Revise: these restrictions might be moved to

sections 2.1 and 2.2 and reference both Figs 1 and 2, as they arise

from the tree-ring o18 network observational restrictions.

2.2.6. l. 164-172: Revise and expand: Michel et al (2020), the

algorithm adapted for the reconstruction, should be specifically cited

at l. 95.  And the description of the RF algorithm needs to be

specifically introduced in section 1, and then the modifications more

specifically described here in section 2.3.  Generally: How does the

RF algorithm work?  What distinguishes it from linear approaches? What

kind of data requirements or assumptions need to be made?  How are

reconstruction uncertainties estimated?  Specifically, expanding from

the modifications described for nesting, are the nests normalized

against each other to counteract heterscedasticity associated with

changing numbers of observatonal records, and their individual skills?

How is the 1x1 degree resolution of the reconstruction target

established and validated for reconstruction by the sparse

observational network of o18 data?  How is validation established (are

a fraction of data withheld for validation for realizations, and if so

how and what fraction?)  How are the reconstructions aggregated?

2.2.7. l. 170-172: CE is a measure of resolved variance and mean

estimate in an independent period; also consider reporting other skill

statistics such as statistics of the ensemble reconstruction

validation bias, correlation, root mean square error.  Report also

statistics for the calibration period to demonstrate the presence or

absence of artificial skill (training statistics much better than

validation statistics) in the results. [Check: is this done in the

results in section 3?]

2.2.8. l. 173-177: Explain in the introduction (see also items

2.1.2-2.1.4) why these comparisons are useful, as they compare

reconstructed VPD to reconstructed temperature, precipitation and

PDSI.  Why should these different climate variables covary with VPD?

As the argument is made in Section 1 that the VPD reconstruction is

novel, also explain, how and why does VPD differ from T, P, PDSI?

Here also explicitly state if the reconstructions are fully

independent from each other.  I am guessing that the T, P, PDSI

reconstructions do not use o18 data, but if they target and are

trained on summer European observations of T, P, PDSI, and the 20CR on

which the VPD reconstructions are also based assimilate observed T

(but probably not observed P, calculated PDSI), then to what extent do

the VPD and T (and possibly P, PDSI) reconstructions correlate because

of these common training data (e.g. Fig 7a,b,c)

2.2.9. l. 177-180: This is a useful exercise, but explain how the

marker years were selected.  Produce a similar exercise, but for the

same selection criterion applied to the 20CR and modern VPD estimates.

This may be useful as a validation of the stationarity of the spatial

patterns in extremes, and discussion of the physical mechanisms

consistent with the patterns, and/or a discussion of the

reconstruction uncertainties and the significances of plotted

diagnostics.

2.2.10. Move Section 2.4 into section 2.2, addressing items 2.2.3, and

you can address items 2.1.1-2.1.3 with the results.  Also the

production of the seasonal averages (l. 195-197) needs to be justified

by the available data and skillful representation of T and RH for

those seasons, and the seasonal representation of the reconstructed

seasons by the tree-ring o18 observations.  I was confused here by the

mention of calculation of non JJA seasons, as if these were input to

summer (JJA?) reconstruction target.  Please clarify if the only

season of 20CRv3 based VPD that is used in the present study is JJA.

How is site information trained non-locally against the possibility

that a site growing season and representation might be, for example,

MAM and JJA, but this is used to reconstruct for a grid point for

which the growing season response might be the same or different?

Here it will again be useful to explain the RF methodology in more

detail (items 2.2.6, 2.2.7) and to produce figures with diagnostics of

the training vs validation ensembles.  This could also be expanded in

discussion: what do we learn from the RF fitting process about the

information contained in the tree-ring o18 records?  This is

potentially another important contribution of the work.

2.3. Results.

2.3.1. l. 201-202: Revise associated results and figure 3: produce the

correlation map for (a) statistics of the training ensemble and (b)

statistics of the validation ensemble: r, bias, RMSE, CE, using the

ensemble of the statistics to estimate statistical significance of the

results and mask nonsignificant results.  Otherwise this presentation

conflates training and validation statistics by including the entire

1900-1994 period used for both training and validation (as yet

unspecified how; see items 2.2.6, 2.2.7).  If they are different,

then we might have an artificially better view of the skill in the

reconstruction, because by definition, the training results have to be

skillful to some extent.  If necessary, revise the associated results

in subsequent sentences of section 3.1.

2.3.2. l. 202-203; Fig 3.  Explain how the correlations are produced.

Are they for the *nearest* gridpoint of VPD estimated from the 20CR

data?  Superimpose the reconstruction grid.  To the extent that this

is not gridded VPD reconstruction, but rather a demonstration that

there is correlation between o18 data and local (nearest gridpoint?)

computed VPD, this figure could more usefully go into section 2.  If

expanded to produce a reasonably analogous training and validation

period, like is done in the RF reconstructioh, the figure could also

be expanded to provide: r, bias, RMSE of standardized values, and CE

of standardized values, for the validation period.  Or these results

could possibly be put into an expanded Table 1.

2.3.3. l. 223-235: Important elements of the methods are given here,

but would be valuably moved to section 2 and respond to items 2.2.6

and 2.2.7.

2.3.4. Add site locations to Fig 4 to help demonstrate the skill local

to th sites, and decline in skill away from them.  I think the

resolution of the target field (1x1 degree) is represented in the

pixelation of the colors.  Is CE normally distributed, or would it be

better to plot the median estimates?  Or the median estimate for which

the 95th percentile confidence interval does not include the null

hypothesis?  I am still left without enough explanation of how

training and validation is established away from the site-nearest

gridpoints, and this should be addressed in Section 2 (items 2.2.6,

2.2.7).  I realize much of this may be developed in Michel et al 2020,

but for use here and its adaptation, methods and signficance

estimation should be briefly summarized in section 2 so the paper

stands on its own.

2.3.5. l. 241-255: To fully understand the results for skillful

nonlocal reconstruction (compare: Figs 1 and 2 to Fig 4), we need more

information about the nature of nonlocal training and validation.

This should be described in section 2 (see items 2.2.6 and 2.2.7).

2.3.6. l. 258-260: The use of +/-2 standard errors implies that the

errors are normally distributed; is this the case?  Better might be

the 95th percentile confidence interval, if not.  See also item 2.3.4.

For the error calculation, this should be done for the validation and

training ensembles separately (see also item 2.3.1 and 2.3.2 notes on

Fig 3 as well).  The use of RMSE here also suggests support for

revising Fig 3 in line with items 2.3.1 and 2.3.2.  Revise Fig 5 to

produce the same statistics and for the validation ensemble of the

statistics.  If the use of only validation statistics is already what

is shown, at l. 201-202 and l. 258 and elsewhere the 1900-1994

interval is noted, please clarify this.

2.3.7. Fig 5: I am unclear as to how to interpret Fig 5 results. I

would revise the figure to show the median RMSD for gridcells for

which the 95th percentile CI does not include the null hypothesis,

masking all other gridcells inthe domain.  Put the o18 record

locations into the figures to help make the case that the skill is

highest nearest the o18 record locations (also Fig 4; l. 241-255;

l. 256-262).

2.3.8. Section 3.3 and l. 271, 275, 292, elsewhere: define beta and m

in Section 2, methods: I believe in all instances, this is a linear

regression slope coefficient, is this correct?  But clarify if these

estimates are regressions on 30 year rolling averages (could they

instead be center-weighted 31-year averages, e.g. a Hamming window or

similar?) Also give the errors on the slope, and plot the regression

fits and the 95th percentile CI error envelopes on Fig 6 for their

evalation by the reader.  

2.3.9. Fig 6, l. 303-305: Clarify: are these averages area-weighted,

and are these averages with nonsignificant results masked out?  From

panels (a,b,c), it appears the answer is no, but I would recommend the

authors indicate the subregions for which VPD is skillfully

reconstructed, and see if the resulting timeseries are different for

skill-masked and non-skill masked estimates.  From comparison with Fig

4, I would recommend the former in revision.  The same masking might

need doing for the P, T and PDSI estimates given their reported

validation skill scores.

2.3.10.  Fig 6, error estimates on time series: based on Figs 1-4,

unclear about 5, I am skeptical that the errors on the time series are

accurately calculated; they seem too small, given results in Fig 3.

Are the nonrandom errors of estimation from the nonlocal estimation

used?  After all there are only 29 data series and many more

gridpoints in each region at 1x1 degree resolution.  Are the errors of

validation included?  Do they vary with the number of o18 series

available over time?  They probably should, according to Fig 4.

Details of the error estimation need to be added in Section 2 (see

also items 2.2.6, 2.2.7).

2.3.11. Fig 6: the presentation of timeseries for like regional

averages of reconstructed VPD: T, P, PDSI: VPD here are reconstructed

anomalies, relative to a historical mean?  If not, explain how the

mean was estimated (section 2), and give the reconstruction

uncertainties for mean estimation.  If for anomalies, give the period

for which anomalies are relative and give details in Section 2.

2.3.12. Fig 6 and l. 293: put error estimates on the 1652 maximum VPD.

2.3.13. l. 290-299: This presentation of results needs introduction in

section 1.  Are the dates relative to solar activity periods

(e.g. Maunder Minimum) and temperature anomalies (e.g. LIA) defined by

the reconstructions or by independent estimation?  It should be the

latter, and these definitions might be usefully described in the

introduction, otherwise they present as potentially ad hoc.

2.3.14. Following from item 2.3.10, it is unclear to me that the

results support the interpretation at l. 292-299, given the decaal

variability and questions concerning the significance of linear trends

on decadal timescales given reconstruction uncertainty.

2.3.15. Fig 7 and Fig 6, Section 3.4.  For Fig 7 and results presented

in the test in section 3.4, are the correlations for specific seasons,

or annual averages?  Specify.  The results in Fig 6 for 30y means

suggest little of the skill is at these timescales.  Calculate

correlations for 31 year centered means, and estimate significances

for the reduced degrees of freedom. Also, since the reconstructions

are all trained on modern observations, which may be non-independent,

it would be good to calculate all correlations for Fig 7 and on

decadal timescales in Fig 6 for the pre-training periods, which I

believe in all cases might be (please check): before 1850.  This is a

more conservative validation test and might resolve the apparent

inconsistency between Figs 4 and 6 and Fig 7, which I don't

understand.  These questions might require some revision of the

sentence at l. 340-344, and see item 2.3.16.

2.3.16. Fig 7, Section 3.4: VPD is in general negative correlated with

P and PDSI, and positively correlated with T2M.  Explain why, here or

in section 4, with reference to section 1 and item 2.2.8, and the

physical links between temperature, precipitation, vapor pressure

deficit and the Palmer Drought severity index, which includes both

temperature and moisture influences.

2.3.17.  Section 3.5, comparison of Figs 6-7: I am having trouble

making sense of this section, because the physical links between the

variables reconstructed have not been introduced (item 2.3.16), and I

am not sure if the changes in the correlation with time and frequency,

on different timescales, do not arise from reconstruction

non-independencies and uncertainties that vary in space and time

(observational density), and frequency (skill?).  I am not sure how to

suggest specific constructive revisions, except to suggest that the

organization by subregion be retained; presentation of the physical

linkages (item 2.3.16) be introduced and explained, checking and

quantifying uncertainties (items 2.3.15, 2.3.1-2.3.12 and

cross-referenced earlier items) be performed, before revisiting and

synthesizing the most reliably estimated features and independent

comparisons with assqociated but not-the-same reconstructed variables

(Figs 6-7), organized by region as is done here.

2.3.18. Section 3.6, Fig 8: This section needs more detail on what was

done.  What does "at the European level" mean, is this an

area-weighted average over skillfully reconstructed gridpoints, for

what season, for the pre-training interval, and considering the

changing uncertainty over time?  These details could go into Section

2, such that the discussion of the results can be done here.  I think

wettest and driest refer to VPD here, but perhaps revise to call them

the three lowest and three highest reconstructed VPD years, estimated

as the average standardized anomalies over all seasons and all

skillfully validated gridcells.  Because in panels (a-c) the sign

changes across regions, skill masking might be important, and the

definition of the extreme years is also important.  If the results do

not change with these considerations, what is striking to me is the

Scandinavian/western European VPD dipole (panels b,c; but not a); and

the consistency in the negative VPD anomalies across all Europe in

panels (d,e,f).  The authors wrote about these features, but maybe

could expand the discussion: are there different mechanisms causing

these two different results?  If so, why?  Perhaps this could be an

organizing question for the discussion of these results.

2.3.19.  Section 3,6, l. 401-406, discussion of the potential for the

Laki eruption to have caused the 1785 extreme events, this seems a

consistent discussion (but I think should refer to Fig 6, not Fig 5),

but it could more comprehensively address the volcanic forcing of VPD,

T, P and drought, by showing a composite of the VPD, T, P and drought

anomaly responses in the years subsequent to all the strong volcanic

events within the time interval of reconstruction.

2.4. Limitations of the reconstruction

2.4.1. In general, revising with respect to items 2.2.2, 2.2.3, 2.2.4,

2.2.6, 2.2.7, 2.2.8, 2.2.10, 2.3.1, 2.3.2, 2.3.4-2.3.10, 2.3.15,

2.3.16 will help support and expand the points here with specifics in

figures than can be referenced.

2.4.2. Alternatively to item 2.4.1: the points in Section 2.4 might

usefully be made in the figures by the suggested revisions, and used

to contextualize the discussion and interpretation of the results, as

suggested in earlier items.

2.4.3. Section 4 could be replaced by a Discussion section.  This

could comprise the central argument of the paper that supports the

statements made in the Conclusion and the Abstract.

2.5. Conclusion and outlook

2.5.1. l. 444-451: Revisit these statements after assessment of the

timeseries with masking for regions of nonsignificant skill, and the

assessment of coherence (correlation as a function of the key

interpreted periodicities, specifically: seasonal, interannual, and

approximately multidecadal.

2.5.2. l. 450-452: "Central Europe and the Mediterranean regions

reveal stronger trends of increasing VPD (highest VPD on average and

highest VPD variability) which can be explained by a precipitation

decrease and a temperature increase."  It is unclear to me how this

summary statement is supported by the results presented in Figs 4-7,

and associated discussion.  What is meant by "trends"?  The

anticorrelation and correlation (or nonlinear associations) between

VPD, P, T (and PDSI), and changes in variability, are not apparent

from the analysis presented in Fig 6 and Fig 7.

2.5.3. l. 452-457: "Our results underline that the European VPD values

have increased over the last decades. Based on the obtained long-term

perspective, we find that this increasing trend in Europe has not

started in 2000, but has already begun a few decades after the Late

Maunder Minimum with a simultaneous increase of the temperature in the

mid-18th century. In the historical context, 455 however, this

long-term trend is unique in magnitude and persistence over the last

400 years. Moreover, the results from our study imply that vegetation

in Europe has been subject to an increase in VPD for a longer period

of time, but that increase has been significantly amplified by recent

climate change, especially in Central Europe and the Mediterranean

regions."  It is not clear to me that these statements are supported

by the results and analysis presented.  Clarify and illustrate the

results in support, in Figs 4-7, addressing prior questions concerning

estimations and uncertainties, or else modify the statements to be

more specific and accurate.  For example, it is not clear that

European VPD values have increased everywhere - not in northern Europe

where there is considerable skill in part of the region (Fig 4, Fig

6).  And in no region is the recent VPD higher than in previous

decades and centuries (Fig 6).

2.5.4. l. 457-461, I think the first half of the sentence is supported

by the logic of the cited references, but the second half depends on

the effect of an increase in VPD on vegetation in Europe since the

Late Maunder Minimum, and this is unclear from the results presented

in Fig 6.

2.5.5. following l. 461, this would be a good place to summarize the

expanded results (see items 3.3.18, 2.3.19) from the analysis of

extrema in the reconstructions and the mechanisms that might have

caused them.

2.5.6. l. 462-467: I think this summary of the validation makes the

case for interpretation of the results, so it belongs at the beginning

of the summary, not at the end.  It also needs to be expanded to

consider the results of further uncertainty quantification as

described in the preceding notes.

2.5.7. There is no summary or discussion of the RF reconstruction and

diagnostics, but there should be, as this is an important new

application of a nonlinear reconstruction technique.

2.5.8. l. 465: Here it seems that summer VPD is what is constructed -

this should be clarified in response to earlier items (e.g. 2.1.3,

2.1.4, 2.2.10).

2.5.9. l. 467-469: the logic of this sentence is unclear.  How does

the gridded reconstruction of European summer VPD minimize statisitcal

uncertainties of historial VPD variability?  How does it provide

long-term context?  (But in Fig 6, if the 20CRv3 VPD and the other

estimates of VPD for these regions, it could do so.)

3. Technical corrections: [compact listing of purely technical

corrections at the very end (typing errors, etc.).]

3.1. l. 62: "Therefore, appears the necessity..." --> "Therefore, it

appears necessary to ..." and see also item 2.1.3.

3.2. l. 71 and throughout: "tree ring-d18O" --> tree-ring d18O"

3.3. l. 80: "opportunity to develop for the first VPD reconstruction

at the European level..." --> "opportunity to develop the first VPD

reconstruction at the European level..."

3.4. Section title might be more generalized: "2. Sample sites and used

climate data" --> "2. Data and Methods".

3.5. l. 403, "favorizing" --> "favoring" 

3.6. l. 465: "f" --> "of"

3.7. l. 466: "d18" --> "d18O"

The reviewed article presents a four-hundred-year reconstruction of the summer vapour pressure deficit (VPD) in the European area. For this purpose, it uses data from previously published research, including the outcomes of two large European projects ISONET and MILLENNIUM.

The research presented has a clear objective, the implementation of which is presented in the following chapters of the reviewed work. The significance of the research undertaken is discussed first. VPD affects the development and growth of plants and thus the life of entire ecosystems. Moreover, the VPD index is an indicator that can be applied to assess the forest mortality, drought occurrence, crop production and wildfire incidence. However, the measurement data of VPD have a very short time range. Therefore, proxies are needed to analyse VPD changes in a long-term context. In this article the measurements of stable oxygen isotope content in cellulose extracted from wood (within individual annual growth-rings) were employed. The data used came from different sites located in Europe, allowing changes in VPD to be traced both over 400 years and across the continent.

In order to reconstruct VPD using proxies, a suitable mathematical model have been developed and then validated. Statistical verification of this model was also presented. The results showed the correctness of the reconstruction performed. On the basis of obtained effects it was possible to relate the contemporary upward trend observed in VPD values into a long-term perspective. Moreover, during conducted research three areas differing in magnitude and in course of VPD values over time were also distinguished: Central Europe, Northern Europe and Mediterranean. The received results were then compared with earlier reconstructions of other meteorological parameters carried out for the last 400 years for European areas. In addition, they were confronted with historical record. New data correlated well with both the earlier reconstructions of temperature, precipitation and Palmer Drought Severity Index, and also showed a high degree of correspondence with the historical evidence.

The reconstruction carried out was based on a large set of measurement data, covering a long period of time and having a large territorial range, was well conducted in terms of the statistical methods used, and was then placed in the context of historical data, and compared with other available reconstructions of climatic parameters.

The work is of great value and tackles a hitherto underexplored scientific problem. The structure of the manuscript is correct, the individual chapters contain a detailed presentation of the material, the research methods utilised, an extensive discussion of the results and a comparison with previously published research data. Both the aims and conclusions are clearly formulated. The results of the presented research constitute a large package of new data that can be used in the future in various fields of science.

I evaluate the work very positively, the only weakness being the figures, the size of which is not always appropriate and does not allow for a comfortable reading of the data presented. I have included the rest of the minor editorial comments in the text of the manuscript (please check that the text highlighted is definitely correct).

The authors have developed a really nice new data for vapor pressure deficit across Europe from 1600-1994, which was spatially gridded. 

The paper was clear and well-written, and should be published after some minor corrections and elaboration of some points. It was enjoyable to read this manuscript, and the authors have done a great job communicating their results. 

The authors do not discuss how changing δ18O of precipitation can influence their record. The correlations to VPD are robust, but could much of the unexplained variance be due to changes in precipitation δ18O? This could be tested by looking at high-resolution speleothem records of δ18O. Bunker Cave, Germany is one example (https://www.mdpi.com/2076-3263/11/4/166), and there should be others. Even over the modern interval, precipitation d18O from the IAEA data set may help place some of the tree-ring d18O records in context, and this would be a useful and welcome addition to the paper.  

I would also like to learn more about the significance of VPD in general for agriculture, ecosystems, etc., in the Introduction. It would also help make the paper have higher impact. Elaboration of why this new data product can be useful will be a good addition to the revised manuscript. For example, testing between a temperature, precipitation, or combined response to produce the VPD would be interesting. 

L75: If leaf δ18O is a function of leaf-to-air VPD, how then does the δ18O signal get transferred to the tree rings? Lines 201-202 help show it is real, but how does it actually happen within the tree? 

L80, delete “for” 

96: how do you test whether 26 series are enough to extract the signal? 

L272 and elsewhere: the reference to the solar irradiance events in L272 was abrupt, and not thoroughly contextualized in the rest of the manuscript. Provide some more context of why you are comparing the reconstructed VPD to irradiance, to give some idea of why you would expect there to be a relationship in the first place. If there were a reasonable correlation, showing the irradiance events on Figure 6 would be helpful. If there is not a clear correlation to TSI, then it would be better to not use TSI as a climate template against which to interpret the VPD record. 

L324: severity, not sensitivity

L336: there are many more types of paleoclimate data than just tree rings. It would be nice to avoid the silo effect by checking into the high-resolution speleothem or lake records of δ18O, because they would be a useful check on the changing δ18O of atmospheric precipitation. Relatedely, what range of δ18O variability is seen in the tree ring records? Including a representative example (or subset) would be useful. A few permil variability would be typical of many speleothem records (and likely in precipitation). How large are the variations in the tree rings?