A neural-network based forward operator for the assimilation of microwave satellite observations with LDAS-Monde

Vural, Jasmin; Vanderbecken, Pierre; Bonan, Bertrand; Rojas-Munoz, Oscar; Calvet, Jean-Christophe

doi:10.5194/egusphere-2026-838

Preprints

https://doi.org/10.5194/egusphere-2026-838

Preprints

25 Feb 2026

| 25 Feb 2026

A neural-network based forward operator for the assimilation of microwave satellite observations with LDAS-Monde

Jasmin Vural, Pierre Vanderbecken, Bertrand Bonan, Oscar Rojas-Munoz, and Jean-Christophe Calvet

Abstract. The knowledge of land-surface variables (LSVs) is essential for an accurate description of the carbon cycle and, hence, for deriving the fraction of anthropogenic CO₂ emissions contributing to global warming. For a better representation of LSVs like the leaf-area index (LAI), we assimilate the observations of the satellite microwave sensors SMAP and AMSR2 with our land-data assimilation system (LDAS). Dedicated observation operators using a neural network (NN) are developed to enable a direct assimilation of the measured brightness temperatures. Direct assimilation is not yet an established method due to the incomplete representation of physical emission processes and the associated computational constraints of physical observation operators. We derive an optimal set of predictors for both instruments resulting in a good match between model equivalent and the observations in the testing period, with correlations up to 0.87/0.93 and a total RMSE of 8.2/3.6 K for SMAP/AMSR2, respectively. The implementation of the derived weights into the LDAS is straightforward and is found to lead to a reasonably good performance of the assimilation system, with a stronger improvement of the departures for SMAP than for AMSR2, which can be attributed to the characteristics of the microwave bands of the observations. For the best assimilation experiments, the verification against LAI observations shows an improvement compared to the open loop on a global scale. For different global cropland regions especially prone to droughts, AMSR2 outperforms SMAP in most cases. Nevertheless, the seasonal and subseasonal variability is still not well represented though due to unsolved issues in our model.

Received: 11 Feb 2026 – Discussion started: 25 Feb 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Jasmin Vural, Pierre Vanderbecken, Bertrand Bonan, Oscar Rojas-Munoz, and Jean-Christophe Calvet

Status: final response (author comments only)

Subscribe to comment alert

RC1:
'Comment on egusphere-2026-838', Anonymous Referee #1, 01 Apr 2026
General comments:
This study proposes observation operators to enable the assimilation of SMAP 1.4 GHz vertically polarized signals and AMSR2 10.65 GHz vertically polarized signals, with the objective of improving LAI and soil moisture retrievals within LDAS-Monde. The analysis focuses on cropland areas at the global scale. To facilitate the direct assimilation of passive microwave brightness temperatures (TB), the authors develop an observation operator based on neural networks, linking geophysical and static predictors to the observed TB.
The implementation of these forward operators within LDAS demonstrates improvements in the departures, particularly for SMAP observations, likely due to the higher variability of TB at L-band compared to X-band. Regarding LAI, the results indicate a positive impact at the global scale. However, at regional scales, the results appear more mixed and depend on the evaluation metric, with improvements mainly observed in terms of bias.
I fully agree with the authors on the importance of such studies aiming at the direct assimilation of observations, and I support the general methodology presented here. I also recognize the amount of work from designing the observation operator to experiments within an assimilation system.
However, I have two major concerns: 1) I find it limiting that only vertical polarization is exploited, while horizontal polarization is also available and contains valuable information. Although I understand the challenges associated with developing observation operators for H-polarization, it would have been beneficial to include results for this polarization, even if their performance is inferior to that of V-polarization. Presenting such results, at least for observation operator, would provide a more complete picture. Scientific progress also relies on documenting difficulties and limitations, which can guide future developments toward assimilating a broader range of observations; 2) The manuscript is sometimes difficult to follow. Certain information appears to be missing or is introduced only later in the text, which affects the overall clarity and understanding especially for reader which is not exactly in the field. More detailed comments are provided in the specific remarks below. I tried to separate them between major and minor even though some might in the middle.

Specific comments:
Major
Title and abstract: It is not very clear for the reader that the study focuses only on cropland areas, despite being conducted at the global scale. This should be explicitly stated in both the title and the abstract to avoid potential misunderstanding.

Section 4.1: From my point view, important information regarding the NN configuration is missing. While the outputs are the brightness temperatures (TBs), it is unclear whether a single NN is used to predict both AMSR2 and SMAP TBs, or whether separate networks are trained for each instrument. This information only appears later in Section 5.1, from which I infer that different NNs are used. This point should be clarified earlier in the manuscript, within the methodological description of the observation operator.

Section 4.1: The study focuses only on vertical polarization. In Corchia et al. (2023), this choice is justified because they rely on ASCAT active instrument, which only provides vertical polarization. However, in the present study, both passive instruments SMAP and AMSR2 provide horizontal polarization, which contains important information, particularly regarding vegetation. Several studies (e.g., [1–6]) demonstrate the added value of combining both V- and H-polarizations for vegetation-related applications. The omission of H-polarization therefore requires stronger justification or, ideally, complementary analysis.

Section 4.1: The statement “likely due to lower emissivity” is somewhat misleading. While it is true that emissivity at H-polarization is generally lower than at V-polarization, the main difficulty in reproducing H-polarized signals arises from their higher variability compared to V-polarization. Moreover, the cited reference attributes the difference in performance to the “lower emissivity variability observed at V polarization,” rather than lower emissivity at H-polarization. Thus, H-polarized signals are more challenging to model because of their strong variability under varying surface conditions (e.g., vegetation, soil moisture, surface water presence, terrain roughness). This point should be clarified or corrected to avoid misinterpretation.

Section 4.1 and Table 1: A clearer and more detailed description of the potential input parameters is needed. Some variables may not be easily understood by readers unfamiliar with the ISBA model or ECOCLIMAP-SG. For example, the choice of using soil water index and soil temperature from layer 2 (and not another layer) is not explained until Section 4.2. Or for example what do you expect the long/short wave radiation brings to observed signal. More generally, it would be beneficial to briefly describe the expected relationships between the presented input parameters and the observed TB. This could be included in Section 3 to better motivate the selection of predictors.

Fig 3: The results obtained with AMSR2 appear to exhibit a negative bias. Could the authors provide insight into the possible reasons for this behaviour, particularly if separate neural networks are trained for each instrument (as suggested in Section 5.1)?

Additionally, the readability of the figure could be improved. The ordering of the experiment numbers seems shuffled, making it even more difficult to relate each case to its corresponding input parameters. As a minor suggestion, inverting the color scale (e.g., blue for lower values) could improve interpretability.

Data Assimilation: Data Assimilation: From what I understood, the assimilation tests are made separately for each sensor, which is important to understand what each observation is bringing. Did you also tried to assimilate both instruments to see if the information may be complementary?

Conclusion (l.367): The conclusion states that SMAP outperforms AMSR2, which seems inconsistent with statements made in the abstract and Section 5. This discrepancy should be clarified to ensure consistency throughout the manuscript.

Conclusion: This point is linked to point 3. The conclusion should explicitly recall that the experiments are conducted using only vertical polarization, and for AMSR2 only at X-band. Given the known sensitivity of horizontal polarization to vegetation, it would be important to acknowledge this limitation and at least briefly discuss its implications.

Minor
Abstract: It would be helpful to briefly specify what is meant by "unsolved issues," as this term can be quite broad. Additionally, the reference to "the model" is somewhat ambiguous, as it could refer either to the land surface model or to the neural networks used for the observation operator.

Introduction (l.23): The expression “providing gridded data” when referring to passive microwave instruments is somewhat misleading. Given the acquisition geometry of these instruments, the data are not inherently gridded but are typically resampled on grid.

AMSR2 observation: Only the 10.65 GHz channel is used in this study. It would have been interesting to include additional frequencies, as their sensitivity to vegetation has been demonstrated in several studies (e.g., 6 GHz [3], 18 GHz [4], 36 GHz [1], or multi-frequency approaches [6]). I list this as a minor comment, as extending the analysis to other frequencies would likely require substantial additional work, but it would be worth mentioning in the discussion or perspectives.

A temporal resolution of 3 hours is mentioned, whereas both SMAP and AMSR2 provide only two overpasses per day. This is somewhat confusing, and the distinction between model time step and observation availability should be clarified.

(l. 135): For readers unfamiliar with ISBA, the terms C3 and C4 are not meaning-full and should be defined.

Fig 1: To facilitate the reading, it would be helpful to include the tile names directly on the map.

Fig 4: For readability, it would be preferable to use full parameter names instead of acronyms.

Paragraph 4.2: The neural network is trained to predict brightness temperatures (TB). Given the radiative transfer equation, it is expected that temperature-related variables play a major role. Did the authors analyse intercorrelations between the input parameters? This could help better interpret the model behaviour.

L 180 181: “This indicates that AMSR2 brightness temperatures are more sensitive to short-range diurnal variations.” Considering that AMSR2 overpass times (≈1:30 AM/PM) differ significantly from those of SMAP (≈6 AM/PM), could this apparent sensitivity be related to observation timing rather than intrinsic sensor characteristics? This comment also relates to the need for a clearer description of input parameters and their physical links to the observations (see major comment above).

L203-206: This behaviour is expected when latitude, longitude, and day-of-year are used as predictors. If certain regions or seasons are underrepresented in the training dataset, the model may struggle to generalize to those conditions.

L230: Should the same observation error be used for both SMOS and SMAP, given that, although they operate at similar frequencies, the instruments have different characteristics?

Fig 8: The comparison is performed at the global scale. However, it might be more rigorous to analyse results separately by hemisphere, as seasonal cycles differ and could impact vegetation-related signals. Given that most croplands are located in the Northern Hemisphere, the figure likely reflects primarily Northern Hemisphere seasonality. Additionally, the legend size could be increased for better readability.

Fig 10: It is not clear over which period these maps are computed. Additionally, the legend title is too small and difficult to read.

Paragraph 5.6: Given that L-band is generally more sensitive to soil moisture than X-band, the reported sensitivity seems somewhat counterintuitive and would benefit from further explanation.

Discussion (l.366-369). The term “unremarkable” is unclear in this context and should be clarified. The analysis presented here is in fact interesting.

Technical corrections:
(l. 43) GPP acronym appears in the introduction but is only defined in section 2.4

(l 77) ‘within one the time interval of 3 hours’: the formulation seems strange; I would have use: within a 3-hour time interval

Paragraph between lines 182-185: Sentences are very long.

References:
[1] Choudhury, B. J., & Tucker, C. J. (1987). Monitoring global vegetation using Nimbus-7 37 GHz Data Some empirical relations. International Journal of Remote Sensing, 8(7), 1085-1090.
[2] Becker, F., & Choudhury, B. J. (1988). Relative sensitivity of normalized difference vegetation index (NDVI) and microwave polarization difference index (MPDI) for vegetation and desertification monitoring. Remote sensing of environment, 24(2), 297-311.
[3] Owe, M., de Jeu, R., & Walker, J. (2002). A methodology for surface soil moisture and vegetation optical depth retrieval using the microwave polarization difference index. IEEE Transactions on Geoscience and Remote Sensing, 39(8), 1643-1654.
[4] Jones, M. O., Jones, L. A., Kimball, J. S., & McDonald, K. C. (2011). Satellite passive microwave remote sensing for monitoring global land surface phenology. Remote Sensing of Environment, 115(4), 1102-1114.
[5] Munchak, S. J., Ringerud, S., Brucker, L., You, Y., De Gelis, I., & Prigent, C. (2020). An active–passive microwave land surface database from GPM. IEEE Transactions on Geoscience and Remote Sensing, 58(9), 6224-6242.
[6] Prigent, C., Jimenez, C., Santoro, M., Cartus, O., & Favrichon, S. (2026). Assessing the combination of passive and active microwave satellite observations (1.4 to 36 GHz) to estimate above ground biomass (AGB) globally. Science of Remote Sensing, 100386.
Citation: https://doi.org/10.5194/egusphere-2026-838-RC1
RC2: 'Comment on egusphere-2026-838', Anonymous Referee #2, 10 Apr 2026

This manuscript presents a neural-network-based observation operator to enable the direct assimilation of passive microwave brightness temperatures from SMAP (L-band) and AMSR2 (X-band) within LDAS-Monde, with the aim of improving soil moisture and leaf area index (LAI) estimates. The approach links model states and auxiliary predictors to observed TB, and evaluates the impact of assimilating each sensor (separately) over cropland-dominated regions at the global scale. Results show reductions in TB departures and some improvements in LAI—more consistent at the global scale than regionally—while the impact on soil moisture and the physical consistency of the framework are less clear. In addition, I found it difficult to reconcile the relatively weak sensitivity of the NN to LAI (compared to SWI) with the magnitude of the resulting LAI increments, which raises further questions about how information is being propagated through the system.
Main concerns (summary):
I have some reservations about the physical consistency of the observation operator as currently presented. The use of observed LAI, along with static predictors (e.g. latitude/longitude) and even forcing variables within the NN, makes it less clear that this is a clean mapping from model state to observation. In particular, it raises questions about how the operator remains dynamically consistent with the model state, and what exactly the NN is learning (physical relationships vs. geographical or climatological patterns). Clarifying the role of these predictors, and why they are required, would strengthen confidence in the approach.
More generally, I found parts of the methodology difficult to follow, with some key elements either not clearly described or introduced relatively late. In particular, it is not always clear whether observed or modeled LAI is used within the operator, how anomalies are defined, what exactly is included in the control vector, and how the assimilation timing is implemented. These are all important for interpretation and reproducibility, and would benefit from a clearer and more explicit description earlier in the manuscript.
Finally, I am not fully convinced by the experimental design and its interpretation. The rationale for assimilating the different sensors separately is not entirely clear, and it would be useful to understand whether their information content is complementary when used together. Similarly, the role of static and forcing predictors, and the apparent mismatch between NN sensitivities and the resulting analysis increments, are not fully explained. Addressing these points would help clarify how the system is behaving and how robust the conclusions are.
Overall, I find the topic timely and the approach promising, and I appreciate the effort involved in developing and implementing the NN-based operator within LDAS-Monde. However, the concerns outlined above—particularly regarding the physical consistency of the observation operator, the clarity of the methodological description, and aspects of the experimental design—make it difficult to fully interpret and trust the results as currently presented. Addressing these points would significantly strengthen the manuscript.
Major comments:
L-1 "A neural-network based forward operator for the assimilation of microwave satellite observations with LDAS-Monde" The title is somewhat generic and does not fully reflect the scope of the study. In particular, the focus on LAI and soil moisture is not evident, despite being central to the results. It would be helpful to revise the title to better highlight the scientific objectives and outcomes of the work, rather than focusing primarily on the methodological aspect.
L1 Abstract. The abstract would benefit from a more balanced description of the results and limitations. In particular, the improvements appear more consistent for LAI than for soil moisture, and vary depending on spatial scale, which is not clearly reflected. It would also be helpful to more explicitly acknowledge key aspects of the experimental setup (e.g. cropland focus), as these define the scope of the conclusions.
L23 “observing in near-real time and providing gridded data” Not sure what “gridded” adds here — this could be clarified or removed.
L53 “assimilating land-surface variables using SMAP and AMSR2 observations” maybe “assimilating land-surface variables using both SMAP and AMSR2 observations”?
L67 “interpolated linearly onto the model grid” This seems a little basic — should this not be part of the observation operator (ie its model states that are interpolated, not observations), rather than just preprocessing? Could be clarified.
L67 What is the resolution of the model grid?
L95 “converted into anomalies by subtracting the mean from the actual value for each data set and scaling with the standard deviation.” Mean over what exactly (grid cell? time period?) — this should be specified.
L106 “The LAI is computed once per day.” So LAI is a fully prognostic variable?
L106 “interaction between CO2 and photosynthesis with regard to stomatal respiration and plant growth” “stomatal respiration” seems like an unusual description — could this be clarified?
L122 “control variable xj (here LAI and soil moistures of layers 2-7, i.e., 0.01- 1 m depth)” It would help to state more explicitly which model variables are included in the control vector, and why they were chosen. Should surface/soil temperature also be included?
L133 “croplands (Corchia et al., 2023; Shan et al., 2024), we use exclusively grid cells that are dominated by crops.” The focus on croplands should be introduced earlier, as it has important implications for the interpretation of the results.
L160 and Table 1 “variable model fields as predictors” It would be helpful to clarify how this list of predictors was selected.
L172 “In addition to the predictors that have already been employed in similar studies, we achieved a significant improvement of the model equivalent when using forcing data.” Clarify exactly what you mean by "forcing data" here.
L175 I’m not sure about having latitude and longitude as static predictors. What do they really represent?
L185 “strong benefit seen from adding the forcing variables” I assume this refers to meteorological forcing used to drive the model? These seem unusual to include in an observation operator, as they break the mapping from model state to observation, and would benefit from further justification.
L213 “observation operator described in Sect. 4” The observation operator is not described clearly enough here — in particular, it is not clear whether observed or modeled LAI is used. It seems the NN is trained with observed LAI?
L218 “LDAS-Monde is set up to work on a 24 h assimilation window and the observations are assimilated at a given hour. To account for the diurnal variability of the observations, we ran four separate assimilation cycles using 6-hour time bins around each of 0300UTC, 0900UTC, 1500UTC, and 2100UTC, which correspond to the model output times in our setup.” Don’t really understand what you have done here – clarify.
L263 Figure 8 is difficult to read in its current form — in particular, SWI2 is indecipherable.
L315 “static predictors are essential for our data sets” This would benefit from further elaboration — in particular, why a coordinate-independent NN is not sufficient in this case. Why is a coordinate-independent NN seemingly not possible with this approach?
Minor Comments:
The manuscript would benefit from some minor improvements in terminology consistency (e.g. “forward operator” vs. “observation operator”, TB vs brightness temperature), clarification of certain terms (e.g. “model equivalent”), and overall readability (sentence length, figure referencing)

Citation: https://doi.org/10.5194/egusphere-2026-838-RC2

Jasmin Vural, Pierre Vanderbecken, Bertrand Bonan, Oscar Rojas-Munoz, and Jean-Christophe Calvet

Viewed

Total article views: 145 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
79	49	17	145	10	19

HTML: 79
PDF: 49
XML: 17
Total: 145
BibTeX: 10
EndNote: 19

Views and downloads (calculated since 25 Feb 2026)

Month	HTML	PDF	XML	Total
Feb 2026	30	8	2	40
Mar 2026	37	31	12	80
Apr 2026	12	10	3	25

Cumulative views and downloads (calculated since 25 Feb 2026)

Month	HTML	PDF	XML	Total
Feb 2026	30	8	2	40
Mar 2026	37	31	12	80
Apr 2026	12	10	3	25

Viewed (geographical distribution)

Total article views: 135 (including HTML, PDF, and XML) Thereof 135 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 14 Apr 2026

Short summary

New satellite observations are used to improve our understanding of agricultural land on a global scale. Innovative artificial intelligence methods are employed to combine the available information. This improvement is evaluated using independent observations and varies depending on the characteristics of instruments and models.


Total:	0
HTML:	0
PDF:	0
XML:	0