Chen et al. generated high-resolution (~1 km) explicit maps of above- and belowground biomass) for woody vegetation in China between 2003 and 2020. I believe the combined use of low-frequency microwaves and laser remote sensing data provides a more accurate estimation of biomass. Overall, the manuscript conducted good work on data collection, statistic analysis, and results presentation. This map could be important for monitoring and estimating woody biomass in China. It has the potential to serve as input or for calibration in Earth System Models. I think it is publishable if several minor issues can be addressed.

This manuscript introduced the annual high-resolution (~1km) maps of woody biomass from both above- and belowground in China from 2003 to 2020. National biomass maps at such high temporal and spatial resolution are very important for ecological studies, earth system modeling, forest and ecosystem managements etc. The author utilized a wide range of remotely sensed and field plots data from different sources to generate the maps. Lots of efforts dedicated to clean data, match datasets of different spatial and temporal resolutions. Limited by the inconsistency among different datasets, a number of assumptions have been proposed to in the generation of final products. Some of them are not reasonable, particularly about the “benchmark map”, and “calibration factor”. Substantial experiments and changes are needed to make the final maps more trustable. Here are more specific comments:

1. Figure 1: Too complicated to read. Summarize the main workflow instead of list all method details. More details can be described in caption and manuscript.
2. Line 110 MAD: Need more details. How were outliers detected and defined?
3. Line 110 “the average weighted by the corresponding canopy cover fractions.” How was the weighted average height calculated? Averaged over 1km pixel or ICESat-2 footprint?  
4. Line 110 “we mapped forest height over China” There are tree height maps at fine resolution over China such as Liu et al, 2022 (https://doi.org/10.1016/j.rse.2021.112844). Have you compared your mapping with these products? These height products can be used directly instead of generating a new height product.
5. Line 140 “it is supposed that the spatial pattern of woody biomass at 1 km resolution would not change much from around 2000 to 2017~2020”: It is a not solid assumption. If the spatial pattern would not change much, why do you want to estimate temporal AGB/BGB?  Land cover change, urbanization, reforestation, restoration and natural forest growth can all lead to changes in spatial pattern of woody biomass.
6. Line 145 “The extreme values (the highest and lowest 1%) were excluded as well. “ The definition of extreme values are pretty arbitrary, sometimes, 1%, sometimes 2.5% (above) need more justifications.
7. Line 160 “AGB maps are for 2017~2020, they are just used as the indicators of the spatial variability rather than the absolute values).” Not a reasonable assumption. The AGB maps for 2017-2020 are more likely to be AGB backmark-2020s rather than 2000s.
8. Line 170: “ Through cumulative distribution function (CDF) matching among different VOD products, the vegetation optical depth climate archive (VODCA) was developed(Moesinger et al., 2020).” Unclear how VODCA was developed. Did you do these or it is a product by itself?
9. Line 215 “Therefore, the mean bias of AMSR2-based VODCA data during 2013~2018 compared to that before 2012 could be estimated as the difference between the mean annual VOD calculated based on the above regression coefficients as well as LAI and VCF data during 2013~2018 and the mean value of the adjusted VODCA’s medians over that period.” Unclear. Unclear what is generated and how or why.
10. Line 220 What is “the CDF matching algorithm”?
11. Line 230 : Many paragraphs with only one/two sentences should be combined.
12. Line 245: “Calibration factor” Is the same "Calibration factor" applied to VOD maps for all years? Seems to be a very risky step. 1. bench mark AGB is not actually from 2003, but actually vary among all years 1990s to 2020. 2. The calibration factor could change among years, depending not only on tree growth difference, but also land cover type changes. Will be more reasonable to use the corresponding year (or every 5years) of true AGB to find the calibration factor for that year of VOD AGB map? 
13. Line 274-275 “AGB decomposition generally followed” The description is very unclear, I suggest author to re-write it and use equations or conceptual figures to explain. The usage grid and pixels are very confusing. Suggesting using AGB-1km grid and LC-100m pixels if this is the correct understandting. “50 pixels”: use % instead, not sure how much is the 50 pixels taken.
14. Line 280: “05, or R2 was below 0”: REALLY?? do you mean R?
15. Line 285: Unclear. “situation. Specifically, for 1/12° grids with less than 50 pixels with forests, but the pixels with shrubland are sufficient, we can reliably estimate the AGB per area shrubland as the ratio of grid average AGB to the mean shrubland area percentage in the grid.” Do you mean assuming those 1/12 grids as all shrubland grids?
16. Line 300: Another ‘calibration factor’: Do you mean the ratio of AGB map in 2017 between pre- and post-decomposition? Anything to do with 2003-2020 years? I am assuming the ratio of forest and shrubland area could change during 2003-2020. Are you using the same ratio from 2017 over the entire period?
17. Figure 4: Mean annual woody biomass is a confusing title, sounds like annually stocked woody biomass. Do you mean "averaged woody biomass in 2003-2020"? please change accordingly. same for c.
18. Line 490: what is "vegetation continuous fields"?
19. Line 490: what this the data that you compared? if not VOD, what it is? Not clear about the logic of this sentence.

Line 505: if the absolute value is incorrect, the spatial and temporal variation will be impacted too. Not a reasonable assumption.  

This study produced a long-term (2003-2020) above- and belowground woody biomass at 1-km resolution in China. Authors did a huge and impressive work to integrate multiple remote sensing data, to evaluate the spatial pattern of AGB, BGB, RSR and to investigate their dynamics. This paper provides a useful dataset for researchers to study the ecological changes in China. But there are still some comments needs to address/answer:

Main comments

1. Line 159. Why you averaged the three sets of simulations rather than put the three LIDAR-based timber volume into one random forest model? You may consider the high relevance of these LIDAR-based timber volume, but why you put three biomass datasets (GlobBiomass, CCI Biomass and GLASS-Biomass v2) into one random forest model? They are also highly correlated with each other.
2. Line 217 ‘Accordingly, by adding this bias to the VODCA VOD data after 2013, we improved its temporal continuity’, why you added the bias to compute AMSR2-VOD and to improve VODCA’s temporal continuity? did you assume there is no vegetation change between two sensors (about 10 years)? I deeply doubt this.
3. Line 410-412 AGB density in ENF > AGB density in EBF? It is opposite to our general knowledge, can you explain it?
4. Figure S5. For the decomposed AGB density in shrubland (b), high values could reach 237 t/ha, is it too high for shrubland?
5. Line 300 ‘the sum of the decomposed AGBs in 2017 was obviously different from the pre-decomposed AGB in other years’. If they were obviously different, how do you prove the robustness of your method?
6. ‘High-resolution…’ it’s hard to say high resolution for 1 km, you can directly say 1-km resolution in title
7. Please improve English in the full text

Minor comments:

1. a) ‘AMSR-2’ should be ‘AMSR2’, b) please find the rectangle written by ‘MODIS VCF (tree cover)’, the arrow near this rectangle is unclear. c) please find 3 rectangle written by ‘Calibrated SMAP / VODCA/ AMSR2…’ The data were processed by denoising and filtering, but without calibrating, so it is inappropriate to say Calibrtaed…
2. Line 86-87, ‘a slightly different algorithm’, please clarify it
3. Line 95-98, it’s different to understand the causality in this sentence, please make it clearer
4. Line 104: ‘ATLAS has three strong and three weak beams.’ But in line 107, you said ‘If the number of strong beam records exceeded 5, then…’ my question is ‘3 strong beams how to exceed 5?’
5. Line 184: ‘filtering out the high-frequency fluctuations’, you mean ‘smooth’? how did you filter? ‘other C-band VOD’ you mean ‘LPRM C-VOD’?
6. Line 194 ‘Second, we virtually filled in the data in 2015 and 2021 by using those in 2016 and 2020” why did you filled 2015 by values in 2016?
7. Line 197 ‘…after setting the thresholds of minimum distance between two peaks, peak height and dominance of peaks to reasonable values’, how did you set the thresholds of minimum distance between two peaks, peak height and dominance of peaks to reasonable values, and please clarify the thresholds of minimum distance between two peaks, peak height and dominance of peaks to reasonable values
8. Line 201-203 not good English, to be written
9. Line 204 please clarify which year/years AMSR2-based VODCA’s VOD came from?
10. Line 234 how did you resampled VODCA C-VOD from 0.25° to 1/12° (=0.083°)?
11. Line 227 ‘AMSR2 VOD’ please clarify which band, and correct in the full text
12. Line 361-364 R²=0.36 was produced by RF when GEDI-derived or GLAS-derived wood volumes involved? And please clarify R² values when you used ATLAS derived tree volume.
13. Line 429 ‘pre-area AGB’ could be modified to ‘AGB density’
14. 1 ‘ENF’s per-area AGB’ not good English, to be written
15. Line 460 ‘those improved AGB maps’ mean AGB in (Su et al., 2016; Huang et al., 2019)? If yes, they are improved than which product? Do they have higher accuracy?

This manuscript describes the development of a new high-resolution biomass (above-ground AGB and below-ground BGB) map based on remote sensing data from several optical/microwaves/lidar data sets and in situ data sets and then analyses the time changes of AGB and BGB over 2003-2020.

I found the manuscript interesting and well written and I recommend its publication after considering the following comments:

Line 82: I’m surprised by the selection of the three data sets: GlobBiomass, CCI-Biomass and GLASS-Biomass v2. For instance, Globbiomass and CCI are developed from the same group (using radar) and it could be interesting to use data sets based on other remote sensing data (as the Saatchi data set for instance).

Line 119: the authors used the “Global Forest Canopy Height 2019 dataset”. Recently released GEDI data sets could have a higher accuracy.

Line 126: I think canopy height is an effective value: it is already an average value that implicitly accounts for tree cover. So, I’m not sure it is good to multiply by TC.

Line 142: “it is supposed that the spatial pattern of woody biomass at 1 km resolution would not change much from around 2000 to 2017~2020, so the time lag problem could be ignored.”

No, I think this issue cannot be ignored: it is a key assumption here. Could the authors discuss the impact of this assumption?

MAJOR COMMENT

Line 156: “The RF model trainings were conducted in MATLAB R2021a software”.

Which RF training do you mean here?

There are many steps of calibration of many models in this study. To make reading easier, the authors should separate more clearly each steps in different subsection, indicating: what is the model? what is the input data? what is the predicted data?  and what is the data used for  calibration?

For instance, 1^st step correspond to the development of the high-res AGB for 2003.

Line 160 “averaging that is weighted by the mean R2”, could you give a reference for this weighting?

MAJOR COMMENT

Line 171: There are well-known issues in using VODCA VOD to monitor biomass over long term period (these issues are discussed here, requiring calibration corrections). So why using VODCA. For instance, the LPDR data set from the Montana University was found better for monitoring biomass. Cf Li et al., 2021, https://10.1016/j.rse.2020.112208, 2021.

MAJOR COMMENT

Line 191, I’m surprised the authors used the DCA VOD product. Li et al., 2022 found strong saturation effects of DCA VOD at high biomass level.  I would strongly recommend using MTDCA or SMAP-IB (INRAE Bordeaux). CF Fig. 9 in https://doi.org/10.1016/j.rse.2022.112921 (SMAP-IB)

-Line 195 “Moreover, we also determined the average annual number of VOD peaks for each grid after setting the thresholds of minimum distance between two peaks, peak height and

dominance of peaks to reasonable values… nearby valid values (Chen et al., 2019b”

What you mean here is not clear at all to me: why you do you estimate those peaks?

Line 245: how can you assume this calibration parameter is constant in time?

Line 358: “The training efficiency is limited by the “ :

It would be much better to use the observations used to make these maps, rather that the AGB product as input to the RF. Could the authors please comment on this issue?

(this is a big change so I do not require here for this reason),

Line 558: recent high-quality China’s forest AGB: Why not using these high quality maps as input of the RF model instead of CCI, Globbiomass, etc ?

Line 490: “A recent study revealed that the variation in VODs is correlated with not only biomass, but also soil moisture availability (Konings et al., 2021).”

To my opinion, this is not a “revelation”. This is just ONE study based ONE particular methodology and some SPECIFIC data sets. Other studies based on different assumptions and different data sets would have found very different results. From what I understood, the main issue is that an AGB data set affected by saturation issues was used to analyse VOD changes, which are not affected by saturation effects…

So, to my opinion, the paper by Konings et al. cannot be used as a reference to analyze the effects of soil moisture on biomass changes (identified by L-VOD) and it is a very unsatisfying idea to use here instantaneous SM as input of RF to model AGB.  SM can have an effect on Biomass but it is generally a long term (delayed) effect…

Minor:

-line 132 : consider revising English here

-line 144, why using ‘however” here?

-line 146: “observe only the canopy”, you mean “forest” here?

Line 170: first retrieval of VOD for biomass monitoring were shown in Wigneron et al. 1993-1995. https://doi.org/10.1016/0034-4257(94)00081-W

Line 175, the issues in merging AMSRE/AMSR2 VOD were discussed and solved in Wang et al. at X-band, https://doi.org/10.1016/j.rse.2021.112556  https://doi.org/10.1016/j.jag.2021.102609

for information, Wang et al. are developing a C-band product based on the same principle

Line 201, What is the HANTS filtering : why do you use it?

Line 206: non-tree

Line 220: Cf my above comments, there are really key issues in data continuity in VODCA. Why using it?

Line 235: why using both mean and median: they are so correlated …

Line 240: 80 000 pixels x 10 km x 10 km would do 8 M km2 which is too much for China…

Line 246: you mean here a 1/12 or 120° grid cell

Line 246-247, you assume height is constant and so biomass is constant… why not assuming directly biomass is constant?

Line 250, you mean forced to zero?

-line 255: “was considered invalid. » How many pixels are invalid

 (number and %)?

-line 467: “yet the two existing long-term datasets predicted 1.26~1.52 Pg. Moreover”. Which “existing long-term datasets’ do you mean here ??

-line 470: define clearly Southern China (province , ha)?

JP Wigneron, INRAE-Bordeaux