the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 28 Apr 2026
Evaluation of ASCAT soil moisture retrievals and their potential to detect intraday variability
Abstract. Accurate sub-daily soil moisture (SM) retrievals from satellite observations remain a major challenge due to sparse temporal sampling and retrieval uncertainties. This study introduces a localized convolutional neural network (CNN-l) framework designed to enhance SM estimates from Advanced SCATterometer (ASCAT) observations by exploiting spatial features and adapting to local conditions. The proposed approach achieves strong agreement with ERA5 reference SM, with total correlation coefficients exceeding 0.9, even at a sub-daily scale. Validation against in situ measurements from 568 monitoring sites across the contiguous United States (CONUS) shows a median temporal correlation of 0.65, compared to 0.59 for the operational ASCAT H120 product. Our CNN-based retrievals also reveal meaningful intraday variability when SM signals exceed retrieval uncertainty, particularly during heavy precipitation events (> 10 mm day−1), offering new insight into short-term hydrological responses. Future efforts should prioritize the integration of complementary satellite observations from multiple instruments to enhance retrieval accuracy, robustness, and temporal resolution. Additionally, strategies to improve retrieval of extremes (such as localization strategies or variable augmentation) should be further developed.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(3694 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(3694 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2026-2360', Anonymous Referee #1, 15 May 2026
-
AC1: 'Reply on RC1', Lan Anh Dinh, 08 Jul 2026
Dear Referee #1,
We thank you for your valuable comments and suggestions. Below, we outline our point-by-point responses. We will implement all corresponding changes and improvements in the revised manuscript during the next stage.
The study presents a novel approach for retrieving ASCAT soil moisture via a convolutional neural network (CNN) that explicitly models spatial dependencies. The validation is robust, using both the ERA5 reanalysis and in situ soil moisture from the ISMN. Furthermore, the performance of the CNN-based ASCAT SM is compared with the H SAF ASCAT SM data record (derived using the change detection approach). The results are convincing, with improvements to the correlation coefficient for the CNN ASCAT SM relative to the H SAF ASCAT SM. However, the study would benefit from a clearer description of the machine learning architecture for reproducibility. Also, stratifying the ISMN validation results according to land cover and vegetation types would be informative.
Response: Thank you for your appreciation of our work. We respond to your two concerns below.
Major comments:
The description of the NN in Sections 3.1 and 3.2 should give more detail on hyperparameter tuning and overfitting prevention. Furthermore, a diagram of the steps would be beneficial for the reader.
Response: We agree that explicitly detailing these technical aspects improves the manuscript's transparency. We will expand the model description in the revised manuscript to highlight the following points:
- Hyperparameter tuning: Due to the computational cost of our large-scale datasets, automated tuning (e.g., Bayesian optimization) was infeasible (Rabiei et al., 2025; Dinh, 2026). Instead, we used a structured manual search: key hyperparameters were initialized using literature standards (for instance, as in Pellet et al. (2025) and Rabiei et al. (2025)) and iteratively refined on the validation set until further adjustments yielded negligible improvements.
- Overfitting prevention: In addition to our strict temporal data split (2016–2018 partitioned into 80% training / 20% validation, with 2019 completely held out for independent testing), we actively mitigated overfitting during training by employing early stopping. As noted in Section 3.2, training ran for a maximum of 200 epochs, but was stopped early if the validation loss showed no improvement after 5 epochs, at which point the best weights were restored. We will ensure this mechanism is emphasized more clearly in the revised text.
- Diagram: We agree this will significantly improve clarity. We will include a detailed flowchart describing the CNN architecture and processing steps in the revised manuscript.
The validation against in situ data in Section 4.2 does not give any information about the land cover and vegetation types for the different networks, and how this affects the relative performance of the SM datasets. From Figure 2 (b) it seems that the CNN soil moisture is particularly advantageous in the mountainous regions in the western US. On the other hand, H120 seems to perform slightly better in the central Great Planes, which is more agricultural.
Response: Thank you for this insightful observation. We agree with your interpretation of the regional patterns. In the revised manuscript, we will include a new performance assessment stratified by land cover and vegetation type for each ISMN site. This will be based on land-cover information from the Annual National Land Cover Database (NLCD) Collection 1, 2019 product (U.S. Geological Survey (USGS), 2024). This additional stratification will enable a quantitative evaluation of how land cover and vegetation types influence regional differences in model performance.
Minor comments:
There are many minor grammatical mistakes, and the paper would benefit from a thorough proofread.
Response: We will carefully proofread the entire manuscript and correct the grammatical errors to improve readability.
The authors should discuss the degree of dependence on ERA5 (training target vs validation independence)
Response: We thank you for highlighting this. Regarding the training dependence, our model is indeed heavily dependent on ERA5, as we used ERA5 soil moisture (SM) as the target. This choice is deliberate and scientifically justified. ERA5 currently provides one of the most robust large-scale SM datasets, resulting from the joint assimilation of multiple satellite observations, atmospheric forcing, and in situ measurements into a physically based land surface model, alongside bias-correction procedures (Soci et al., 2024). Using such a comprehensive reanalysis as a reference for supervised learning is a well-established practice in remote sensing (Aires et al., 2005; Rodríguez-Fernández et al., 2015; Pellet et al., 2025).
To ensure an independent evaluation, we validated the model against in situ observations from the ISMN, which were not used during training. Although the model was trained to reproduce the ERA5 SM target, validation against independent in situ measurements provides a robust assessment of its ability to retrieve physically meaningful SM information beyond the ERA5 training space. The fact that the retrieval shows comparable, and in some locations improved, agreement with in situ observations relative to ERA5 indicates that the model is able to extract additional information from the ASCAT observations beyond simply reproducing the ERA5 background.
In the revised manuscript, we will add a dedicated paragraph to explicitly clarify this dependence and the role of our independent validation.
In Section 1 the authors should define what the surface soil moisture represents (i.e. top few centimetres of soil)
Response: We agree with the reviewer. We will add this definition into the manuscript.
Line 139: It seems this should sentence should start with a bullet point like the previous two bullet points; Line 168: The second ypred should be yref; Line 233: The red point is shown in Figure 4 (a), not Figure 4 (b); Line 252: section 55.1 typo
Response: Thank you, we will correct all these errors in the revised manuscript.
Line 269: ERA5-24h should be defined before the abbreviation is first used
Response: Yes, actually this has been explained earlier in the original manuscript in line 229: “For reference, the full 24-hour SM amplitude of ERA5 (ERA5-24h).”
References:
- Rabiei S, Babaeian E, Grunwald S (2025). Deep learning-based short- and mid-term surface and subsurface soil moisture projections from remote sensing and digital soil maps. Remote Sens. 17, 3219. https://doi.org/10.3390/rs17183219
- Pellet V, Aires F, Boucher E, Volden E (2025). Enhancing soil moisture statistical retrieval from SMOS using partial convolutions and localization strategies. J. Appl. Meteorol. Climatol. 64, 1951–1965. https://doi.org/10.1175/JAMC-D-25-0041.1
- Dinh LA (2026). ASCAT soil moisture retrieval using deep learning: A focus on localization strategy. Front. Remote Sens. 6, 1718353. https://doi.org/10.3389/frsen.2025.1718353
- U.S. Geological Survey (2024). Annual NLCD Collection 1 Science Products. https://doi.org/10.5066/P94UXNTS
- Soci C, Hersbach H, Simmons A, Poli P, Bell B, Berrisford P, et al. (2024). The ERA5 global reanalysis from 1940 to 2022. Q. J. R. Meteorol. Soc. 150, 4014–4048. https://doi.org/10.1002/qj.4803
- Aires F, Prigent C, Rossow WB (2005). Sensitivity of satellite microwave and infrared observations to soil moisture at a global scale: 2. Global statistical relationships. J. Geophys. Res. Atmos. 110. https://doi.org/10.1029/2004JD005094
- Rodriguez-Fernandez N, Aires F, Richaume P, Kerr Y H, Prigent C, Kolassa J, et al. (2015). Soil moisture retrieval using neural networks: Application to SMOS. IEEE Transactions on Geoscience and Remote Sensing 53, 5991–6007. https://doi.org/10.1109/TGRS.2015.2430845
- Rodriguez-Fernandez N, de Rosnay P, Albergel C, Richaume P, Aires F, Prigent C, et al. (2019). SMOS neural network soil moisture data assimilation in a land surface model and atmospheric impact. Remote Sens. 11. https://doi.org/10.3390/rs11111334
Citation: https://doi.org/10.5194/egusphere-2026-2360-AC1
-
AC1: 'Reply on RC1', Lan Anh Dinh, 08 Jul 2026
-
RC2: 'Comment on egusphere-2026-2360', Anonymous Referee #2, 07 Jun 2026
Review of
Evaluation of ASCAT soil moisture retrievals and their potential to detect intraday variability
by Dinh et al.
General comments:
This is an interesting paper demonstrating the potential of machine learning to enhance existing soil moisture products derived from EUMETSAT's ASCAT instruments on MetOp satellites. A neural network (NN) is trained using ERA5 simulations as a reference. The resulting soil moisture retrievals are less noisy than the original ASCAT soil moisture product. In addition, the authors provide some evidence suggesting that sub-daily soil moisture variability could be observed through NN retrievals. While the paper is well written, the addition of a diagram presenting the NN training and retrievals with inputs and outputs would be useful. Also missing is a table summarising mean score values per experiment with respect to in situ observations. In addition to scores covering the entire four-year period, the table should also indicate seasonal scores (e.g. DJF, MAM, JJA and SON across years). Finally, the approach is somewhat circular since ERA5-derived variables (ST and LAI) are used to predict soil moisture, as well as ASCAT backscatters. This shortcoming should be clearly acknowledged, and the reasons for it should be made clear.
Recommendation: major revisions.
Particular comment:
- L. 41: "SM estimates at sub-daily resolution" is not clear. Do you mean that several ASCAT observations are available for a given location on the same day? Or does the time of observation vary from one day to the next?
- L. 44: ASCAT observations are global. Why focusing on CONUS? Any reason for that?
- L. 102: Does “the ERA5 LAI product” exist? As far as I am aware, ERA5 neither simulates dynamic LAI nor integrates LAI observations. This sentence needs to be clarified.
- L. 225: Replace "extracting a reliable diurnal signal is a true challenge!" by "extracting a reliable signal is challenging".
- L. 232 (Fig. 4): ERA5 precipitation data can be subject to significant biases. Could you include in situ precipitation observations in Fig. 4b?
- L. 240: This study does not actually demonstrate the intraday capability of the method. In situ observations of precipitation would provide a more convincing demonstration.
- L. 252: "section 55.1"?
Citation: https://doi.org/10.5194/egusphere-2026-2360-RC2 -
AC2: 'Reply on RC2', Lan Anh Dinh, 08 Jul 2026
Dear Referee #2,
Thank you for your thorough review and insightful suggestions. Below, we provide our point-by-point responses to your comments for this interactive discussion. We look forward to implementing these outlined changes in the revised version of the manuscript in the next phase.
General comments:
This is an interesting paper demonstrating the potential of machine learning to enhance existing soil moisture products derived from EUMETSAT's ASCAT instruments on MetOp satellites. A neural network (NN) is trained using ERA5 simulations as a reference. The resulting soil moisture retrievals are less noisy than the original ASCAT soil moisture product. In addition, the authors provide some evidence suggesting that sub-daily soil moisture variability could be observed through NN retrievals. While the paper is well written, the addition of a diagram presenting the NN training and retrievals with inputs and outputs would be useful.
Response: Thank you for your appreciation of our work. As also mentioned in the response to Referee #1, we will expand the model description and add a detailed diagram describing the CNN architecture and processing steps in the revised manuscript.
Also missing is a table summarising mean score values per experiment with respect to in situ observations. In addition to scores covering the entire four-year period, the table should also indicate seasonal scores (e.g. DJF, MAM, JJA and SON across years).
Response: Thank you for this helpful suggestion. We agree that this addition will improve the presentation of the results, and this information will be included in the revised manuscript.
Finally, the approach is somewhat circular since ERA5-derived variables (ST and LAI) are used to predict soil moisture, as well as ASCAT backscatters. This shortcoming should be clearly acknowledged, and the reasons for it should be made clear.
Response: We agree that this aspect of the experimental design should be more clearly acknowledged and justified. Our framework uses ERA5 SM as the training target, together with ERA5 ST and LAI as ancillary inputs, thereby introducing dependence on the ERA5 framework. The choice of ERA5 SM as the training target is motivated by several considerations. ERA5 provides one of the most comprehensive large-scale SM estimates currently available, combining a physically based land surface model with the assimilation of multiple satellite observations, atmospheric forcing, and available in situ information (Soci et al., 2024). Consequently, it offers a spatially and temporally consistent reference dataset for supervised learning, an approach that has been widely adopted in remote sensing applications (Aires et al., 2005; Rodríguez-Fernández et al., 2015; Pellet et al., 2025).
The inclusion of ST and LAI is also physically motivated. Microwave backscatter is highly sensitive to vegetation attenuation and the thermodynamic state of the land surface. These variables thus provide boundary conditions that help the model separate SM effects from other sources of variability in the ASCAT signal. Methodologically, using ERA5 ancillary variables ensures the retrieved ASCAT product remains fully consistent with the ERA5 background space, which is a standard requirement for data assimilation preparatory frameworks (e.g., Rodríguez-Fernández et al., 2019). We acknowledge that the retrieved product inherits the large-scale climatology and background state of ERA5. However, the sub-daily and event-scale temporal variations are primarily constrained by the independent ASCAT observations, which provide information that is not contained in the ERA5 ancillary variables alone.
Finally, model performance was evaluated using independent in situ SM observations from the ISMN, which were not used during training. Although this evaluation cannot eliminate the dependence on ERA5 introduced during model development, it provides an independent assessment of whether the retrieval produces physically meaningful SM estimates. The fact that the retrieval shows comparable, and in some locations improved, agreement with in situ observations relative to ERA5 indicates that the model is able to extract additional information from the ASCAT observations beyond simply reproducing the ERA5 background.
We will revise the manuscript to explicitly acknowledge this dependence on ERA5 and to clarify the physical and methodological rationale for the chosen framework.
Particular comments:
- L. 41: "SM estimates at sub-daily resolution" is not clear. Do you mean that several ASCAT observations are available for a given location on the same day? Or does the time of observation vary from one day to the next?
Response: We are referring to the availability of multiple ASCAT observations for a given location on the same day (corresponding to native overpass times), rather than daily aggregated soil moisture. We will make this distinction clearer in the revised manuscript.
- L. 44: ASCAT observations are global. Why focusing on CONUS? Any reason for that?
Response: Here we selected the CONUS as it offers a highly dense network of in situ SM measurements (Marinescu et al., 2024), which are essential for robust validation. Also, CONUS encompasses pronounced spatial and climatic variability, making it an ideal testbed (Bernhardt et al., 2018). We will explicitly state this rationale in the introduction section (after L. 44) of the revised manuscript to clarify our selection for the readers.
- L. 102: Does “the ERA5 LAI product” exist? As far as I am aware, ERA5 neither simulates dynamic LAI nor integrates LAI observations. This sentence needs to be clarified.
Response: Thank you for pointing this out. We fully agree; ERA5 does not simulate a dynamic LAI nor assimilate LAI observations, but rather uses a prescribed monthly climatology (Duveiller et al., 2023). Our original phrasing ("the ERA5 LAI product") was inadvertently misleading, as it implied a dynamically varying dataset.
However, it is important to note here that our rationale for using this specific climatological background, rather than dynamic external satellite datasets, was to ensure strict physical consistency with our training target. Feeding the model dynamic vegetation anomalies that the ERA5 soil moisture target inherently does not possess would introduce an input-target mismatch. We will clarify this point in the revised manuscript.
- L. 225: Replace "extracting a reliable diurnal signal is a true challenge!" by "extracting a reliable signal is challenging".
Response: We agree with this change.
- L. 232 (Fig. 4): ERA5 precipitation data can be subject to significant biases. Could you include in situ precipitation observations in Fig. 4b? ; - L. 240: This study does not actually demonstrate the intraday capability of the method. In situ observations of precipitation would provide a more convincing demonstration.
Response: We thank you for this valuable suggestion. The primary purpose of using ERA5 precipitation (as in Fig. 4a) was to illustrate the spatial pattern of the precipitation even over the CONUS. However, we agree that including an independent observational precipitation record would strengthen the interpretation of the case study presented in Fig. 4b.
In the revised manuscript, we will attempt to obtain precipitation observations for the selected location from available in situ gauge networks, such as the NOAA Global Historical Climatology Network-hourly (GHCNh) or the ISMN, where available. If no suitable gauge observations are available for the selected event, we will instead use a high-resolution gauge-adjusted precipitation product (e.g., NOAA Stage IV) to provide an observational reference. This will allow us to independently verify the timing and magnitude of the rainfall event, thereby strengthening the interpretation of the SM response without relying solely on ERA5 data.
We also agree that the original wording in Line 240 overstated the conclusions that can be drawn from a single case study. Accordingly, we will revise the text to clarify that this example is intended to illustrate the model’s behavior in capturing intraday SM variability during a heavy precipitation event, rather than to provide a formal demonstration of model capability. By incorporating an independent observational precipitation dataset, we will also provide an objective reference for assessing whether the timing of the retrieved SM response is consistent with the observed rainfall, rather than reflecting only the ERA5 forcing used during model development.
- L. 252: "section 55.1"?
Response: We referred to section 5.1 here. We will correct this in the revised manuscript.
References:
- Marinescu P J, Abdi D, Hilburn K, Jankov I, Lin L (2024). An Evaluation of NOAA Modeled and In Situ Soil Moisture Values and Variability across the Continental United States. Weather and Forecasting, 39(3), 523-540. https://doi.org/10.1175/WAF-D-23-0136.1
- Bernhardt J, Carleton A M, LaMagna C (2018). A comparison of daily temperature-averaging methods: spatial variability and recent change for the conus. J. Clim.31, 979–996. https://doi.org/10.1175/JCLI-D-17-0089.1
- Duveiller G, Pickering M, Muñoz-Sabater J, Caporaso L, Boussetta S, Balsamo G, Cescatti A (2023). Getting the leaves right matters for estimating temperature extremes, Geosci. Model Dev., 16, 7357–7373, https://doi.org/10.5194/gmd-16-7357-2023
- Soci C, Hersbach H, Simmons A, Poli P, Bell B, Berrisford P, et al. (2024). The ERA5 global reanalysis from 1940 to 2022. Q. J. R. Meteorol. Soc. 150, 4014–4048. https://doi.org/10.1002/qj.4803
- Aires F, Prigent C, Rossow WB (2005). Sensitivity of satellite microwave and infrared observations to soil moisture at a global scale: 2. Global statistical relationships. J. Geophys. Res. Atmos. 110. https://doi.org/10.1029/2004JD005094
- Rodriguez-Fernandez N, Aires F, Richaume P, Kerr Y H, Prigent C, Kolassa J, et al. (2015). Soil moisture retrieval using neural networks: Application to SMOS. IEEE Transactions on Geoscience and Remote Sensing 53, 5991–6007. https://doi.org/10.1109/TGRS.2015.2430845
- Rodriguez-Fernandez N, de Rosnay P, Albergel C, Richaume P, Aires F, Prigent C, et al. (2019). SMOS neural network soil moisture data assimilation in a land surface model and atmospheric impact. Remote Sens. 11. https://doi.org/10.3390/rs11111334
- Pellet V, Aires F, Boucher E, Volden E (2025). Enhancing soil moisture statistical retrieval from SMOS using partial convolutions and localization strategies. J. Appl. Meteorol. Climatol. 64, 1951–1965. https://doi.org/10.1175/JAMC-D-25-0041.1
Citation: https://doi.org/10.5194/egusphere-2026-2360-AC2
-
AC2: 'Reply on RC2', Lan Anh Dinh, 08 Jul 2026
Peer review completion
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2026-2360', Anonymous Referee #1, 15 May 2026
The study presents a novel approach for retrieving ASCAT soil moisture via a convolutional neural network (CNN) that explicitly models spatial dependencies. The validation is robust, using both the ERA5 reanalysis and in situ soil moisture from the ISMN. Furthermore, the performance of the CNN-based ASCAT SM is compared with the H SAF ASCAT SM data record (derived using the change detection approach). The results are convincing, with improvements to the correlation coefficient for the CNN ASCAT SM relative to the H SAF ASCAT SM. However, the study would benefit from a clearer description of the machine learning architecture for reproducibility. Also, stratifying the ISMN validation results according to land cover and vegetation types would be informative.
Major comments:
- The description of the NN in Sections 3.1 and 3.2 should give more detail on hyperparameter tuning and overfitting prevention. Furthermore, a diagram of the steps would be beneficial for the reader.
- The validation against in situ data in Section 4.2 does not give any information about the land cover and vegetation types for the different networks, and how this affects the relative performance of the SM datasets. From Figure 2 (b) it seems that the CNN soil moisture is particularly advantageous in the mountainous regions in the western US. On the other hand, H120 seems to perform slightly better in the central Great Planes, which is more agricultural.
Minor comments:
- There are many minor grammatical mistakes and the paper would benefit from a thorough proof read
- The authors should discuss the degree of dependence on ERA5 (training target vs validation independence)
- In Section 1 the authors should define what the surface soil moisture represents (i.e. top few centimetres of soil)
- Line 139: It seems this should sentence should start with a bullet point like the previous two bullet points
- Line 168: The second ypred should be yref
- Line 233: The red point is shown in Figure 4 (a), not Figure 4 (b)
- Line 252: section 55.1 typo
- Line 269: ERA5-24h should be defined before the abbreviation is first used
Citation: https://doi.org/10.5194/egusphere-2026-2360-RC1 -
AC1: 'Reply on RC1', Lan Anh Dinh, 08 Jul 2026
Dear Referee #1,
We thank you for your valuable comments and suggestions. Below, we outline our point-by-point responses. We will implement all corresponding changes and improvements in the revised manuscript during the next stage.
The study presents a novel approach for retrieving ASCAT soil moisture via a convolutional neural network (CNN) that explicitly models spatial dependencies. The validation is robust, using both the ERA5 reanalysis and in situ soil moisture from the ISMN. Furthermore, the performance of the CNN-based ASCAT SM is compared with the H SAF ASCAT SM data record (derived using the change detection approach). The results are convincing, with improvements to the correlation coefficient for the CNN ASCAT SM relative to the H SAF ASCAT SM. However, the study would benefit from a clearer description of the machine learning architecture for reproducibility. Also, stratifying the ISMN validation results according to land cover and vegetation types would be informative.
Response: Thank you for your appreciation of our work. We respond to your two concerns below.
Major comments:
The description of the NN in Sections 3.1 and 3.2 should give more detail on hyperparameter tuning and overfitting prevention. Furthermore, a diagram of the steps would be beneficial for the reader.
Response: We agree that explicitly detailing these technical aspects improves the manuscript's transparency. We will expand the model description in the revised manuscript to highlight the following points:
- Hyperparameter tuning: Due to the computational cost of our large-scale datasets, automated tuning (e.g., Bayesian optimization) was infeasible (Rabiei et al., 2025; Dinh, 2026). Instead, we used a structured manual search: key hyperparameters were initialized using literature standards (for instance, as in Pellet et al. (2025) and Rabiei et al. (2025)) and iteratively refined on the validation set until further adjustments yielded negligible improvements.
- Overfitting prevention: In addition to our strict temporal data split (2016–2018 partitioned into 80% training / 20% validation, with 2019 completely held out for independent testing), we actively mitigated overfitting during training by employing early stopping. As noted in Section 3.2, training ran for a maximum of 200 epochs, but was stopped early if the validation loss showed no improvement after 5 epochs, at which point the best weights were restored. We will ensure this mechanism is emphasized more clearly in the revised text.
- Diagram: We agree this will significantly improve clarity. We will include a detailed flowchart describing the CNN architecture and processing steps in the revised manuscript.
The validation against in situ data in Section 4.2 does not give any information about the land cover and vegetation types for the different networks, and how this affects the relative performance of the SM datasets. From Figure 2 (b) it seems that the CNN soil moisture is particularly advantageous in the mountainous regions in the western US. On the other hand, H120 seems to perform slightly better in the central Great Planes, which is more agricultural.
Response: Thank you for this insightful observation. We agree with your interpretation of the regional patterns. In the revised manuscript, we will include a new performance assessment stratified by land cover and vegetation type for each ISMN site. This will be based on land-cover information from the Annual National Land Cover Database (NLCD) Collection 1, 2019 product (U.S. Geological Survey (USGS), 2024). This additional stratification will enable a quantitative evaluation of how land cover and vegetation types influence regional differences in model performance.
Minor comments:
There are many minor grammatical mistakes, and the paper would benefit from a thorough proofread.
Response: We will carefully proofread the entire manuscript and correct the grammatical errors to improve readability.
The authors should discuss the degree of dependence on ERA5 (training target vs validation independence)
Response: We thank you for highlighting this. Regarding the training dependence, our model is indeed heavily dependent on ERA5, as we used ERA5 soil moisture (SM) as the target. This choice is deliberate and scientifically justified. ERA5 currently provides one of the most robust large-scale SM datasets, resulting from the joint assimilation of multiple satellite observations, atmospheric forcing, and in situ measurements into a physically based land surface model, alongside bias-correction procedures (Soci et al., 2024). Using such a comprehensive reanalysis as a reference for supervised learning is a well-established practice in remote sensing (Aires et al., 2005; Rodríguez-Fernández et al., 2015; Pellet et al., 2025).
To ensure an independent evaluation, we validated the model against in situ observations from the ISMN, which were not used during training. Although the model was trained to reproduce the ERA5 SM target, validation against independent in situ measurements provides a robust assessment of its ability to retrieve physically meaningful SM information beyond the ERA5 training space. The fact that the retrieval shows comparable, and in some locations improved, agreement with in situ observations relative to ERA5 indicates that the model is able to extract additional information from the ASCAT observations beyond simply reproducing the ERA5 background.
In the revised manuscript, we will add a dedicated paragraph to explicitly clarify this dependence and the role of our independent validation.
In Section 1 the authors should define what the surface soil moisture represents (i.e. top few centimetres of soil)
Response: We agree with the reviewer. We will add this definition into the manuscript.
Line 139: It seems this should sentence should start with a bullet point like the previous two bullet points; Line 168: The second ypred should be yref; Line 233: The red point is shown in Figure 4 (a), not Figure 4 (b); Line 252: section 55.1 typo
Response: Thank you, we will correct all these errors in the revised manuscript.
Line 269: ERA5-24h should be defined before the abbreviation is first used
Response: Yes, actually this has been explained earlier in the original manuscript in line 229: “For reference, the full 24-hour SM amplitude of ERA5 (ERA5-24h).”
References:
- Rabiei S, Babaeian E, Grunwald S (2025). Deep learning-based short- and mid-term surface and subsurface soil moisture projections from remote sensing and digital soil maps. Remote Sens. 17, 3219. https://doi.org/10.3390/rs17183219
- Pellet V, Aires F, Boucher E, Volden E (2025). Enhancing soil moisture statistical retrieval from SMOS using partial convolutions and localization strategies. J. Appl. Meteorol. Climatol. 64, 1951–1965. https://doi.org/10.1175/JAMC-D-25-0041.1
- Dinh LA (2026). ASCAT soil moisture retrieval using deep learning: A focus on localization strategy. Front. Remote Sens. 6, 1718353. https://doi.org/10.3389/frsen.2025.1718353
- U.S. Geological Survey (2024). Annual NLCD Collection 1 Science Products. https://doi.org/10.5066/P94UXNTS
- Soci C, Hersbach H, Simmons A, Poli P, Bell B, Berrisford P, et al. (2024). The ERA5 global reanalysis from 1940 to 2022. Q. J. R. Meteorol. Soc. 150, 4014–4048. https://doi.org/10.1002/qj.4803
- Aires F, Prigent C, Rossow WB (2005). Sensitivity of satellite microwave and infrared observations to soil moisture at a global scale: 2. Global statistical relationships. J. Geophys. Res. Atmos. 110. https://doi.org/10.1029/2004JD005094
- Rodriguez-Fernandez N, Aires F, Richaume P, Kerr Y H, Prigent C, Kolassa J, et al. (2015). Soil moisture retrieval using neural networks: Application to SMOS. IEEE Transactions on Geoscience and Remote Sensing 53, 5991–6007. https://doi.org/10.1109/TGRS.2015.2430845
- Rodriguez-Fernandez N, de Rosnay P, Albergel C, Richaume P, Aires F, Prigent C, et al. (2019). SMOS neural network soil moisture data assimilation in a land surface model and atmospheric impact. Remote Sens. 11. https://doi.org/10.3390/rs11111334
Citation: https://doi.org/10.5194/egusphere-2026-2360-AC1
-
RC2: 'Comment on egusphere-2026-2360', Anonymous Referee #2, 07 Jun 2026
Review of
Evaluation of ASCAT soil moisture retrievals and their potential to detect intraday variability
by Dinh et al.
General comments:
This is an interesting paper demonstrating the potential of machine learning to enhance existing soil moisture products derived from EUMETSAT's ASCAT instruments on MetOp satellites. A neural network (NN) is trained using ERA5 simulations as a reference. The resulting soil moisture retrievals are less noisy than the original ASCAT soil moisture product. In addition, the authors provide some evidence suggesting that sub-daily soil moisture variability could be observed through NN retrievals. While the paper is well written, the addition of a diagram presenting the NN training and retrievals with inputs and outputs would be useful. Also missing is a table summarising mean score values per experiment with respect to in situ observations. In addition to scores covering the entire four-year period, the table should also indicate seasonal scores (e.g. DJF, MAM, JJA and SON across years). Finally, the approach is somewhat circular since ERA5-derived variables (ST and LAI) are used to predict soil moisture, as well as ASCAT backscatters. This shortcoming should be clearly acknowledged, and the reasons for it should be made clear.
Recommendation: major revisions.
Particular comment:
- L. 41: "SM estimates at sub-daily resolution" is not clear. Do you mean that several ASCAT observations are available for a given location on the same day? Or does the time of observation vary from one day to the next?
- L. 44: ASCAT observations are global. Why focusing on CONUS? Any reason for that?
- L. 102: Does “the ERA5 LAI product” exist? As far as I am aware, ERA5 neither simulates dynamic LAI nor integrates LAI observations. This sentence needs to be clarified.
- L. 225: Replace "extracting a reliable diurnal signal is a true challenge!" by "extracting a reliable signal is challenging".
- L. 232 (Fig. 4): ERA5 precipitation data can be subject to significant biases. Could you include in situ precipitation observations in Fig. 4b?
- L. 240: This study does not actually demonstrate the intraday capability of the method. In situ observations of precipitation would provide a more convincing demonstration.
- L. 252: "section 55.1"?
Citation: https://doi.org/10.5194/egusphere-2026-2360-RC2 -
AC2: 'Reply on RC2', Lan Anh Dinh, 08 Jul 2026
Dear Referee #2,
Thank you for your thorough review and insightful suggestions. Below, we provide our point-by-point responses to your comments for this interactive discussion. We look forward to implementing these outlined changes in the revised version of the manuscript in the next phase.
General comments:
This is an interesting paper demonstrating the potential of machine learning to enhance existing soil moisture products derived from EUMETSAT's ASCAT instruments on MetOp satellites. A neural network (NN) is trained using ERA5 simulations as a reference. The resulting soil moisture retrievals are less noisy than the original ASCAT soil moisture product. In addition, the authors provide some evidence suggesting that sub-daily soil moisture variability could be observed through NN retrievals. While the paper is well written, the addition of a diagram presenting the NN training and retrievals with inputs and outputs would be useful.
Response: Thank you for your appreciation of our work. As also mentioned in the response to Referee #1, we will expand the model description and add a detailed diagram describing the CNN architecture and processing steps in the revised manuscript.
Also missing is a table summarising mean score values per experiment with respect to in situ observations. In addition to scores covering the entire four-year period, the table should also indicate seasonal scores (e.g. DJF, MAM, JJA and SON across years).
Response: Thank you for this helpful suggestion. We agree that this addition will improve the presentation of the results, and this information will be included in the revised manuscript.
Finally, the approach is somewhat circular since ERA5-derived variables (ST and LAI) are used to predict soil moisture, as well as ASCAT backscatters. This shortcoming should be clearly acknowledged, and the reasons for it should be made clear.
Response: We agree that this aspect of the experimental design should be more clearly acknowledged and justified. Our framework uses ERA5 SM as the training target, together with ERA5 ST and LAI as ancillary inputs, thereby introducing dependence on the ERA5 framework. The choice of ERA5 SM as the training target is motivated by several considerations. ERA5 provides one of the most comprehensive large-scale SM estimates currently available, combining a physically based land surface model with the assimilation of multiple satellite observations, atmospheric forcing, and available in situ information (Soci et al., 2024). Consequently, it offers a spatially and temporally consistent reference dataset for supervised learning, an approach that has been widely adopted in remote sensing applications (Aires et al., 2005; Rodríguez-Fernández et al., 2015; Pellet et al., 2025).
The inclusion of ST and LAI is also physically motivated. Microwave backscatter is highly sensitive to vegetation attenuation and the thermodynamic state of the land surface. These variables thus provide boundary conditions that help the model separate SM effects from other sources of variability in the ASCAT signal. Methodologically, using ERA5 ancillary variables ensures the retrieved ASCAT product remains fully consistent with the ERA5 background space, which is a standard requirement for data assimilation preparatory frameworks (e.g., Rodríguez-Fernández et al., 2019). We acknowledge that the retrieved product inherits the large-scale climatology and background state of ERA5. However, the sub-daily and event-scale temporal variations are primarily constrained by the independent ASCAT observations, which provide information that is not contained in the ERA5 ancillary variables alone.
Finally, model performance was evaluated using independent in situ SM observations from the ISMN, which were not used during training. Although this evaluation cannot eliminate the dependence on ERA5 introduced during model development, it provides an independent assessment of whether the retrieval produces physically meaningful SM estimates. The fact that the retrieval shows comparable, and in some locations improved, agreement with in situ observations relative to ERA5 indicates that the model is able to extract additional information from the ASCAT observations beyond simply reproducing the ERA5 background.
We will revise the manuscript to explicitly acknowledge this dependence on ERA5 and to clarify the physical and methodological rationale for the chosen framework.
Particular comments:
- L. 41: "SM estimates at sub-daily resolution" is not clear. Do you mean that several ASCAT observations are available for a given location on the same day? Or does the time of observation vary from one day to the next?
Response: We are referring to the availability of multiple ASCAT observations for a given location on the same day (corresponding to native overpass times), rather than daily aggregated soil moisture. We will make this distinction clearer in the revised manuscript.
- L. 44: ASCAT observations are global. Why focusing on CONUS? Any reason for that?
Response: Here we selected the CONUS as it offers a highly dense network of in situ SM measurements (Marinescu et al., 2024), which are essential for robust validation. Also, CONUS encompasses pronounced spatial and climatic variability, making it an ideal testbed (Bernhardt et al., 2018). We will explicitly state this rationale in the introduction section (after L. 44) of the revised manuscript to clarify our selection for the readers.
- L. 102: Does “the ERA5 LAI product” exist? As far as I am aware, ERA5 neither simulates dynamic LAI nor integrates LAI observations. This sentence needs to be clarified.
Response: Thank you for pointing this out. We fully agree; ERA5 does not simulate a dynamic LAI nor assimilate LAI observations, but rather uses a prescribed monthly climatology (Duveiller et al., 2023). Our original phrasing ("the ERA5 LAI product") was inadvertently misleading, as it implied a dynamically varying dataset.
However, it is important to note here that our rationale for using this specific climatological background, rather than dynamic external satellite datasets, was to ensure strict physical consistency with our training target. Feeding the model dynamic vegetation anomalies that the ERA5 soil moisture target inherently does not possess would introduce an input-target mismatch. We will clarify this point in the revised manuscript.
- L. 225: Replace "extracting a reliable diurnal signal is a true challenge!" by "extracting a reliable signal is challenging".
Response: We agree with this change.
- L. 232 (Fig. 4): ERA5 precipitation data can be subject to significant biases. Could you include in situ precipitation observations in Fig. 4b? ; - L. 240: This study does not actually demonstrate the intraday capability of the method. In situ observations of precipitation would provide a more convincing demonstration.
Response: We thank you for this valuable suggestion. The primary purpose of using ERA5 precipitation (as in Fig. 4a) was to illustrate the spatial pattern of the precipitation even over the CONUS. However, we agree that including an independent observational precipitation record would strengthen the interpretation of the case study presented in Fig. 4b.
In the revised manuscript, we will attempt to obtain precipitation observations for the selected location from available in situ gauge networks, such as the NOAA Global Historical Climatology Network-hourly (GHCNh) or the ISMN, where available. If no suitable gauge observations are available for the selected event, we will instead use a high-resolution gauge-adjusted precipitation product (e.g., NOAA Stage IV) to provide an observational reference. This will allow us to independently verify the timing and magnitude of the rainfall event, thereby strengthening the interpretation of the SM response without relying solely on ERA5 data.
We also agree that the original wording in Line 240 overstated the conclusions that can be drawn from a single case study. Accordingly, we will revise the text to clarify that this example is intended to illustrate the model’s behavior in capturing intraday SM variability during a heavy precipitation event, rather than to provide a formal demonstration of model capability. By incorporating an independent observational precipitation dataset, we will also provide an objective reference for assessing whether the timing of the retrieved SM response is consistent with the observed rainfall, rather than reflecting only the ERA5 forcing used during model development.
- L. 252: "section 55.1"?
Response: We referred to section 5.1 here. We will correct this in the revised manuscript.
References:
- Marinescu P J, Abdi D, Hilburn K, Jankov I, Lin L (2024). An Evaluation of NOAA Modeled and In Situ Soil Moisture Values and Variability across the Continental United States. Weather and Forecasting, 39(3), 523-540. https://doi.org/10.1175/WAF-D-23-0136.1
- Bernhardt J, Carleton A M, LaMagna C (2018). A comparison of daily temperature-averaging methods: spatial variability and recent change for the conus. J. Clim.31, 979–996. https://doi.org/10.1175/JCLI-D-17-0089.1
- Duveiller G, Pickering M, Muñoz-Sabater J, Caporaso L, Boussetta S, Balsamo G, Cescatti A (2023). Getting the leaves right matters for estimating temperature extremes, Geosci. Model Dev., 16, 7357–7373, https://doi.org/10.5194/gmd-16-7357-2023
- Soci C, Hersbach H, Simmons A, Poli P, Bell B, Berrisford P, et al. (2024). The ERA5 global reanalysis from 1940 to 2022. Q. J. R. Meteorol. Soc. 150, 4014–4048. https://doi.org/10.1002/qj.4803
- Aires F, Prigent C, Rossow WB (2005). Sensitivity of satellite microwave and infrared observations to soil moisture at a global scale: 2. Global statistical relationships. J. Geophys. Res. Atmos. 110. https://doi.org/10.1029/2004JD005094
- Rodriguez-Fernandez N, Aires F, Richaume P, Kerr Y H, Prigent C, Kolassa J, et al. (2015). Soil moisture retrieval using neural networks: Application to SMOS. IEEE Transactions on Geoscience and Remote Sensing 53, 5991–6007. https://doi.org/10.1109/TGRS.2015.2430845
- Rodriguez-Fernandez N, de Rosnay P, Albergel C, Richaume P, Aires F, Prigent C, et al. (2019). SMOS neural network soil moisture data assimilation in a land surface model and atmospheric impact. Remote Sens. 11. https://doi.org/10.3390/rs11111334
- Pellet V, Aires F, Boucher E, Volden E (2025). Enhancing soil moisture statistical retrieval from SMOS using partial convolutions and localization strategies. J. Appl. Meteorol. Climatol. 64, 1951–1965. https://doi.org/10.1175/JAMC-D-25-0041.1
Citation: https://doi.org/10.5194/egusphere-2026-2360-AC2
-
AC2: 'Reply on RC2', Lan Anh Dinh, 08 Jul 2026
Peer review completion
Journal article(s) based on this preprint
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 298 | 124 | 45 | 467 | 47 | 53 |
- HTML: 298
- PDF: 124
- XML: 45
- Total: 467
- BibTeX: 47
- EndNote: 53
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Filipe Aires
Victor Pellet
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(3694 KB) - Metadata XML
The study presents a novel approach for retrieving ASCAT soil moisture via a convolutional neural network (CNN) that explicitly models spatial dependencies. The validation is robust, using both the ERA5 reanalysis and in situ soil moisture from the ISMN. Furthermore, the performance of the CNN-based ASCAT SM is compared with the H SAF ASCAT SM data record (derived using the change detection approach). The results are convincing, with improvements to the correlation coefficient for the CNN ASCAT SM relative to the H SAF ASCAT SM. However, the study would benefit from a clearer description of the machine learning architecture for reproducibility. Also, stratifying the ISMN validation results according to land cover and vegetation types would be informative.
Major comments:
Minor comments: