the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 24 Nov 2025
Spatio-temporal Monitoring of Agricultural Drought in China Based on Downscaled Soil Moisture Data
Abstract. Agricultural drought threatens China's food and ecological security, and accurate spatio-temporal monitoring is key for disaster mitigation. Soil moisture is critical for drought assessment, but the accuracy of existing remote sensing-based SM products in China remains to be improved. This study develops a framework that synergistically integrates a spatiotemporally adaptive gap-filling algorithm with a machine learning-based downscaling approach, generating a seamless 0.05° monthly SM dataset for China from 2003 to 2023. The methodology harnesses the complementary strengths of random forest modeling and spatiotemporal reconstruction techniques to effectively fuse multi-source satellite observations, achieving dual improvements in SM data accuracy and spatial coverage. Using this dataset, the standardized soil moisture index was applied to characterize the spatio-temporal evolution of agricultural drought. Results demonstrate that (1) The downscaled SM dataset achieves significant improvements in both spatial resolution and accuracy, showing a 2.3–34.4 % reduction in ubRMSE and 1.2–52.7 % improvement in correlation coefficients compared to benchmark datasets. (2) Drought characterization based on the downscaled SM dataset and SSI accurately identified the extent of agricultural drought, showing a significant spatiotemporal consistency with agricultural disaster area. (3) Agricultural drought intensified significantly across China during the study period, characterized by northward migration of drought center and spatially heterogeneous aridification patterns – decreasing severity from northwest to southeast while increasing from northeast to southwest. High-frequency drought zones were predominantly clustered in ecologically vulnerable regions, particularly the agro-pastoral ecotone of northern China. (4) Distinct intra-annual drought dynamics emerged, with a southwest-to-northeast expansion dominating from January to June, followed by bidirectional propagation from the Yellow River-Huaihe River Basin (YRB-HRB) to northwestern and southeastern regions from June to December. This study provides high-accuracy data support for agricultural drought monitoring and offers scientific insights for developing regional differentiated drought mitigation strategies, which are of great significance for ensuring national food security.
- Preprint
(2156 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 21 Mar 2026)
-
CC1: 'Comment on egusphere-2025-5122', Nima Zafarmomen, 25 Dec 2025
reply
-
AC1: 'Reply on CC1', Mengmeng Cao, 27 Dec 2025
reply
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2025/egusphere-2025-5122/egusphere-2025-5122-AC1-supplement.pdf
-
AC1: 'Reply on CC1', Mengmeng Cao, 27 Dec 2025
reply
-
RC1: 'Comment on egusphere-2025-5122', Anonymous Referee #1, 03 Feb 2026
reply
This manuscript presents a comprehensive framework for developing a spatially and temporally continuous soil moisture dataset over China. The methodology is technically rigorous, well justified, and the resulting product is of clear relevance for drought monitoring and hydroclimatic research. The following comments are intended to clarify specific methodological aspects and improve the presentation and interpretability of the results.
Comments:
Outlier/QC description is too vague (Page 4, lines 157–161). The authors state that outliers were excluded via “rigorous quality control,” but they do not specify the method (e.g., SD-based screening, median absolute deviation, percentile/IQR filtering, temporal consistency checks, etc.). Please explicitly describe the QC procedure and provide a citation.
Baseline datasets should be introduced earlier and motivated. The two benchmark soil moisture products (Zhang et al., 2023; Meng et al., 2021) are used for comparison, but the Introduction would benefit from a short description of what these datasets are, and why they may be insufficient for agricultural drought monitoring in China (e.g., continuity limitations, regional bias, coarse spatial resolution, limited temporal coverage). This is especially important for Meng et al. (2021), which is China-specific, because the novelty here appears to be extended temporal coverage and finer target resolution (0.25° → 0.05°).
DEM resampling + slope derivation need methodological clarity (Page 4, lines 181–183). The paper states that SRTM 1 km DEM was resampled to 0.05°, and slope was derived using a 3×3 neighborhood gradient algorithm, but the resampling method is not provided. Bilinear interpolation is often preferable to nearest-neighbor for continuous topography to avoid blocky artifacts at 0.05°. Also, please clarify whether slope was computed after resampling to 0.05° (recommended to reduce artifacts) or computed at 1 km and then resampled.
Figure 2 readability/layout. Figure 2 is comprehensive but difficult to read at its current size and vertical arrangement. Consider using a layout that better leverages horizontal space (e.g., moving Steps 1 and 2 beside Step 3, or increasing font size and simplifying inset elements) to improve readability.
Figure 4 color convention + inset choice. Consider reversing the palette so that higher soil moisture is visually intuitive (commonly blue = wetter, red = drier). Also, the inset region shown does not strongly demonstrate “local-scale” improvements; choosing a more heterogeneous boundary/transition zone (e.g., shrubland–cropland or complex terrain/land cover) would better illustrate the benefit of the downscaled product.
I am not sure what is meant by this sentence on Page 12, lines 403–405, "As shown in the zoomed-in area of Fig. 4, critical subgrid features that were previously obscured in passive microwave observations are now resolved with a precision of 0.038 m³/m³.". Please clarify what precision means here.
Pearson's R inconsistency (Page 15, lines, 449–450). The manuscript writes that there is significant "negative correlation" but the numbers reported are positive (Pearson’s R=0.015 to 0.277, p<0.05).
Expand the discussion on generalizability. The manuscript argues that the China-specific hydroclimatic drivers mitigate biases relative to generic global products (Page 13, lines 431–433). It would strengthen the Discussion to briefly frame how this workflow could be transferred to other drought-prone regions (and what region-specific drivers would likely be needed), which could broaden impact/citability.
Their data availability statement lists input sources, but it’s not clear (from the statement excerpt) whether the new seamless 0.05° SM dataset and processing code are archived somewhere persistent.
Citation: https://doi.org/10.5194/egusphere-2025-5122-RC1 -
RC2: 'Comment on egusphere-2025-5122', Anonymous Referee #2, 19 Feb 2026
reply
In this paper, the authors developed satellite-based high spatial resolution soil moisture data and applied it to drought analysis in China. They show better accuracy of their newly developed monthly soil moisture estimation than previous global products. In addition, they reveal the long-term trend of soil moisture in China.
The topic of this paper is suitable for HESS. However, I believe that this paper does not show enough novelty to be published. I do not recommend this paper for publication.
First, it is unclear to me if there is anything methodologically new in this paper. Although their method was carefully designed and their evaluation results look sound, the authors basically rely on the widely adopted triangle relationship between vegetation dynamics and temperature. I think their method is not intrinsically new, and their advantage is that they focus only on data in China to train their model and outperform the existing global estimates in China, which may not be recognized as a novel contribution from potential HESS readers.
Second, the connection between high resolution soil moisture data and drought analysis is largely unclear. The authors did not take any advantage of their accurate and high-resolution soil moisture data. I believe that their analysis can be performed by conventional soil moisture data. The authors may elaborate more on the advantages of their new data toward the assessment of long-term drought hazards.
Third, I believe that their computation of SSI has an error. I agree that monthly soil moisture follows Gaussian distribution, so that any transformation is unnecessary. However, their mean and standard deviation should be calculated each month during study period, not using all months in study period. Figure 9 clearly indicates that their SSI has a seasonal cycle, which is not consistent with drought indices conventionally implemented. Seasonally dry periods are not normally called drought. In most cases, drought is evaluated as the deviation (or anomaly) from the average conditions in each month. I do not think their approach is aligned with the previous efforts on drought quantification.
Specific comments:
Major points:
L149-156: There are several AMSR-based SM products. Which did the authors use? Is it the JAXA standard product? Please refer to the appropriate reference.
L356-362: As I pointed out above, mean and standard deviation are computed using time series of each month in most of the previous works. I do not intend to say that the authors’ computation is wrong. But I can say that it largely deviates from previous exercises. The authors need to explain what they intended or fix this error.
L422-436: I think the authors would like to refer Figure 6 in this paragraph. Also, I do not think Figure 6 is easy to understand. Why not just show table to report evaluation metrics?
L443-453: I think the authors elaborate more on evaluating their drought index. The current result is not so convincing that their SSI may not be able to reproduce real drought events. First, as I discussed above, they may design the evaluation activity to feature high spatial resolution of their data, for example, show spatial distribution of drought maps and SSI. Please show that higher resolution data is more useful than previous works. Second, please consider the other indicator of droughts such as (detrended) crop yield to be compared with SSI. Also, clarify what “officially reported crop drought-affected areas” indicate. Are there any solid criteria to identify drought affected areas? I did not understand what the authors exactly compare. Third, I could not agree that yearly SSI was useful to quantify real drought events. The authors may focus more on monthly SSI in growing and/or harvesting seasons.
Minor points:
L44: hm2 --> km2
Citation: https://doi.org/10.5194/egusphere-2025-5122-RC2 -
RC3: 'Comment on egusphere-2025-5122', Anonymous Referee #3, 22 Feb 2026
reply
Summary of comments:
This manuscript (MS herein) developed a downscaled monthly soil moisture (SM) dataset at 0.05° resolution by combining spato-temporal gap-filling algorithms and machine learning downscaling approach. The MS used the downscaled SM product to compute the standardised soil moisture drought index (SSI) in z-scores and applied SSI to assess the spatio-temporal dynamics of agricultural drought across China. The downscaled SM product achieved improved accuracy against benchmarking SM products. The MS then analysed spatial patterns, interannual variability and seasonal variations of agricultural drought derived from the SSI metric.
The topic of this MS is suitable to the science scope of HESS. The SM downscaling approach was sophisticated and logically sound, and the derived high-resolution products showed potential for agricultural drought assessments. However, a much more in-depth engagement with the relevant literature is needed in Introduction and Discussion sections to articulate the scientific context and evaluate the novelty of this research. Moreover, the drought index definition needs to be revisited, which may lead to changes in the drought assessment findings and conclusions.
After considering my major and specific comments below, I recommend a reject with invitation to resubmit.
Major comments:
- The novelty of this research was unclear. There was limited literature review for existing downscaled SM-based drought studies in the Introduction to justify the research gap. Moreover, the authors reported the key interpretation of the results without critical discussion against existing literature, partly due to the lack of a standalone Discussion section. The authors partly showed the improvement by their downscaled SM product in the validation statistics, but did not demonstrate how such high-resolution SM data adds to agricultural drought monitoring. Think critically on this question: findings in this MS on drought patterns can also be obtained using the existing SM products – why do we need a new downscaled SM product to achieve this objective?
- The drought index definition likely has a fundamental conceptual issue. The SSI calculation used the long-term mean and standard deviation and thus didn’t account for the seasonal variations in SM. By this construct, the drought assessment results in Section 4.2-4.3 were confounded with the seasonal climate, and normal dry-season conditions seem to have been systematically misclassified as drought.
- The Methodology section was extensive and sophisticated, but it read more like a technical documentation of individual steps and included too many details. The effects and added value of each step on the SM product was not clear. There was limited justification for the methodological choices or the improvement over existing methods.
Specific comments
Introduction
L60-L85: The main idea for text here is that microwave remote sensing observations of SM have coarse spatial resolution and thus requires downscaling for finer-scale, more accurate agricultural drought monitoring. Consider shortening the context to make this point sharp and succinct. For example, the detailed introduction of each microwave SM sensor can be simplified, while the ’25 to 50 km resolution’ can be moved to earlier in the paragraph.
L99-L100: DISPATCH method ‘which synergizes physical insights with statistical formulations to enhance interpretability while maintaining operational efficiency’: This description was far too generic and repeats the summary of semi-empirical approach in the earlier sentence. What are the actual variables considered in DISPATCH (e.g., soil evaporative efficiency; SEE).
L108-126: The review of past studies was too generic and lacked specific references, leading to a quite weak logical linkage between existing work and the proposed objectives and methods. Specifically, the authors made a point on the limitations of current applications of downscaled SM in agricultural drought monitoring, but only provided generic description of issues in current SM product characteristics (algorithm, spatio-temporal continuity, accuracy of global products). To really substantiate the point, this paragraph should synthesise past studies using downscaled SM products for agricultural drought monitoring and then discuss any challenges/limitations they had in producing reliable drought assessments. This will provide a logically clearer and more powerful justification for the new SM downscaling approach proposed in this MS.
Also, more specific explanation is preferable to generic wording, and specific references need to be added to support the argument. For example, for ‘regional dependencies’ for existing algorithms, the authors can quantify the spatial extent in previous studies and see if the extent is too narrow for spatial generalisability.
Materials
L127: Section heading = should only be “Materials” as Section 3 is the Methods.
L162-166 Benchmarking datasets of downscaled SM products. More specific information about the 2 benchmarking products are needed, as current wordings like “integrating multiple datasets”, “reconstruction-based” are too generic. Consider using a Table to clearly summarise and compare the key information 2 benchmark products and the proposed new product, such as spatio-temporal resolution and coverage, input data and downscaling methods, which enables straightforward comparison. Additionally, it could be worthwhile to describe the 2 products earlier in the Introduction and critically explain why they have limitations for agricultural drought monitoring.
Methods
L192-204: The writing style of this method summary paragraph was very similar to a technical note rather than a scientific paper, such as the note-taking type words like “Stage 1: Gap-filling of passive microwave SM products – A spatiotemporally adaptive gap-filling algorithm …”. Since Figure 2 was provided, there is no need to repeat the name of each step. For the example sentence above, it could be re-written as “Firstly, we developed a spatiotemporally adaptive gap-filling algorithm to …”
Section 3.1.1 Reconstruction of missing values in passive microwave remote sensing data: While I appreciate that the methodology for spatio-temporal gap-filling was sophisticated, this section is too long. The summary L209-L215 for the 3 major steps was redundant. Moreover, It will be much more informative to provide a new figure that shows the sequential changes in the spatio-temporal completeness of SM data by the 3 steps (temporal Savitzky-Golay filtering, GWR spatial gap-filling, harmonisation). This can help demonstrate and quantify the effect of each step on the SM data completeness, and thus provide much better justification for the selected spatio-temporal gap-filling methods.
Section 3.1.2 Downscaling of passive microwave remote sensing data: Given the sub-headings in this section (e.g., “(1) Calculation of the TVDI”) have already broken down the key steps, the summary L303-317 can be removed or integrated into the method description to shorten the text.
L319-L326 calculation of temperature-vegetation dryness index (TVDI): It’s unclear how the dry/wet edge and the coefficients for TVDI are defined. Are they estimated on a per-pixel basis? What is the temporal period used to define the LST-NDVI feature space?
L356-L359 Section 3.2, SSI drought index definition: The mean and standard deviation (SD) in the equation 13 are calculated from all months in the 2003-2023 study period. This failed to account for seasonality in SM dynamics and the variations in SSI will mix both the interannual and seasonal variability. For eastern and southern parts of China influenced by monsoons, warmer wet seasons will have higher rainfall and SM on average than cooler dry seasons, so SSI will have artefacts of low values for ‘drought conditions’ in cooler dry seasons. In other words, SSI captured climate aridity rather than a true drought event that shows an abnormal water deficit relative to long-term conditions.
As a solution, the authors can compute the mean and SD for each calendar month and re-calculate the SSI with respect to the month-specific mean and SD. This will account for the seasonality for SM and highlight the interannual variability due to climatic fluctuations, which would better capture the occurrence of drought events and trigger more scientifically meaningful interpretations.
Results
L396-L400 and Fig 4: SM before and after downscaling showed similar fluctuations in the west-east latitudinal gradient, so it did not provide useful evidence to support the advantage of downscaled SM. The authors may improve Figure 4 if they zoom into multiple small regions (e.g., 1 x 1 degree) with strong spatial heterogeneity (e.g., cropping regions, boundaries between distinct land cover types) and produce inset maps that compare the SM spatial patterns before and after downscaling. This will better support the argument that the downscaling method "significantly enhancing the representation of localized hydrological heterogeneity".
L404-405: What is the definition of “precision”? Please specify.
L432-433: How can the improvement over PI benchmark from Zhang et al. (2023) lead to the conclusion “This underscores the critical role of localized algorithm calibration in mitigating biases from generic global models”? More specific description of SM downscaling benchmark datasets will be useful to make this point clearer.
Fig 6: The figure was hard to understand. In the inner circle, the statistics for the downscaling product (D) repeat for 3 times, due to the comparison against the 3 benchmarks. The figure can be easily simplified into a table to report evaluation metric values, or a multi-part, side-by-side barplot of R and RMSE for 4 products separated by the 7 land cover types.
Fig 7: SSI vs crop drought-affected areas. There are multiple issues in this figure. The letters for each sub-part were problematic, e.g., Yunan has a sub-part letter (abc), where it should be (c); Shandong should be part (h) not (egh).The y-axis scales for SSI were concerning. SSI was fitted by normal distribution and converted into z-scores. For standard normal random variables, 99.9% of data will be between -3 and 3, so why the area-averaged SSI values reached outside of -5 and 5 in many cases? The authors should revisit their calculations. Moreover, there were many instances where the SSI’s performance was not convincing. For example, in Henan, 2012-2014 had high drought-affected areas, but SSI was the highest among the record. Can the authors critically discuss these instances in better details? Finally, consider adding an annotation of the temporal correlation or R-squared for each sub-region to aid with the readers’ interpretation.
L472-L473: Please define the location for the “tri-junction hotspots” for readers who are not familiar with China’s landscapes.
L501-502 Intra-annual variations of SSI: All the results and interpretation in this part were very likely to suffer from the artefacts of SSI construction and thus require a new iteration. For example, the authors found two drought peaks in spring (March–May) and late autumn to early winter (October–December). These “drought peaks” were not really drought events, but the inherent characteristics of monsoonal climates where the cooler seasons (October to May in China) have lower rainfall and thus lower SM. Similarly, the “winter–spring drought and summer–autumn wetness” (L512-513) is already known due to the seasonal monsoon cycle, so there is limited extra information provided by the SSI.
It is expected that the identified spatio-temporal drought patterns will be quite different if the SSI is re-defined by standardising against the month-specific mean and SD instead of the whole-period statistics. When SSI is re-defined, the analyses in Fig 7 to 10 will need to be repeated to assess the consistencies and discrepancies in the findings.
L539: Please navigate the readers regarding the locations for "China’s key agro-pastoral transition zones and ecologically vulnerable regions". Also, this could be a practically important result for the downscaled SM product, so more in-depth discussions referring to previous drought studies for such ecological vulnerable regions will be useful.
Discussion and conclusion
The combined “5. Discussion and conclusion” section basically functioned as a conventional Conclusion section that reiterated the main findings. The last paragraph briefly discussed methodological limitations, but was too generic. Moreover, it lacked critical comparison with previous studies and deeper interpretation of the results in the broader scientific context. It is unclear how the downscaled SM product and subsequent analyses in this research provided new scientific insights to the agricultural drought monitoring in China, compared to previous studies in this field. I will really encourage a separate Discussion and Conclusion section respectively, where the Discussion is expanded to critically compare their findings with relevant literature on SM downscaling applications in drought monitoring, examine possible biophysical explanations for observed drought patterns, and clarify the broader implications and limitations of the results. This will substantially improve the scientific quality and clarity of the MS so that it better matches the standards and requirements by HESS.
L572: The first sentence starting with ‘Although’ is incomplete, and should be integrated with the next sentence.
.
Citation: https://doi.org/10.5194/egusphere-2025-5122-RC3 -
RC4: 'Comment on egusphere-2025-5122', Anonymous Referee #4, 22 Feb 2026
reply
The manuscript entitled “Spatio-temporal Monitoring of Agricultural Drought in China Based on Downscaled Soil Moisture Data” by Luo et al. investigates the downscaling of remotely sensed soil moisture products from a spatial resolution of approximately 0.25° to 0.05° for the purpose of agricultural drought monitoring. The authors present a comprehensive effort to design and assess the proposed downscaling framework. The subject is both timely and highly relevant, given the increasing importance of accurate drought monitoring for water resources management and policy decisions. Overall, the study fits well within the scope of Hydrology and Earth System Sciences and demonstrates clear potential to make a meaningful contribution to the field. Nevertheless, several methodological aspects and minor technical details would benefit from additional clarification and refinement to improve the transparency and robustness of the findings. The comments below are intended to support the authors in further strengthening the manuscript’s clarity.
Specific comments:
- Gap-filling strategy across AMSR-E, SMOS, and AMSR-2 datasets
The general approach used to fill the gaps between AMSR-E, SMOS, and AMSR-2 soil moisture products requires further clarification.In addition to the variations in ascending and descending times, diurnal variability may also impact the monthly mean soil moisture. These sensors use different methods to calculate soil moisture from microwave signals, including various algorithms, models, frequency bands, calibration steps, and processing systems. Such methodological differences may introduce structural inconsistencies and biases in the merged time series. The manuscript would benefit from a more detailed discussion of how these inter-sensor differences were addressed, harmonized, or evaluated prior to merging the datasets. In addition, clearer documentation and references describing the specific soil moisture products used would improve transparency and reproducibility.
- Comparison with previous machine learning studies
The manuscript compares the performance of the proposed method with results reported in previous studies that applied various machine learning approaches. Although these comparisons offer valuable context, they do not ensure the accuracy or reliability of the proposed method. Differences in study area, input datasets, preprocessing procedures, spatial and temporal resolution, training strategies, and validation frameworks can substantially influence reported performance metrics. Therefore, a more cautious interpretation of these comparisons is recommended. The manuscript would benefit from a more detailed explanation of how similar the experimental settings are in different studies and a more direct discussion of the uncertainties that come with the modeling framework and input data. This would strengthen the validity of the conclusions drawn from the comparative analysis. - Interpretation of correlation strength between SSI and drought-affected areas
The manuscript reports statistically significant negative correlations between area-averaged SSI and governmental crop drought-affected areas (Pearson’s R = 0.015–0.277, p < 0.05). While statistical significance is achieved, the reported correlation coefficients indicate very weak to weak relationships. In particular, an R value as low as 0.015 suggests negligible practical association, despite being statistically significant. Therefore, the conclusion that the downscaled SSI “can effectively capture agricultural drought patterns across China” appears somewhat overstated.
Technical corrections:
In general, the figures in the manuscript would benefit from consistent fonts, sizes, and styles across all panels and labels in accordance with the journal’s figure requirements.
- Figure 1: The caption could be rephrased to clarify that the figure shows the locations of the in-situ soil moisture (SM) stations rather than the in-situ SM distribution. Additionally, it may be worth verifying the elevation color scale, as it appears to be reversed (e.g., coastal areas appear higher than the Tibetan Plateau and the Himalayas in the west). "Province boundary" is written twice in the legend.
- Figure 2: It provides a comprehensive overview of the methodology but is somewhat dense due to the large number of formulas included. The text already explains many equations, so the figures could benefit from a simplified layout that omits some formulas to improve readability.
- Figure 4: It would be helpful to use the same maximum and minimum limits for the color bar in both the original and downscaled images for easier comparison. The axis of horizontal profile should include units (e.g., consider changing Data Value to Soil Moisture (m-3 m-3), Transect(unit)). While I assume the inset corresponds to the horizontal profile along the transect, it is currently difficult to relate the x-axis values to the transect location on the map. Is this profile also showing the average value along the latitudinal axis?
- Figure 5: The unit on the figure is currently written as m^3/m^3. For clarity and proper SI notation, it should be formatted as m-3 m-3.
- Figure 7: It is recommended to name each panel with the river basin name and its abbreviation instead of using letters (e.g. (h) Shandong (YRB-HRB-HRYB)) as the current panel naming can be confusing.
- Figure 8: Please use the same maximum and minimum range to observe the drought evolution throughout the years.
Citation: https://doi.org/10.5194/egusphere-2025-5122-RC4 - Gap-filling strategy across AMSR-E, SMOS, and AMSR-2 datasets
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 336 | 202 | 26 | 564 | 231 | 234 |
- HTML: 336
- PDF: 202
- XML: 26
- Total: 564
- BibTeX: 231
- EndNote: 234
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This study presents a robust and scientifically significant contribution to the field of hydro-climatology and precision agriculture. The novelty of this work lies in its integrated methodological framework that successfully bridges the gap between coarse-resolution satellite observations and the high-resolution requirements of regional drought assessment.
Minor Comments for Improvement
While the manuscript is well-structured, the following minor points should be addressed to enhance clarity and depth:
1. The term "file-scale" appears to be a typo. It should likely be corrected to "field-scale" to accurately reflect the context of agricultural monitoring.
2. In the comparison between original and downscaled images, the "zoomed-in" areas are very helpful. However, it would be beneficial to explicitly state the specific geographic location (e.g., coordinates or province) of the zoomed-in sub-regions in the figure caption to provide better context.
3. You mention that TVDI and precipitation were dominant predictors. It would be valuable to include a small Feature Importance Plot (perhaps as a sub-panel in Figure 2 or 3) to quantitatively show the contribution of each environmental covariate.
4. In Section 5, you mention that the study is constrained by "static SSI thresholds." Briefly expanding on how dynamic thresholds (accounting for different crop phenological stages) might change the results in the future would strengthen the "Future Work" portion of the discussion.
5. Finally, while the study correctly identifies groundwater depletion and irrigation expansion as compounding drivers of drought in the YRB-HRB regions (Line 475), the discussion would be significantly strengthened by acknowledging the role of high-resolution vegetation data in modeling these interactions. I strongly recommend considering and discussing the implications of studies such as "Assimilation of sentinel‐based leaf area index for modeling surface‐ground water interactions in irrigation districts". Integrating such perspectives would provide a deeper theoretical link between satellite-derived vegetation indices (like the LAI used in your RF model) and the complex subterranean water dynamics that govern agricultural moisture availability in China’s intensive irrigation zones.