This is a well-prepared manuscript. The authors focused on mapping several key plant functional traits in China by integrating three kinds of machine learning algorithms and climate, soil, and vegetation variables. Comprehensive experiments were implemented and all necessary technical details were properly introduced. It could be of great interest to those who are interested in trait ecology, and global vegetation modeling. However, at present, some technical details need to be added and the language of the paper needs to be further improved. In its current form, major revisions are needed before this manuscript could be accepted, thereby further improving the quality and legibility of this manuscript. The main comments are as follows:

• We are aware that the plant functional traits have strong seasonal variability. However, it seems that the issue of seasonality was not taken into account in the synthesized plant functional traits database by the authors. As a result, I don't know which time period of these estimated plant functional trait maps. Could you please provide some additional explanations regarding the temporal information associated with these plant trait maps?
• It is really good you compiled a large plant trait database with more than 50 thousand samples, spanning large geographic regions and species, please present the number of samples for each selected plant functional trait. And how many samples are for model calibration and validation?
• There are many choices of climate variable products and each product carries varying levels of uncertainty. Why did you choose the WorldClim dataset and did you assess the uncertainties of these datasets?
• I found that the time period for bioclimate variables and RAD is from 1970 to 2000, while the AI data is from 1950 to 2000 and the vegetation indices are 2000-2018 & 2002-2011, please explain why the time period of different input variables are not consistent.
• The authors used the nearest neighborhood method to resample all the input data to a consistent spatial resolution of 1 km. It is fine for the original resolution of the data below 1 km to upscale to 1km. However, Downscaling data to 1km resolution using this method is not meant for datasets with spatial resolutions greater than 1km such as MTCI with 4.63 km spatial resolution.
• Did you build separate models for each plant trait, or estimated these traits simultaneously? How did you consider the covariance of these traits when you were modeling?
• For the calculations of community-weighted mean values, you first build the relationships between the observed trait values and the input variables with 1km spatial resolution. I think your predicted values of traits present the values of 1km grid cells, so my question is how you applied CWM using the abundance of each PFT in each 1km grid cell.
• What is the ensemble model mean? how to merge the results of RF and boosted regression trees as you presented in line 263. Could you please present more details of the ensemble model?
• You just build the models between the 1km spatial resolution data and the filed measurement data. Do you think that is there any spatial mismatch between them?
• Please add the units of these traits in Table 2 although you present them in your supplementary table. And I suggest that it is better to use nRMSE in the realm of leaf functional trait estimation (nRMSE = RMSE/range of estimated plant traits).
• When analyzing the spatial patterns of plant functional traits, it is better to have a map to show the readers where the locations you mentioned in the manuscript like Yunnan, Loes Plateau, etc. are.
• For the accuracy of these estimated plant functional traits. The sampling of WD, LPC and SLA is dense, it’s reasonable that these three traits have relatively high performance. But LNC and LA also show relatively dense sampling across China as shown in Fig. 5, Could you please tell me why LNC and LA show relatively poor performance?
• I suggest that the authors may consider excluding the plant trait of SM and Height, despite their significance in many terrestrial ecosystem processes. The sampling for these traits seems too sparse to accurately represent the trait variability across the entire region of China. As a result, it becomes difficult for me to place trust in the obtained results.

This manuscript presented a study on mapping eight key plant traits at 1 km spatial resolution across China using field measurements, environmental variables and vegetation indices. Two machine learning methods were used to develop the trait prediction models. This study is well written and is interesting to the community. The trait dataset of this study has great potential to advance trait-based ecology. However, the methods are not clearly described. Also, it is recommended to perform a quantitative comparison between the trait maps of this study with those from previous studies. I hope that the following comments are helpful to improve the quality of the manuscript.

Specific comments:

Line 67: “PROPECT model” should be “PROSPECT model”.

Line 125: It is interesting to know the proportion of data excluded by criteria #3. Since the trait values of individual plant were aggregated to community-weighted trait values within 1km, including these data can be helpful to increase the number of measurements.

Line 135: SLA of sun and shade leaves can be quite different, which may lead to uncertainties for later analysis.

Line 159: Specify the full name of AI.

Line 167: The soil data was from Shangguan et al., 2013. Please justify that soil properties are time invariant, or their variation across time has little influence on the plant traits.

Lines 167-169: Were the soil properties of eight layers averaged? If the topsoil properties are important, it would be good to simply use the soil properties of the first layer (0-45cm).

Line 200: Please explain why the MTCT/MIR/etc. of January were used. They are not within the growing season.

Lines 239-241: What is the difference between the 10-fold cross validation and 80%/20% data split?

Line 252: Please describe the way of obtaining permuted values in more detail.

Line 259: It is not clear how the trait values of individual plant were aggregated with PFT to community-weighted trait values within 1km. Please describe the method in more detail.

Lines 263-264: It is not clear how the predictions of the two methods were merged. Did the authors set a threshold for the cross-validated R2? If the accuracies of predictions of one method were too low, it may not be necessary to include them.

Lines 403-417: It would be great to perform a quantitative comparison with previous trait maps, for instance, the differences between the trait maps from this study and those of previous studies can be calculated. From such maps, one can easily tell the main differences among the datasets.

This ms sets out to assess the potential for synergies between biodiversity conservation and

NCP preservation, by identifying “hotspot” areas that face high risk for both under different

scenarios. I have two major concerns with this ms, and some minor ones.

First, while the attempt to apply the novel framework of IPBES, specifically the notion of

Nature’s Contributions to People, is timely and interesting, there seem to be major flaws in the

understanding of the concept. NCP, according to the original definition by Diaz et al. 2018

(Science) and its application in IPBES (e.g. IPBES 2019, Diaz et al. 2019 Science, Brauman et

al. 2021 PNAS) are benefits or detriments emanating from biological elements (species, guilds,

vegetation patches). This ms seems to interpret direct changes in atmospheric conditions due to

climate change (increased temperature, decreased humidity, higher frequency of fires) as

changes in NCP (e.g. L 84, L 88, L 321-331 and several more) which is conceptually wrong, or

at least does not correspond to the concept of NCP. For example, an increase of CO2 (and

presumably in particulates) in the air will decrease air quality directly, not the air quality

regulation that vegetation provides by e.g. emitting certain volatiles or intercepting particulates,

which could constitute a NCP provided by the vegetation. Similarly, the fact that extreme events

such as fires, droughts, etc. will happen more frequently is different from the capacity of

elements of living nature (e.g. riverine forests, low-combustion vegetation) to regulate them

(making their impact worse or lesser). One can have the same frequency of extreme events, e.g.

floods, but different impacts by them if there are elements of nature regulating them (e.g. coastal

mangroves, well-developed riverine forests). In the methods (L 151 and following), it is not at

all clear why mostly environmental variables (temperature, precipitation, water availability,

aridity index, actual evapotranspiration, burned area) have been used as proxies of NCP.

In sum, this ms does a good job at modelling extinction risk of mammals, but a poor one in

trying to connect it with NCP. Moreover, it is highly surprising that, considering all the

prominent recent literature in NCP mediation by mammals (especially but not exclusively

megafauna), there is no attempt to address it or explanation why it has not been considered.

Second, one could recommend to strip the ms from the NCP dimension and refocus solely in the

extinction risks. But in such case one wonders what really is the main novelty that would justify

publication in GCB. The results, namely the fact that extinction risk is mainly associated to life-

history traits repeatedly identified in the literature, are not surprising, as the authors admit. Also,

the overlap of important areas for water, C and biodiversity at the global level has been very

convincingly showed by Jung et al. 2021. It is not clear what precisely are the advancements in

this respect offered by this hew ms.

The introduction is long and repetitive, with plenty of room for making tighter. Just state once

that climate change is affecting the number of species, the size of their populations, they

distribution and also many NCP, and use fires as example.

Most importantly, I have two major concerns with the quality of the source dataset:

First, it is not clear what proportion of the functional trait data is original and which is a

compilation of already published sources. If the main objective of the paper is the mapping of

NCP and the trait data are just ancillary, then this is not an issue. But if this is mostly a data

paper, then how many data re really new is a central question.

Second, I think the source dataset has not been checked/curated with enough thoroughness.

Firstly, many rows do not have a source. Second, I believe the values in some cases might have

not been re-checked thoroughly. This is a crucial issue: people are likely to use it, even if they

are not directly interested in the main focus of the paper (spatial distribution of NCP). And I

fear many mistakes or indeterminations might have slipped in it. I cannot afford the time needed

for a row by row inspection, so here are some examples of values I found intriguing in

principle; they might be correct, but they need re-checking. In some cases, I suspect there have

been mistakes in the original units of measurement, which are very common when compiling

heterogeneous datasets. In some cases the values are repeated to the last decimal place, so they

cannot be independent values.

See the following examples, all related to one trait (LA=leaf area):

I am not familiar with Juniperus squamata var. squamata, but are the authors sure this LA in the

order of magnitude of 0.01 cm2 is correct, or there is a mistake in the units?

Data ID 4079 to 4679: the leaf area of Hypoxis aurea appears to be repeated many times (same

source for all of them). Same happens with Data ID 4337-5012, for Hemipragma

heterophyllum.

It is well known that LA is a highly variable trait, but it is rare more than an order of magnitude

for adult healthy individuals of the same species, unless the environmental conditions are

really extreme. However, see the range of TWO orders of magnitude here, for the same

species, and no source:

7215 Heteropappus altaicus 0.2668

7235 Heteropappus altaicus 21.344

7261 Heteropappus altaicus 16.675

7283 Heteropappus altaicus 0.2668

A similar problem seems to occur here:

6740 Acer pictum 2401.2

6813 Acer pictum 16.675

And here (2 orders of magnitude fiference between two of the values, and three more values

are identical):

7197 Kochia prostrata 3.2016

7216 Kochia prostrata 16.675

7284 Kochia prostrata 1.0672

7308 Kochia prostrata 0.05

7796 Kochia prostrata 0.2668

7833 Kochia prostrata 0.2668

7855 Kochia prostrata 0.2668

And here:

7755 Juglans mandshurica 1.334

8705 Juglans mandshurica 119.9692

12824 Juglans mandshurica

222 Wei L P. 2014 Juglans mandshurica 714.437

And here:

2645

Krober W, et

al., 2012

Elaeocarpus

japonicus 309.65

3768

Iida et al.,

2014

Elaeocarpus

japonicus 32.0103

An

And here (huge differences, and repeated values):

779

7

Krascheninnikovia

ceratoides 0.03

780

8

Krascheninnikovia

ceratoides 1.0005

783

4

Krascheninnikovia

ceratoides 1.0005

784

6

Krascheninnikovia

ceratoides 1.0005

786

3

Krascheninnikovia

ceratoides 1.0005

788

4

Krascheninnikovia

ceratoides 1.0005

948

5

Krascheninnikovia

ceratoides 3.4831

More minor comments:

The abastract (L 20) mentiones the two dimensional spectrum of plant form and function. Note

that some of the traits used correspond to the six traits used in the original publications (e.g.

Diaz et al. 2016) to define it, but other do not. So this statement needs to be modified.

The choice of periods (1970-2000 and then 2041-2060) with a crucial 40-year gap in between, is

not justified.

L 200: “attitude” is not the right word in this context.

The random forest model is described in minute detail, but how the life history traits are used,

and why each of the traits is important to the main question is not mentioned.

Results in L 267-270 are fully confirmatory, as the authors admit.

The change in threat risk (L 294-296) appear very low (between 0.02 and > 0.04). If this is not

the case, it deserves discussion.

DATA ID 149 and 150 one leaf area is around 9 cm2 and the other 0.4.

Leaf area for Acer burgerianum several between 2 and 9 and one around 150.

Acer pictum: all LA are between 20 and 28 cm2 but there is one of around 2400

Plant functional traits must be sampled and measured according to standardized measurement

134 procedures (Perez-Harguindeguy et al., 2013) to reduce the variation and uncertainty among

135 different data sources. In this study, we included SLA measurements on both sun-leaves and shade136

leaves, WD measurements on both heartwood and sapwood of tree species, SM measurements on

137 both seeds and fruits, and plant height measurements on both vegetative and generative organs.

Enough new data?

Second, illogical values,

143 repeated values and outliers were removed, which were defined by observations exceeding 1.5

144 standard deviations of the mean trait value for a given species (Kattge et al., 2011).

The use of continuous functional trait data are proposed as a step forward with respect to PFTs. However, PFTs still seem to

play an important and unclear role in the modelling: “To calculate community weighted mean trait values, the abundance of

individual PFT within 1 km grid cell was estimated using a land cover map with a spatial resolution of 100 m. The final

community weighted mean trait values were calculated according to the predicted trait values and

corresponding abundance of each PFT.”

Cross-validation showed that the performance of the predictive models differed greatly among the plant traits and in no case were

higher than .68, and were variable and often quite low. This casts doubts on the applicability and confidence to be put in all the

results.

Not sure about the usefulness of the spatially continuous at a 1-km spatial resolution using machine learning methods in

combination with field measurements, environmental variables and vegetation indices

L 468: following sentence unclear: “In addition, due to the challenges of measurements for small shrubs

and low vegetation, WD data is mainly confined to eastern forests (Perez-Harguindeguy et al., 2013)”. The handbook

mentioned does not, to the best of my knowledge, warn about difficulties of mentioning WD in shrubs. The methods should

be pretty much the same as for trees.

Section 4 on applications is unconvincing. However, this is not crucial for the paper.

The manuscript provides a novel large dataset of 52477 trait measurements on 4291 species for eight relevant traits from 1541 sites across China, compiled from existing datasets and an extensive literature search. Based on these trait data, the authors use environmental drivers, satellite-derived vegetation indices and plant functional type association and abundance to derive high-resolution maps (1km x 1km) across China for these traits. The authors evaluate the maps.

The trait measurements and maps presented fit the scope of the journal. 

I have one major and a few minor comments.

My major comment: I was not able to completely follow the up-scaling workflow from the leaf-level data to the gridded maps. A figure indicating the different resources and steps might help.

Minor comments:

- The maps should not be called 'data', as they are rather data products. I would suggest just calling them 'maps'.

- Line 67: probably the PROSPECT model (not PROPECT).

- Line 131: The measurement date or/and time are not provided with the leaf level data. 

- Figure 4: I do not understand the values of the density axes.

- For vegetation modelling it would be excellent to additionally provide a separate map for each PFT per trait.