This manuscript by Dong et al. provided an intriguing study on quantifying soil organic carbon (SOC) at various depths of 0-20 cm and 0-100 cm in China from 1985 and 2020. The SOC change is highly important for the terrestrial ecosystem carbon cycle. It’s innovative to use climate zones to improve the performance of this model, with comparable accuracy to other published SOC datasets. In addition, this dataset expands the temporal availability of SOCD in China at high spatial resolution, which can be highly applied in related studies. Overall, this manuscript has clear scientific innovations in understanding SOC changes. Please find my detailed comments below.

1. Section 3.2. Could you give more explanation about the principles of selecting variables? For example, from Fig.2, the R between AH and SOCD is almost 0, why select this variable? And only 18 variables have been shown on Fig. 2 without CLCU, how to select CLCU as an input predictor?
2. Section 4.2. Fig. 8 and Line 250: The discussion of different features for SOCD estimations is comprehensive, which can help us to understand the important factors of SOCD variations. But it’s very interesting to find that the features have different important values in the two depth models. Please try to discuss more about these differences.
3. Section 4.5. Fig. 13 and Line 315: “This may be the result of the topsoil being more susceptible to the direct effects of soil management practices and environmental changes.” Which types of management practices contribute to the changes of SOCD in topsoil? Please add more details (policies or references). As shown in Fig. 13(b), the SOCD estimation in 0-100 cm from this study has a higher value than others. Please add some validation for SOCD in 0-100 cm as mentioned previously. In addition, the SOCD in deep soil should increase if SOCD in topsoil increases. So, please give possible reasons for SOCD in 0-100 cm to be stable from the 1990s to 2020s. Fig. 14 (d) and Fig. 15 (d): In Xinjiang province, the SOCD in 2000-2005 seems to change a lot when compared to another period. Is this due to the model itself, or has some event happened during this period to make a significant change in SOCD? Please give reasonable explanations in this part.
4. Section 2.1 “brown soil, brown soil”. Duplicate
5.   Section 2.2. Line 95: The SOCD data from Song or Xu? Please check it carefully.
6. Line 125: Generally, the spatial interpolation results are reliable if stations are evenly distributed. How about the spatial distribution of these meteorological data used for interpolation?
7. Line 130: Please add the produced time or effective period of the published soil datasets.
8.   Section 4.2. Line 225: There is no need to write the full name of the statistical metrics, which have been mentioned previously. Fig. 6: Could you add the sample number in Fig. 6? Please add unit for RMSE both in Figures and the manuscript.
9.   Section 4.4. Fig. 11: Please add a unit for colorbar for (b), (d), (f), and note the Time (which year). Is it the annual average or any specific year? Please add the validation results for 0-100 cm SOCD in the manuscript or Supplementary.

The manuscript attempts to produce a 1km SOCD data product of China during 1985-2020. Yet, the novelty of the manuscript is very limited, and the reliability of the produced data product is still difficult to judge, since the manuscript has many inaccurate descriptions on data source and the methodology. My specific comments were given below:

(1) Feature optimization? I think it should be feature selection. Yet, random forest (RF) may represent extremely complicated nonlinear relationship between SOCD and their drivers (i.e., the covariates that you used), why did you select features based on Pearson correlation coefficients? Besides, RF is some insensitive to feature selection!

(2) The use of climate zone in this study is really unnecessary since the temperature and precipitation that you used to define the climate zone have already used as features in your RF model!

(3) How did you separate your train and test samples? How many samples for each of the year? is it a balance sampling across years? This information is very important, since the readers want know if your extrapolation beyond the year of observation.

(4) The descriptions for building your space-time RF model is very confusing! I think your RF model should be a space-time model, otherwise, you can not get time series of SOC from 1985 to 2020. Or you just model the SOCD during each time period separately？if that was true, this manuscript would have no any novelty. if you built the space-time RF model through space-for-time (see, Heuvelink et al. 2020. Machine learning in space and time for modelling soil organic carbon change. Eur J Soil Sci.),  are the covariates like vegetation and land use considered as “dynamic covariates? how did you represent the lagging effects of dynamical covariates (i.e., the effects of temperature on SOC state is lagged, vegetation and land use as well), and the memory effects of SOC (i.e., the state of SOC in this year depends on last year)? these information is essential for modelling changes or dynamics of SOC using machine learning (ML) method like RF, as the ML is pure data-driven method.

(5) the validation across different time period is missing, thus, it is difficult to judge the trend in SOC change.

(6) Source of data is confusing. How many soil profiles for each of the year, as we should check the balance of data across time. Your DEM was generated from topographic maps or resampled from SRTM DEM? are weather data monthly or yearly? What’s the beginning year of your weather data.

(7)Line 152： “the measured data in the 2000s is SOCD”？ I don’t think so, sine SOCD was calculated from SOC, bulk density, and coarse fragment, not directly measured.

(8) Line 153 : ” Second National Soil Census”, Census is usually for economics, here should be “survey”? many English words for such kind of description (for source of data) were inaccurate or confusing.

(9) Line 158. since you calculate the SOCDs of Chinese sampling points using bulk density and volume percentage of coarse fragments from the SoilGrids 2.0 data product, it is the very reason that your products are highly correlated to the SOCD of SoilGrids 2.0!

(10) Line 160: “coarse fractions proportion”, I think here is not “proportion”, since your CF was divided 100 in equation 7.

General Comment

The authors developed a long time series SOCD dataset for China by using Random Forest Approach. The dataset may be useful. However, they failed to prove the superiority of this dataset over existing similar datasets. The independent validation was only performed in a small region, while the correlation with existing SOC or SOCD maps cannot indicate a higher accuracy. In addition, the novelty is unclear. It’s unreasonable to incorporate climate zoning as predictors of the machine learning models when climatic variables have already been applied as predictors. Thirdly, although the authors produced a long time series dataset, there exists inherent problems that will contribute to large uncertainties in the interannual variation of SOCD. This issue is further exarcerbated by the authors’ failure to justify the reliability of the interannual variation. Finally, significance test is missing for most of the comparisons in this paper. For above reasons, I recommend rejection. The detailed comments are attached below. 

Major comments

1)I don’t understand why the authors calculated and incorporated vegetation and water indexes  as predictors instead of using the surface reflectances at some key bands as predictors of the Random Forest Model. As we know, machine learning models can learn the complicated and nolinear relationships.

2)The exact years when the soil samples were collected seems to be unknown. So, how did you select the corresponding annual predictors, such as the indexes derived from Landsat images? I doubt the lack of exact sample collection time could lead to uncertainty in the final annual SOC products, especially in terms of interannual variation.

3)Line 145: You should highlight what are the potential shortcomings of SOC Dynamics ML dataset. Without detailed description on this, it’s unclear why you would like to produce a similar dataset.

4)Section 3.2: Have you taken into account the auto-correlation or interdependence among the potential predictors?

5)Section 3.3: You have already included temperature and precipitation as predictors of Random Forest Model. Why use climate zoning as well? It’s unreasonable! Probably climate zoning promotes the training efficiency, but that’s because the Random Forest Model you developed in this study is not efficient enough to fully learn the relationships. If you try deep learning models or AI foundation models, you may find out that the training effiency with and without climate zoning will be approximate.

6)Many results, such as comparison results, are unconvincing, since significance test is lacking.

7)Section 4.3: You developed the SOC map for China, but only did the independent validation in Heihe River Basin, which is quite small, and the climate condition there is quite different from many other parts of China (e.g., southern China).

8)Section 4.4: The relatively high correlation with existing datasets cannot justify that your dataset is much better than existing ones. Moreover, some of these existing datasets you compared with are global-scale maps.

9)Section 4.5: Have you compared the temporal variation of SOC to other existing datasets or in-situ measurements? Can you validate your SOC temporal variation? Can you justify that your SOC dataset is much better than existing SOC maps in terms of temporal variation?

Minor comments

1)RMSEs should have units;

2)Lines 239-240: what does ‘before zone’, ‘after zone’ mean? Please polish your writing.

3)Line 231: R2 for 0-20 cm is 0.43-0.59; R2 for 0-100 cm is 0.50-0.54. How can you conclude that the fitting or correlation is slightly worse for 0-100 cm compared to 0-20 cm?

4)Lines 236-238: Suggest using relative RMSE (rRMSE, RMSE/mean value) instead of RMSE, since the SOC in arid regions can be quite low.

5)Lines 242-244: I thought the performances are similar. Have you performed significance test? In addition, the explanation is not quite convincing. Can you provide more robust proof?