Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Reviewer comments for authors: -

This study conducts a detailed assessment of deep learning models for predicting agricultural drought severity in Guangdong Province, employing ERA5-Land and SMAP-L3 datasets. Although the findings contribute meaningful insights, there remain opportunities to improve the study’s methodological robustness and practical relevance.

1. The study does not specify the experimental design or the number of replications used.
2. The study recognizes the limitations of SMAP-L3 data in regions with dense vegetation or extensive urban development, as well as the spatial and temporal discontinuities resulting from its revisit cycle. Although random forest interpolation was employed to compensate for missing data, the extent to which this method affects prediction accuracy—particularly in areas with substantial data gaps—has not been thoroughly examined. A more in-depth evaluation of the uncertainty introduced by this approach would enhance the study’s robustness.
3. Despite ERA5-Land’s advantage of providing consistent global coverage and reliable temporal data, its spatial resolution of 0.1° × 0.1° (approximately 9 km × 9 km) may be insufficient for capturing localized agricultural drought conditions, especially in areas with complex terrain or heterogeneous land use. The study could benefit from a discussion on how this resolution might influence the detection of microclimatic variations and small-scale drought phenomena.
4. While the study emphasizes soil moisture (SM) data alongside atmospheric and static variables, agricultural drought is shaped by a wider array of influences, such as vegetation health indicators (e.g., NDVI, EVI), actual evapotranspiration, and socio-economic conditions. Integrating these additional factors could enhance the comprehensiveness and precision of drought prediction models, particularly for long-term forecasting and distinguishing between various drought types.
5. The study employs the Soil Water Deficit Index (SWDI) to categorize drought conditions into five distinct levels. Although this method is widely used, the fixed thresholds may not be universally suitable or sufficiently sensitive to the diverse agricultural environments across Guangdong Province. Conducting a sensitivity analysis of these thresholds or comparing SWDI with alternative drought indices could lead to more robust and context-specific classifications.
6. The research evaluates four variants of the Long Short-Term Memory (LSTM) model—standard LSTM, AttLSTM, EDLSTM, and AEDLSTM. While this comparison is informative, the study could be further strengthened by including traditional machine learning algorithms (e.g., Support Vector Machines, Random Forests) and statistical models (e.g., ARIMA) in the analysis. Such a comparison would more clearly highlight the advantages and enhanced performance of the proposed deep learning approaches in drought prediction.
7. The study is confined to Guangdong Province, which is characterized by distinct monsoonal and hydrological dynamics. However, the applicability of the proposed models and their optimal configurations—such as data sources and model architectures—to agriculturally diverse regions with varying climatic conditions remains unexplored. Including a discussion on model transferability or the necessity for region-specific calibration would enhance the study’s generalizability.
8. The paper observes that model predictions tend to either overestimate or underestimate drought severity in certain regions or under specific environmental conditions. A more comprehensive analysis of the underlying causes—such as localized environmental characteristics, data quality issues, or limitations in model architecture—would offer valuable insights for refining future predictive frameworks.
9. Although the introduction highlights the broad impacts of drought on agriculture, industry, and society, the study relies solely on technical evaluation metrics such as correlation coefficient (R), unbiased root mean square error (ubRMSE), and accuracy (ACC). Incorporating indicators that capture the economic or social consequences of prediction accuracy would enhance the practical significance and policy relevance of the research findings.
10. Although the paper acknowledges that prediction uncertainty increases with longer lead times, it lacks a clear quantification or visualization of this uncertainty. Incorporating methods such as confidence intervals or probabilistic forecasting would offer decision-makers a more transparent view of the reliability and limitations of the predictions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Water  3940040 -1 review

4 DL models, within 3 forecasting frameworks, were tested with 2 data sets to assess prediction of 3 drought indices over severe drought conditions in  2020 in 5 drought categories.  The results were analyzed for accuracy and bias over time scales.  Results suggest the best model combinations varies with vegetation, topography and wetness.  The study was very well executed, and the results, well displayed and presented.  Discussion was very relevant and easy to follow.  Well Done.

A few notes: 

Reference 14 seems to be entered twice in the reference list.

References 34 and 35.  Be consistent in type.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The presented article is aimed to develop a methodology for forecasting soil water content available to agricultural plants for an time period of about 1-14 days. For this purpose, datasets from the reanalysis of meteorological data (ERA5-Land) and data from space microwave sensing of soil surface moisture (SMAP-L3) during agricultural droughts in 4 districts of Guangdong Province, located in Southern China, were used. To assess the impact of the developed neural network structure on accuracy and stability, 4 drought forecasting models based on the LSTM (Long Short Term Memory) architecture were developed.. The SWDI Soil Water Deficit Index, which is closely related to the atmospheric water demand parameter, was used as the target parameter of the corresponding reliability estimates. For each of these models, a custom raster calculation framework was developed using meteorological monitoring datasets, as well as additional layers of terrain and vegetation data. Three metrics were used to assess the reliability of the simulation. Their use made it possible to evaluate the quality of modeling for different time forecasting horizons. Based on the calculation results, a fairly detailed analysis of the data sets obtained, both in terms of metrics and their spatial distribution, was carried out. As a result of the analysis of systematic errors and correlation coefficients, it was shown that using one of the frameworks (FWB), the best forecasting results were obtained than the others, especially over longer time horizons. These important results were obtained by studying the effect of the degree of depth of droughts on the quality of forecasting. As a result of further analysis, the authors came to the conclusion that flexible forecasting systems adapted to the growing season phases of crops should be used to monitor drought in agricultural regions. Thus, the integration of monitoring data from several sources, as well as the use of new calculation architectures, made it possible to improve the methodology for predicting agricultural droughts using the example of Guangdong Province in China.

Remarks

Figure

1 – If it shows the average values of precipitation rates for 2015-2020, then this should be emphasized in the title. Figure 12 – Poor visibility of labels, it is necessary to increase the size of the pin. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I appreciate the authors’ diligent efforts in responding to all comments. The revisions are satisfactory and have enhanced the overall quality of the manuscript.