I enjoyed this paper. The knowledge gap is well established, the methodology is solid, the results and discussion address the study objectives. I have a few comments for the authors' consideration.

Specific Comments:

1. Since this study focuses specifically on Qinghai and Tibet (not the entire Tibetan Plateau), I recommend adjusting the title to more accurately reflect this scope. Additionally, I suggest adding a boundary outlining these two regions in Figure 1 to clearly define the study area. For Figures 9 and 10, a gray shadow as the background could enhance the visibility of the study area and improve contrast, as the current color is hard to discern. Adding the boundary and shaded areas will help readers understand the study area more effectively and avoid potential misinterpretations. For example, based on Figures 9 and 10, I initially concluded that there is no cultivated pasture in Xinjiang, Gansu, and Sichuan, which could be misleading.

2. The Methods section could be made clearer with the following revisions:

   • The growing season is first introduced in Section 3.1, while the description of quantile extraction appears in Section 3.2. I suggest merging these two sections to ensure a smoother logical flow. For instance, begin with an introduction to SR and the seven spectral indices, followed by an explanation of the satellite products and the growing season sampling process, and conclude with the topography data. Feel free to disregard this suggestion if it doesn’t fit the structure of the paper.
   • The use of quantiles is not entirely clear. Were they used as separate model inputs, or did they interact in some way? If they were used individually, the importance of each quantile likely varies for different pixels. Clarifying this would improve understanding.
   • Several types of cultivated pasture were used as training data. Since the model is binary (cultivated vs. other), how were these different types handled in the model? Were they treated as equivalent to cultivated pasture, or did the model account for their distinctions? Additionally, how did the model perform across these various types?
   • The performance of the model should be presented more explicitly, particularly regarding the importance of different input drivers.
   • The field records used to train the model cover only a portion of the study area. Would it be feasible to extend the training data by using high-resolution satellite images for non-pilot regions?
   • Spectral and topographic data alone may not be sufficient to accurately predict cultivated pasture, especially over time. I suggest considering additional drivers such as climate variables, soil properties, and human or livestock populations in the modeling process.

3. The predicted area in Figure 8(a) appears to be consistently smaller than government statistics. Figure 11(b) also shows that the prediction for Qinghai is under-estimated. Are these discrepancies related to the limitations mentioned in point 1? They need to be clarified. 

Technical Comments:

• Figure 3: For better visualization, use two distinguishable colors for the two categories. This will improve clarity and contrast.

General Comments：

Cultivated pastures maps are essential for effective grassland resource management. This study uses a Random Forest (RF) binary classification model to identify cultivated pastures distribution on the Tibetan Plateau and distinguish between artificial grasslands, natural grasslands, and croplands. The authors have produced annual distribution maps of cultivated pastures in Qinghai Province and the Tibet Autonomous Region from 1988 to 2021. I personally think this research is significant but challenging. After carefully reading this paper, I have serious concerns about the credibility of its data.

Major concerns:

Line24: The abstract states that there are 2000 independent validation samples, but in Figure 4 and line 207, the number is given as 1000. The authors should check this discrepancy. Moreover, among these 2000 independent validation samples, only 93 are for cultivated pastures, which is clearly too small.

Line160~230: a) The authors mention in Lines 160-163 that they surveyed 81 cultivated pastures in Qinghai Province and 114 samples in Tibet. From Table 2, it appears that all field survey samples were used for model training. So, the origin of the 93 independent validation samples for cultivated pastures in Table 3 is unclear and needs to be verified by the authors. b) The proportion of other types of data is 91% in the training sample and even higher, at 95%, in the validation data, while the proportion of cultivated pastures is very low. This distribution is unreasonable and may affect the accuracy and reliability of the model. c) In machine learning studies similar to this study, it is common to use 70% of the data for training and 30% for validation. However, in this study, the training sample size is 2089 and the independent sample size is 2000. The authors should explain why they chose this proportion to divide the training and validation datasets.

Line195-205: I agree with the author's description in the introduction section that cultivated pastures differ from croplands in their shorter duration of use. However, the study failed to incorporate this feature as a criterion for distinguishing cultivated pastures from croplands when building the model. As a result, the study's extraction of cultivated pastures cannot be clearly differentiated from croplands.

Line250: a) As the images in Figure 5/6 show, it is difficult to distinguish cultivated pastures from croplands. So, what were the criteria for distinguishing between croplands and cultivated pastures when the authors did visual interpretation? b) There are only 34 datasets from 1988 to 2021. So, why does the maximum value in Figure 5/6b exceed 34?

Line273: a) In Table4, the total proportion of cultivated pastures and other types in Qinghai Province does not add up to 1. This inconsistency need to be clarified by the authors. b) The overall accuracy of 97.05% ± 0.4% mentioned in the abstract and discussion is only presented in those sections and not in the main text. Moreover, this accuracy represents the model's performance rather than the dataset's inherent quality. Moreover, they should also address the difference between model accuracy and dataset precision to ensure the reliability of the study's results.

Line294: a) Figure 10 shows that many areas in Tibet have cultivated pastures with a record time of less than 5 years, which seems inconsistent with general expectations. The authors should verify this phenomenon and provide a reasonable explanation. b) Figures 9 and S3 show that artificial grasslands suddenly disappear and reappear in the following year, which is inconsistent with the conventional pattern of grassland resource use. The authors should explain the reasons for this phenomenon.

Line 295-298: a) The authors only mention that the study results are close to government statistics between 2010 and 2017, but do not provide a quantitative error analysis. It is recommended that the authors calculate error indicators such as MAE and MRE to more clearly evaluate data accuracy. b) In Table S1, the authors compare the results of their dataset with government statistics at the county level. The table shows significant differences between the two, such as an MAE of 3.24×10⁴ and an MRE of 223% in some counties. Such high errors cast doubt on the data quality. The authors should further analyze and explain these discrepancies to enhance the credibility of their study.

Line363-367: Comparing data from 2021 with government statistics from 2017 is clearly unreasonable.

Line381-395: The discussion of the limitations of this study by the authors is not in-depth enough. Several key aspects require more thorough analysis and discussion, such as the technical limitations related to remote sensing data and modeling process.

Minor concerns:

Line25: What does F1 refer to?

Line326: The authors' discussion of the methods used to map cultivated pastures is too superficial.

Line341-344:  What are the ratios of independent validation, training samples, and validation samples?

Figure S1:  Why do the number of good observations show such distinct striping patterns?