the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 17 Mar 2025
All-Sky Temperature and Humidity Retrieval from the MWRI-RM Onboard the FY-3G Satellite
Abstract. To investigate the application of deep learning in satellite remote sensing, this study employs brightness temperature observations from the remapped Micro-Wave Radiation Imager-Rainfall Mission (MWRI-RM) onboard the Fengyun-3G (FY-3G) satellite as input data, while temperature and humidity profiles (ranging from 1000 hPa to 100 hPa) obtained from ERA5 reanalysis data are used as label data. An Advanced Residual Convolutional Neural Network (AR-CNN) model was developed to retrieve atmospheric temperature and humidity profile data. The results show that: (1) The retrieval of temperature profiles has a root mean square error (RMSE) of approximately 1.24 K, and the RMSE for humidity profile retrieval is 12.98 %. (2) A comparison between predicted and labeled samples reveals consistent results for temperature retrieval but inconsistencies in high-humidity regions, indicating that further refinement of the model is needed in these areas. (3) Gradient backpropagation and perturbation experiments demonstrate that channels near 118 GHz are critical for retrieving upper-level temperatures, and those near 183 GHz mainly affect mid-to-lower atmospheric temperature retrieval. For humidity, channels near 183 GHz are essential for detecting mid-to-lower water vapor, and the 118 GHz oxygen absorption channel is indispensable for upper-level humidity retrieval. This suggests that the model possesses a certain degree of interpretability and stability.
-
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
(8585 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(8585 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2025-680', Anonymous Referee #1, 10 Apr 2025
-
AC1: 'Reply on RC1', Wei Han, 23 Apr 2025
Thank you for your insightful feedback. We appreciate your concerns and address them as follows:
- Novelty and Evidence for AR-CNN
Our study focuses on evaluating the contributions of the 118 GHz and 183 GHz microwave channels to temperature and humidity profile retrievals, leveraging the enhanced capabilities of the FY-3G/MWRI-RM. While machine learning (ML) has been applied to atmospheric retrievals, the novelty lies in the Advanced Residual CNN (AR-CNN) architecture, which integrates residual blocks and adaptive pooling to better capture spatial features from 26-channel microwave observations. To validate AR-CNN’s superiority, we compared it with MLP and standard CNN models (not explicitly tabulated in the original manuscript). The AR-CNN achieved a lower RMSE (1.24 K for temperature, 12.98% for humidity) compared to MLP (1.52 K, 15.3%) and CNN (1.38 K, 14.1%), demonstrating its effectiveness in handling high-dimensional, nonlinear satellite data. - Structure and Methodology Clarification
Regarding the arrangement of the paragraph, we apologize for any confusion caused by the unedited paragraphs. We will thoroughly edit the manuscript to ensure a clear and logical structure. We will also add more paragraphs to support the use of the AR-CNN architecture and provide a detailed discussion of the results. - Physical Interpretability and Appendix Figures
The results (Sections 4.1–4.2) are grounded in gradient backpropagation and perturbation experiments, which quantitatively link 118 GHz and 183 GHz channels to their respective atmospheric layers (e.g., 118 GHz for upper-level retrievals, 183 GHz for mid-to-lower layers). These align with Jacobian analyses (Figs. 5, 9) and confirm the model’s interpretability. The appendix figures (e.g., global maps in Fig. A1–A2) visually validate spatial consistency between retrievals and ERA5, particularly in regions with high humidity gradients or dynamic processes. Future revisions will explicitly reference these figures in the discussion. - Purpose of ERA5-Based Retrieval
The goal is not to replicate ERA5 but to explore satellite-driven retrievals with higher spatiotemporal resolution. ERA5 serves as a reanalysis benchmark due to its global coverage and assimilation of multi-source data. While ERA5 uncertainties exist, our model reduces systematic biases and provides independent retrievals for regions with sparse radiosonde data.
Citation: https://doi.org/10.5194/egusphere-2025-680-AC1 - Novelty and Evidence for AR-CNN
-
AC1: 'Reply on RC1', Wei Han, 23 Apr 2025
-
RC2: 'Comment on egusphere-2025-680', Anonymous Referee #2, 29 Apr 2025
This paper presents a algorithm for use on the microwave imager instrument aboard FY-3G to retrieve profiles of temperature and humidity. The algorithm is a convolutional neural network (CNN) that was trained with ERA5 data, and because it is in the microwave part of the spectrum it can retrieve all sky conditions. Neural network retrievals are actively being improved in the literature, especially multidimensional algorithms such as convolutional neural networks, so this paper in its ideas is suitable for publication in AMT.
Broader feedback:
From what I can discern, the validation here is strictly whether the CNN is able to reproduce the input data based on the sample learning period. While it is good to utilize this as a check of the algorithm behavior, there should be a validation performed with observations from alternative sensors/algorithms or ground-based observations (including ground-launched radiosondes).
Why was relative humidity (RH) chosen over water vapor mixing ratio or specific humidity? While I am not arguing that RH is invalid to retrieve, RH is not an absolute quantity and is dependent on temperature (which entangles retrieval of waver vapor and temperature). RH makes it more difficult to compare results against other literature in the microwave remote sensing community. Retrieving a water vapor value specifically would make the two values more distinct and provide better illumination of what channels are sensitive to which physical values.
The indication that 183 GHz provides higher information content for temperature sounding than all but one 118 GHz channel is a surprise to me (this statement is based on my interpretation of Figure 5 along with corresponding text). One citation is given for the reasoning behind this but is unfortunately in another language and inaccessible to me. Also, L239-242 states water vapor attenuates the signal (and cites a paper describing an infrared retrieval paper), which seems contrary to the stated conclusion. Detailed information about this analysis methodology and better tying the conclusions to relevant principles should be provided.
The methodology explanation does not line up with explanations in the results sections. Up until section 4.2, I thought the retrieval was only over ocean (viz. L137) given that the training database was constructed from ocean only profiles and the validation dataset is part of the training database. The appendix plots also show only over ocean data. Yet Section 4.2 describes land cover characteristics when discussing the humidity retrieval performance. More detailed information about the database construction (perhaps a spatial map of where the database is compiled from) should be provided and the text clarified to line up with the methods.
Figures 3 (c-h) and 7 (c-h) are not very illustrative because of the density of points. Rather than shading by absolute error, a two-dimensional histogram should be performed and shading indicate the counts in the bins. Additionally, the line plotted on each subplot has no label -- whether it is a least squares fit or a 1-to-1 slope, that should be indicated in the figures. And speaking of least squares fit, are any fitting statistics available for the scatter plots?
In surveying the literature, much of the historical work mentioned was on infrared retrievals. More citations on microwave regime development should be included beyond why the regime is important for all-sky retrievals.
Below is more specific copyediting feedback (which can be rejected where deemed), but overall the wording should be refined to focus on the scientific conclusions drawn from the results at hand without extra pomp or overbroad explanations.
The introduction section was difficult to parse. It is a three page-long paragraph. Consider splitting it up into thematic paragraphs. (This may be simply because of the template used.)
L47-50: Is it the case that the spatiotemporal resolution is higher indeed for microwave instruments, or that the spatial sampling and therefore temporal resolution is higher given all-sky conditions sensing?
L55+: I would not personally exclude neural networks when defining conventional satellite retrievals as done in the separation. While neural networks are still being developed and improved, they are not a super novel concept. In example, L85 cites a neural network retrieval from 15 years ago.
L79-81: This sentence has too much puffery.
Table 1 should be properly cited with a note to the references section; the current listing in the title is inadequate.
L135-139: The first sentence of this paragraph should be reworded to indicate that brightness temperatures from all channels are used--or whatever methodology is appropriate. The static values should be further described.
L142-L144: There is some redundancy in these two sentences about interpolation.
Figure 2: Font size should be increased on within-figure text.
L185-186: This wording is awkward.
L187-188: I would argue this is more a matter of instrument sensitivity vertically than model performance.
L198-200: I would rephrase or better convey the argument here.
L208-219: This is a very broad set of statements that, while touch on the fundamental uncertainties and error sources in remote sensing, are not fully tied into the results at hand. Further, how would you plan to account for this or mitigate this in the future?
L228-231: This is a very abrupt statement as an off-hand comment in parentheses along with the introduction sentence. This should be expanded into a paragraph of its own methodology as I do not understand what has occurred. What is model.conv1? What spatial dimensions?
L231-L246: This paragraph oscillates between highly detailed and very broad. I would reconsider how it is phrased and the statements used to weave together the points. Overall, I understand and appreciate the information content study here.
Figure 6: To directly compare all of these perturbations, the x axes should be the same on all the plots.
L272-282: Similar to L208-219, this is another very broad set of statements that do not fully tie into the results or indicate how/if you plan to address these concerns in the future.
L286-288: These sentences are partially redundant.
L290-295: Repeated phrasing. Further, I would argue that the "complex meteorological process" are the local humidity distributions undergoing alteration--not that the humidity change is an indirect consequence of some other phenomenon.
L307-329: More redundancies, especially when including explanations earlier in this section.
Figure 10: To directly compare all of these perturbations, the x axes should be the same on all the plots.
Citation: https://doi.org/10.5194/egusphere-2025-680-RC2 -
AC2: 'Reply on RC2', Wei Han, 23 May 2025
Overview
This paper presents a algorithm for use on the microwave imager instrument aboard FY-3G to retrieve profiles of temperature and humidity. The algorithm is a convolutional neural network (CNN) that was trained with ERA5 data, and because it is in the microwave part of the spectrum it can retrieve all sky conditions. Neural network retrievals are actively being improved in the literature, especially multidimensional algorithms such as convolutional neural networks, so this paper in its ideas is suitable for publication in AMT.
From what I can discern, the validation here is strictly whether the CNN is able to reproduce the input data based on the sample learning period. While it is good to utilize this as a check of the algorithm behavior, there should be a validation performed with observations from alternative sensors/algorithms or ground-based observations (including ground-launched radiosondes).
Response:
Thank you for your insightful feedback on our validation methodology. We acknowledge the need to strengthen transparency and have clarified our current work’s scope. Specifically, quantitative metrics (e.g., RMSE) for different model architectures are presented in Appendix A, Table A1 of the revised manuscript. These serve as preliminary model-selection indicators during development, without in-depth interpretation, as the study prioritizes demonstrating the framework’s feasibility over exhaustive benchmarking. Additionally, the scarcity of in situ marine observations and spatiotemporal mismatches with FY3G retrievals currently limit robust maritime validation. In follow-up research, we will systematically address these via spatio-temporal alignment algorithms and high-resolution satellite datasets to ensure comprehensive cross-environment evaluation.
Why was relative humidity (RH) chosen over water vapor mixing ratio or specific humidity? While I am not arguing that RH is invalid to retrieve, RH is not an absolute quantity and is dependent on temperature (which entangles retrieval of waver vapor and temperature). RH makes it more difficult to compare results against other literature in the microwave remote sensing community. Retrieving a water vapor value specifically would make the two values more distinct and provide better illumination of what channels are sensitive to which physical values.
Response:
Thank you for your comment on RH selection. This choice is grounded in two key considerations. First, as a standard ERA5 output, RH offers direct usability as label data and aligns with meteorological operational practices, ensuring compatibility with public observations and model inputs. Second, this study represents an initial effort to develop a temperature-humidity joint inversion model using FY-3G/MWRI-RM full-channel data. While RH’s temperature dependence introduces complexity, it enables exploration of dynamic temperature-humidity interactions (e.g., synergies between 118 GHz and 183 GHz channels; see Sections 4.2 and Conclusions in the revised manuscript). We agree that specific humidity retrieval would enhance physical clarity and plan to expand the model’s output to include water vapor parameters in future work, alongside error propagation analysis to decouple temperature-humidity contributions.
The indication that 183 GHz provides higher information content for temperature sounding than all but one 118 GHz channel is a surprise to me (this statement is based on my interpretation of Figure 5 along with corresponding text). One citation is given for the reasoning behind this but is unfortunately in another language and inaccessible to me. Also, L239-242 states water vapor attenuates the signal (and cites a paper describing an infrared retrieval paper), which seems contrary to the stated conclusion. Detailed information about this analysis methodology and better tying the conclusions to relevant principles should be provided.
Response:
To deepen the analysis of 183 GHz channel sensitivity, we have enhanced both physical explanations and literature integration. Through gradient backpropagation and perturbation experiments (Sections 4.1–4.2), we clarify that the 183 GHz water vapor channel indirectly constrains temperature via vapor-radiation feedback: in high-humidity regions (e.g., boundary layers), water vapor’s absorption/emission modifies radiation transfer, influencing temperature through latent heat release. In contrast, the 118 GHz oxygen channel directly senses mid-upper tropospheric temperature gradients, with humidity impacts mediated by dynamics like vertical advection and atmospheric stability. Furthermore, we replaced some Chinese references with English counterparts and added discussions on temperature-humidity coupling (Lines 209–215 in Section 4.1; Lines 279–284 in Section 4.2 of the revised manuscript). The “model weight diffusion in convolutional layers” describes potential propagation of 118 GHz temperature-sensitive features to humidity parameters via spatial feature mixing, a phenomenon warranting further study with physics-constrained neural networks.
The methodology explanation does not line up with explanations in the results sections. Up until section 4.2, I thought the retrieval was only over ocean (viz. L137) given that the training database was constructed from ocean only profiles and the validation dataset is part of the training database. The appendix plots also show only over ocean data. Yet Section 4.2 describes land cover characteristics when discussing the humidity retrieval performance. More detailed information about the database construction (perhaps a spatial map of where the database is compiled from) should be provided and the text clarified to line up with the methods.
Response:
Thank you for identifying the inconsistency in our data description. The term “land cover characteristics” was a typo; both training and validation datasets exclusively use oceanic observations (detailed in Section 2.2). We have corrected relevant paragraphs and removed land-data discussions, explicitly stated in the methodology: “To reduce preprocessing complexity, this study only selects oceanic data.”, highlighted marine data coverage in Appendix Figs. A1–A2. Future research will extend to land scenarios via standardized surface parameter preprocessing (e.g., soil moisture, vegetation index).
Figures 3 (c-h) and 7 (c-h) are not very illustrative because of the density of points. Rather than shading by absolute error, a two-dimensional histogram should be performed and shading indicate the counts in the bins. Additionally, the line plotted on each subplot has no label -- whether it is a least squares fit or a 1-to-1 slope, that should be indicated in the figures. And speaking of least squares fit, are any fitting statistics available for the scatter plots?
Response:
Thank you for your thoughtful suggestions, which have been instrumental in enhancing the clarity and informativeness of our figures.
To address the density of points in Figures 3(c–h) and 7(c–h), we have replaced the absolute error shading with two-dimensional histograms, where color intensity now represents the count of data points in each bin. This modification improves visual interpretability by highlighting data distribution patterns while retaining error magnitude insights. Additionally, the black dashed line in each subplot is explicitly labeled as the “1:1 consistency line” to clarify that it denotes perfect agreement between retrieved and target values, rather than a least squares fit. To further quantify the scatter plot relationships, we have included three statistical metrics—Root Mean Square Error (RMSE), Bias, and Pearson correlation coefficient (r)—in the figure captions and corresponding discussions. These metrics systematically evaluate error magnitude, systematic deviation, and linear correlation, respectively, providing robust statistical support for our validation results.
For other figures, we have implemented the following improvements to ensure consistency and readability:
Figure 2: Font sizes were adjusted to enhance legibility across all text elements, ensuring key labels and annotations are easily distinguishable.
Figures 5 and 9: To emphasize critical channels, Channel 22 (118.75±1.2 GHz) and 26 (183.31±7 GHz) are now plotted with solid lines, while other channels use dash-dot lines for clear visual differentiation. This distinction highlights their unique roles in temperature-humidity retrieval as discussed in Section 4.2.
Figures 6 and 10: The layouts were revised to incorporate pressure-level distributions of percentage changes induced by perturbations. This addition clarifies how the model responds to input variations across different atmospheric layers, strengthening the mechanistic insights presented.
These adjustments collectively aim to align our figures with scientific visualization best practices, ensuring they effectively communicate the study’s key findings while addressing the nuances raised in your feedback. We appreciate your guidance in refining our graphical representations.
In surveying the literature, much of the historical work mentioned was on infrared retrievals. More citations on microwave regime development should be included beyond why the regime is important for all-sky retrievals.
Response:
In response to your suggestion, we revised the introduction to enhance contextual depth:
Lines 55-68 in the revised manuscript: Expanded discussions on microwave remote sensing applications for atmospheric variable retrieval, integrating recent advancements and key studies to underscore the research’s scientific significance.
This revision strengthens the rationale for our approach by connecting it to broader disciplinary progress.Below is more specific copyediting feedback (which can be rejected where deemed), but overall the wording should be refined to focus on the scientific conclusions drawn from the results at hand without extra pomp or overbroad explanations.
Response:
We sincerely appreciate your meticulous copyediting suggestions, which have been invaluable in refining the manuscript’s clarity and scientific rigor. Below is a summary of the key revisions made to address your feedback, with a focus on enhancing readability, streamlining redundancy, and strengthening logical flow. All line numbers cited below correspond to the revised manuscript:
The original three-page introductory paragraph has been restructured into thematic subsections to improve navigability. We also expanded the literature review on microwave remote sensing (e.g., added citations in Lines 55–68) to better contextualize our approach, while adjusting the description of neural networks to acknowledge their established role in the field (avoiding the implication of excluding them from “conventional” methods).
Lines180-187 and Lines254-263: We recognize the need to further integrate broad theoretical discussions with our specific findings, particularly regarding error sources and future mitigation strategies. These aspects will be prioritized in follow-up analyses, including planned investigations into physics-constrained neural network architectures to address model uncertainties.
Broad statements about remote sensing uncertainties were either tied to our specific results or condensed, with clearer links to future mitigation strategies.
Lines199–201: The parenthetical comment about "model.conv1" was expanded into a dedicated paragraph explaining the convolutional layer architecture and spatial dimension considerations, improving methodological transparency.
-
AC2: 'Reply on RC2', Wei Han, 23 May 2025
-
EC1: 'Comment on egusphere-2025-680', S. Joseph Munchak, 10 Mar 2026
Following the decision for minor revisions with editor review, here are my specific comments to address prior to publication:
1. Lines 229 and 328 - the 118.75 +- 1.2 GHz channel is described as being sensitive to temperature gradients. This is not specific - are these horizontal or vertical gradients? For vertical gradients (lapse rate), the channel differences (compared to a channel closer or farther from the 118.75 O2 line) would actually be directly indicative of the gradient. Horizontal gradients would be captured by gradients in the 5x5 patch that is used as input - but it is not clear how this is relevant to the Jacobians. I think a better phrase would be "temperature distributions" instead of "temperature gradients" in these sentences.2. Table A1 - the units of the AR-CNN results are incorrect (and inconsistent with the other rows, which are correct).
Citation: https://doi.org/10.5194/egusphere-2025-680-EC1 -
AC3: 'Reply on EC1', Wei Han, 12 Mar 2026
Dear Editor,
We would like to express our sincere gratitude for your detailed and insightful comments. Your feedback, particularly regarding the precise physical interpretation of the microwave channels, has been immensely helpful in improving the rigor and clarity of our manuscript. We have carefully studied your suggestions and made the corresponding revisions.
Below is a point-by-point response to your comments.
Editor Comment 1: Lines 229 and 328 - the 118.75 +- 1.2 GHz channel is described as being sensitive to temperature gradients. This is not specific - are these horizontal or vertical gradients? ... I think a better phrase would be “temperature distributions” instead of “temperature gradients” in these sentences.
Response: We completely agree with your assessment. You are absolutely correct that describing a single channel’s Jacobian as sensitive to “temperature gradients” is physically inaccurate in this context. As you accurately pointed out, the Jacobians reflect the sensitivity to the state variable itself at specific pressure levels, rather than its spatial derivatives (which would require channel differencing for vertical gradients or spatial feature analysis for horizontal gradients). We highly appreciate you pointing out this conceptual ambiguity.
Editor Comment 2: Table A1 - the units of the AR-CNN results are incorrect (and inconsistent with the other rows, which are correct).
Response: Thank you for catching this oversight. We apologize for the typo in the formatting of the AR-CNN results. We have thoroughly checked Table A1 and corrected the units for the AR-CNN row to ensure complete consistency with the rest of the table.
Furthermore, inspired by your focus on physical rigor, we conducted a thorough review of the manuscript and made two additional proactive refinements to ensure our interpretations of the deep learning model’s behavior are fundamentally sound:
- We revised a sentence regarding the relationship between the 118 GHz channel and humidity. We replaced the original explanation involving “dynamical coupling”and “vertical advection” with a more accurate description of the “thermodynamic correlation.”
- We changed the term “radiometric observations”to “radiance observations” when discussing deep convection areas, as “radiance” is a more precise term for the physical quantity interacting with atmospheric hydrometeors.
Modifications in the Manuscript:
- Line 229 (Revised):“This highlights the model’s utilization of the oxygen absorption line near 118 GHz, a spectral feature with significant sensitivity to mid-to-upper tropospheric temperature distributions (500-200 hPa) and...”
- Line 328 (Revised):“In contrast, the 118.75±1.2 GHz channel is primarily sensitive to mid-to-upper tropospheric temperature distributions (500-200 hPa) via...”
- Table A1:The units for the AR-CNN row have been corrected to perfectly match the baseline formats.
- Additional Revision 1(Line 332): “This apparent discrepancy can be attributed to the thermodynamic correlation between temperature and humidity profiles." (The unsupported second half of the sentence regarding vertical advection was removed).
- Additional Revision 2(Line 287): “These processes, especially in areas with deep convection or liquid water droplets and ice crystals, obscure the direct link between radiance observations and humidity, challenging the model's ability to accurately retrieve humidity values solely from TBs.”
Citation: https://doi.org/10.5194/egusphere-2025-680-AC3
-
AC3: 'Reply on EC1', Wei Han, 12 Mar 2026
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2025-680', Anonymous Referee #1, 10 Apr 2025
The paper used machine learning and microwave channels to retrieve temperature and humidity profiles. However, the paper did not provide enough evidence in the analysis to support the arguments. There is nothing new in the paper in retrieving temperature and humidity profiles using machine learning. The paper should be thoroughly edited, and the authors should provide more evidence.
- The arrangement of the paragraph is chaotic. It seems put a bunch of unedited paragraphs together.
- Not enough paragraph to support of using AR-CNN architecture. What if using other machine learning architecture? The authors did not show the performance in the retrieval algorithm validation.
- A lot of the arguments mentioned in the results are based on the speculation. No other future examination. No discussion to the appendix figures.
- What is the purpose of making an ERA5 temperature and humidity estimator? At best, the model is to get ERA5 temperature and humidity profiles without improving the biases and uncertainties inherent in ERA5.
Citation: https://doi.org/10.5194/egusphere-2025-680-RC1 -
AC1: 'Reply on RC1', Wei Han, 23 Apr 2025
Thank you for your insightful feedback. We appreciate your concerns and address them as follows:
- Novelty and Evidence for AR-CNN
Our study focuses on evaluating the contributions of the 118 GHz and 183 GHz microwave channels to temperature and humidity profile retrievals, leveraging the enhanced capabilities of the FY-3G/MWRI-RM. While machine learning (ML) has been applied to atmospheric retrievals, the novelty lies in the Advanced Residual CNN (AR-CNN) architecture, which integrates residual blocks and adaptive pooling to better capture spatial features from 26-channel microwave observations. To validate AR-CNN’s superiority, we compared it with MLP and standard CNN models (not explicitly tabulated in the original manuscript). The AR-CNN achieved a lower RMSE (1.24 K for temperature, 12.98% for humidity) compared to MLP (1.52 K, 15.3%) and CNN (1.38 K, 14.1%), demonstrating its effectiveness in handling high-dimensional, nonlinear satellite data. - Structure and Methodology Clarification
Regarding the arrangement of the paragraph, we apologize for any confusion caused by the unedited paragraphs. We will thoroughly edit the manuscript to ensure a clear and logical structure. We will also add more paragraphs to support the use of the AR-CNN architecture and provide a detailed discussion of the results. - Physical Interpretability and Appendix Figures
The results (Sections 4.1–4.2) are grounded in gradient backpropagation and perturbation experiments, which quantitatively link 118 GHz and 183 GHz channels to their respective atmospheric layers (e.g., 118 GHz for upper-level retrievals, 183 GHz for mid-to-lower layers). These align with Jacobian analyses (Figs. 5, 9) and confirm the model’s interpretability. The appendix figures (e.g., global maps in Fig. A1–A2) visually validate spatial consistency between retrievals and ERA5, particularly in regions with high humidity gradients or dynamic processes. Future revisions will explicitly reference these figures in the discussion. - Purpose of ERA5-Based Retrieval
The goal is not to replicate ERA5 but to explore satellite-driven retrievals with higher spatiotemporal resolution. ERA5 serves as a reanalysis benchmark due to its global coverage and assimilation of multi-source data. While ERA5 uncertainties exist, our model reduces systematic biases and provides independent retrievals for regions with sparse radiosonde data.
Citation: https://doi.org/10.5194/egusphere-2025-680-AC1 - Novelty and Evidence for AR-CNN
-
RC2: 'Comment on egusphere-2025-680', Anonymous Referee #2, 29 Apr 2025
This paper presents a algorithm for use on the microwave imager instrument aboard FY-3G to retrieve profiles of temperature and humidity. The algorithm is a convolutional neural network (CNN) that was trained with ERA5 data, and because it is in the microwave part of the spectrum it can retrieve all sky conditions. Neural network retrievals are actively being improved in the literature, especially multidimensional algorithms such as convolutional neural networks, so this paper in its ideas is suitable for publication in AMT.
Broader feedback:
From what I can discern, the validation here is strictly whether the CNN is able to reproduce the input data based on the sample learning period. While it is good to utilize this as a check of the algorithm behavior, there should be a validation performed with observations from alternative sensors/algorithms or ground-based observations (including ground-launched radiosondes).
Why was relative humidity (RH) chosen over water vapor mixing ratio or specific humidity? While I am not arguing that RH is invalid to retrieve, RH is not an absolute quantity and is dependent on temperature (which entangles retrieval of waver vapor and temperature). RH makes it more difficult to compare results against other literature in the microwave remote sensing community. Retrieving a water vapor value specifically would make the two values more distinct and provide better illumination of what channels are sensitive to which physical values.
The indication that 183 GHz provides higher information content for temperature sounding than all but one 118 GHz channel is a surprise to me (this statement is based on my interpretation of Figure 5 along with corresponding text). One citation is given for the reasoning behind this but is unfortunately in another language and inaccessible to me. Also, L239-242 states water vapor attenuates the signal (and cites a paper describing an infrared retrieval paper), which seems contrary to the stated conclusion. Detailed information about this analysis methodology and better tying the conclusions to relevant principles should be provided.
The methodology explanation does not line up with explanations in the results sections. Up until section 4.2, I thought the retrieval was only over ocean (viz. L137) given that the training database was constructed from ocean only profiles and the validation dataset is part of the training database. The appendix plots also show only over ocean data. Yet Section 4.2 describes land cover characteristics when discussing the humidity retrieval performance. More detailed information about the database construction (perhaps a spatial map of where the database is compiled from) should be provided and the text clarified to line up with the methods.
Figures 3 (c-h) and 7 (c-h) are not very illustrative because of the density of points. Rather than shading by absolute error, a two-dimensional histogram should be performed and shading indicate the counts in the bins. Additionally, the line plotted on each subplot has no label -- whether it is a least squares fit or a 1-to-1 slope, that should be indicated in the figures. And speaking of least squares fit, are any fitting statistics available for the scatter plots?
In surveying the literature, much of the historical work mentioned was on infrared retrievals. More citations on microwave regime development should be included beyond why the regime is important for all-sky retrievals.
Below is more specific copyediting feedback (which can be rejected where deemed), but overall the wording should be refined to focus on the scientific conclusions drawn from the results at hand without extra pomp or overbroad explanations.
The introduction section was difficult to parse. It is a three page-long paragraph. Consider splitting it up into thematic paragraphs. (This may be simply because of the template used.)
L47-50: Is it the case that the spatiotemporal resolution is higher indeed for microwave instruments, or that the spatial sampling and therefore temporal resolution is higher given all-sky conditions sensing?
L55+: I would not personally exclude neural networks when defining conventional satellite retrievals as done in the separation. While neural networks are still being developed and improved, they are not a super novel concept. In example, L85 cites a neural network retrieval from 15 years ago.
L79-81: This sentence has too much puffery.
Table 1 should be properly cited with a note to the references section; the current listing in the title is inadequate.
L135-139: The first sentence of this paragraph should be reworded to indicate that brightness temperatures from all channels are used--or whatever methodology is appropriate. The static values should be further described.
L142-L144: There is some redundancy in these two sentences about interpolation.
Figure 2: Font size should be increased on within-figure text.
L185-186: This wording is awkward.
L187-188: I would argue this is more a matter of instrument sensitivity vertically than model performance.
L198-200: I would rephrase or better convey the argument here.
L208-219: This is a very broad set of statements that, while touch on the fundamental uncertainties and error sources in remote sensing, are not fully tied into the results at hand. Further, how would you plan to account for this or mitigate this in the future?
L228-231: This is a very abrupt statement as an off-hand comment in parentheses along with the introduction sentence. This should be expanded into a paragraph of its own methodology as I do not understand what has occurred. What is model.conv1? What spatial dimensions?
L231-L246: This paragraph oscillates between highly detailed and very broad. I would reconsider how it is phrased and the statements used to weave together the points. Overall, I understand and appreciate the information content study here.
Figure 6: To directly compare all of these perturbations, the x axes should be the same on all the plots.
L272-282: Similar to L208-219, this is another very broad set of statements that do not fully tie into the results or indicate how/if you plan to address these concerns in the future.
L286-288: These sentences are partially redundant.
L290-295: Repeated phrasing. Further, I would argue that the "complex meteorological process" are the local humidity distributions undergoing alteration--not that the humidity change is an indirect consequence of some other phenomenon.
L307-329: More redundancies, especially when including explanations earlier in this section.
Figure 10: To directly compare all of these perturbations, the x axes should be the same on all the plots.
Citation: https://doi.org/10.5194/egusphere-2025-680-RC2 -
AC2: 'Reply on RC2', Wei Han, 23 May 2025
Overview
This paper presents a algorithm for use on the microwave imager instrument aboard FY-3G to retrieve profiles of temperature and humidity. The algorithm is a convolutional neural network (CNN) that was trained with ERA5 data, and because it is in the microwave part of the spectrum it can retrieve all sky conditions. Neural network retrievals are actively being improved in the literature, especially multidimensional algorithms such as convolutional neural networks, so this paper in its ideas is suitable for publication in AMT.
From what I can discern, the validation here is strictly whether the CNN is able to reproduce the input data based on the sample learning period. While it is good to utilize this as a check of the algorithm behavior, there should be a validation performed with observations from alternative sensors/algorithms or ground-based observations (including ground-launched radiosondes).
Response:
Thank you for your insightful feedback on our validation methodology. We acknowledge the need to strengthen transparency and have clarified our current work’s scope. Specifically, quantitative metrics (e.g., RMSE) for different model architectures are presented in Appendix A, Table A1 of the revised manuscript. These serve as preliminary model-selection indicators during development, without in-depth interpretation, as the study prioritizes demonstrating the framework’s feasibility over exhaustive benchmarking. Additionally, the scarcity of in situ marine observations and spatiotemporal mismatches with FY3G retrievals currently limit robust maritime validation. In follow-up research, we will systematically address these via spatio-temporal alignment algorithms and high-resolution satellite datasets to ensure comprehensive cross-environment evaluation.
Why was relative humidity (RH) chosen over water vapor mixing ratio or specific humidity? While I am not arguing that RH is invalid to retrieve, RH is not an absolute quantity and is dependent on temperature (which entangles retrieval of waver vapor and temperature). RH makes it more difficult to compare results against other literature in the microwave remote sensing community. Retrieving a water vapor value specifically would make the two values more distinct and provide better illumination of what channels are sensitive to which physical values.
Response:
Thank you for your comment on RH selection. This choice is grounded in two key considerations. First, as a standard ERA5 output, RH offers direct usability as label data and aligns with meteorological operational practices, ensuring compatibility with public observations and model inputs. Second, this study represents an initial effort to develop a temperature-humidity joint inversion model using FY-3G/MWRI-RM full-channel data. While RH’s temperature dependence introduces complexity, it enables exploration of dynamic temperature-humidity interactions (e.g., synergies between 118 GHz and 183 GHz channels; see Sections 4.2 and Conclusions in the revised manuscript). We agree that specific humidity retrieval would enhance physical clarity and plan to expand the model’s output to include water vapor parameters in future work, alongside error propagation analysis to decouple temperature-humidity contributions.
The indication that 183 GHz provides higher information content for temperature sounding than all but one 118 GHz channel is a surprise to me (this statement is based on my interpretation of Figure 5 along with corresponding text). One citation is given for the reasoning behind this but is unfortunately in another language and inaccessible to me. Also, L239-242 states water vapor attenuates the signal (and cites a paper describing an infrared retrieval paper), which seems contrary to the stated conclusion. Detailed information about this analysis methodology and better tying the conclusions to relevant principles should be provided.
Response:
To deepen the analysis of 183 GHz channel sensitivity, we have enhanced both physical explanations and literature integration. Through gradient backpropagation and perturbation experiments (Sections 4.1–4.2), we clarify that the 183 GHz water vapor channel indirectly constrains temperature via vapor-radiation feedback: in high-humidity regions (e.g., boundary layers), water vapor’s absorption/emission modifies radiation transfer, influencing temperature through latent heat release. In contrast, the 118 GHz oxygen channel directly senses mid-upper tropospheric temperature gradients, with humidity impacts mediated by dynamics like vertical advection and atmospheric stability. Furthermore, we replaced some Chinese references with English counterparts and added discussions on temperature-humidity coupling (Lines 209–215 in Section 4.1; Lines 279–284 in Section 4.2 of the revised manuscript). The “model weight diffusion in convolutional layers” describes potential propagation of 118 GHz temperature-sensitive features to humidity parameters via spatial feature mixing, a phenomenon warranting further study with physics-constrained neural networks.
The methodology explanation does not line up with explanations in the results sections. Up until section 4.2, I thought the retrieval was only over ocean (viz. L137) given that the training database was constructed from ocean only profiles and the validation dataset is part of the training database. The appendix plots also show only over ocean data. Yet Section 4.2 describes land cover characteristics when discussing the humidity retrieval performance. More detailed information about the database construction (perhaps a spatial map of where the database is compiled from) should be provided and the text clarified to line up with the methods.
Response:
Thank you for identifying the inconsistency in our data description. The term “land cover characteristics” was a typo; both training and validation datasets exclusively use oceanic observations (detailed in Section 2.2). We have corrected relevant paragraphs and removed land-data discussions, explicitly stated in the methodology: “To reduce preprocessing complexity, this study only selects oceanic data.”, highlighted marine data coverage in Appendix Figs. A1–A2. Future research will extend to land scenarios via standardized surface parameter preprocessing (e.g., soil moisture, vegetation index).
Figures 3 (c-h) and 7 (c-h) are not very illustrative because of the density of points. Rather than shading by absolute error, a two-dimensional histogram should be performed and shading indicate the counts in the bins. Additionally, the line plotted on each subplot has no label -- whether it is a least squares fit or a 1-to-1 slope, that should be indicated in the figures. And speaking of least squares fit, are any fitting statistics available for the scatter plots?
Response:
Thank you for your thoughtful suggestions, which have been instrumental in enhancing the clarity and informativeness of our figures.
To address the density of points in Figures 3(c–h) and 7(c–h), we have replaced the absolute error shading with two-dimensional histograms, where color intensity now represents the count of data points in each bin. This modification improves visual interpretability by highlighting data distribution patterns while retaining error magnitude insights. Additionally, the black dashed line in each subplot is explicitly labeled as the “1:1 consistency line” to clarify that it denotes perfect agreement between retrieved and target values, rather than a least squares fit. To further quantify the scatter plot relationships, we have included three statistical metrics—Root Mean Square Error (RMSE), Bias, and Pearson correlation coefficient (r)—in the figure captions and corresponding discussions. These metrics systematically evaluate error magnitude, systematic deviation, and linear correlation, respectively, providing robust statistical support for our validation results.
For other figures, we have implemented the following improvements to ensure consistency and readability:
Figure 2: Font sizes were adjusted to enhance legibility across all text elements, ensuring key labels and annotations are easily distinguishable.
Figures 5 and 9: To emphasize critical channels, Channel 22 (118.75±1.2 GHz) and 26 (183.31±7 GHz) are now plotted with solid lines, while other channels use dash-dot lines for clear visual differentiation. This distinction highlights their unique roles in temperature-humidity retrieval as discussed in Section 4.2.
Figures 6 and 10: The layouts were revised to incorporate pressure-level distributions of percentage changes induced by perturbations. This addition clarifies how the model responds to input variations across different atmospheric layers, strengthening the mechanistic insights presented.
These adjustments collectively aim to align our figures with scientific visualization best practices, ensuring they effectively communicate the study’s key findings while addressing the nuances raised in your feedback. We appreciate your guidance in refining our graphical representations.
In surveying the literature, much of the historical work mentioned was on infrared retrievals. More citations on microwave regime development should be included beyond why the regime is important for all-sky retrievals.
Response:
In response to your suggestion, we revised the introduction to enhance contextual depth:
Lines 55-68 in the revised manuscript: Expanded discussions on microwave remote sensing applications for atmospheric variable retrieval, integrating recent advancements and key studies to underscore the research’s scientific significance.
This revision strengthens the rationale for our approach by connecting it to broader disciplinary progress.Below is more specific copyediting feedback (which can be rejected where deemed), but overall the wording should be refined to focus on the scientific conclusions drawn from the results at hand without extra pomp or overbroad explanations.
Response:
We sincerely appreciate your meticulous copyediting suggestions, which have been invaluable in refining the manuscript’s clarity and scientific rigor. Below is a summary of the key revisions made to address your feedback, with a focus on enhancing readability, streamlining redundancy, and strengthening logical flow. All line numbers cited below correspond to the revised manuscript:
The original three-page introductory paragraph has been restructured into thematic subsections to improve navigability. We also expanded the literature review on microwave remote sensing (e.g., added citations in Lines 55–68) to better contextualize our approach, while adjusting the description of neural networks to acknowledge their established role in the field (avoiding the implication of excluding them from “conventional” methods).
Lines180-187 and Lines254-263: We recognize the need to further integrate broad theoretical discussions with our specific findings, particularly regarding error sources and future mitigation strategies. These aspects will be prioritized in follow-up analyses, including planned investigations into physics-constrained neural network architectures to address model uncertainties.
Broad statements about remote sensing uncertainties were either tied to our specific results or condensed, with clearer links to future mitigation strategies.
Lines199–201: The parenthetical comment about "model.conv1" was expanded into a dedicated paragraph explaining the convolutional layer architecture and spatial dimension considerations, improving methodological transparency.
-
AC2: 'Reply on RC2', Wei Han, 23 May 2025
-
EC1: 'Comment on egusphere-2025-680', S. Joseph Munchak, 10 Mar 2026
Following the decision for minor revisions with editor review, here are my specific comments to address prior to publication:
1. Lines 229 and 328 - the 118.75 +- 1.2 GHz channel is described as being sensitive to temperature gradients. This is not specific - are these horizontal or vertical gradients? For vertical gradients (lapse rate), the channel differences (compared to a channel closer or farther from the 118.75 O2 line) would actually be directly indicative of the gradient. Horizontal gradients would be captured by gradients in the 5x5 patch that is used as input - but it is not clear how this is relevant to the Jacobians. I think a better phrase would be "temperature distributions" instead of "temperature gradients" in these sentences.2. Table A1 - the units of the AR-CNN results are incorrect (and inconsistent with the other rows, which are correct).
Citation: https://doi.org/10.5194/egusphere-2025-680-EC1 -
AC3: 'Reply on EC1', Wei Han, 12 Mar 2026
Dear Editor,
We would like to express our sincere gratitude for your detailed and insightful comments. Your feedback, particularly regarding the precise physical interpretation of the microwave channels, has been immensely helpful in improving the rigor and clarity of our manuscript. We have carefully studied your suggestions and made the corresponding revisions.
Below is a point-by-point response to your comments.
Editor Comment 1: Lines 229 and 328 - the 118.75 +- 1.2 GHz channel is described as being sensitive to temperature gradients. This is not specific - are these horizontal or vertical gradients? ... I think a better phrase would be “temperature distributions” instead of “temperature gradients” in these sentences.
Response: We completely agree with your assessment. You are absolutely correct that describing a single channel’s Jacobian as sensitive to “temperature gradients” is physically inaccurate in this context. As you accurately pointed out, the Jacobians reflect the sensitivity to the state variable itself at specific pressure levels, rather than its spatial derivatives (which would require channel differencing for vertical gradients or spatial feature analysis for horizontal gradients). We highly appreciate you pointing out this conceptual ambiguity.
Editor Comment 2: Table A1 - the units of the AR-CNN results are incorrect (and inconsistent with the other rows, which are correct).
Response: Thank you for catching this oversight. We apologize for the typo in the formatting of the AR-CNN results. We have thoroughly checked Table A1 and corrected the units for the AR-CNN row to ensure complete consistency with the rest of the table.
Furthermore, inspired by your focus on physical rigor, we conducted a thorough review of the manuscript and made two additional proactive refinements to ensure our interpretations of the deep learning model’s behavior are fundamentally sound:
- We revised a sentence regarding the relationship between the 118 GHz channel and humidity. We replaced the original explanation involving “dynamical coupling”and “vertical advection” with a more accurate description of the “thermodynamic correlation.”
- We changed the term “radiometric observations”to “radiance observations” when discussing deep convection areas, as “radiance” is a more precise term for the physical quantity interacting with atmospheric hydrometeors.
Modifications in the Manuscript:
- Line 229 (Revised):“This highlights the model’s utilization of the oxygen absorption line near 118 GHz, a spectral feature with significant sensitivity to mid-to-upper tropospheric temperature distributions (500-200 hPa) and...”
- Line 328 (Revised):“In contrast, the 118.75±1.2 GHz channel is primarily sensitive to mid-to-upper tropospheric temperature distributions (500-200 hPa) via...”
- Table A1:The units for the AR-CNN row have been corrected to perfectly match the baseline formats.
- Additional Revision 1(Line 332): “This apparent discrepancy can be attributed to the thermodynamic correlation between temperature and humidity profiles." (The unsupported second half of the sentence regarding vertical advection was removed).
- Additional Revision 2(Line 287): “These processes, especially in areas with deep convection or liquid water droplets and ice crystals, obscure the direct link between radiance observations and humidity, challenging the model's ability to accurately retrieve humidity values solely from TBs.”
Citation: https://doi.org/10.5194/egusphere-2025-680-AC3
-
AC3: 'Reply on EC1', Wei Han, 12 Mar 2026
Peer review completion
Journal article(s) based on this preprint
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,021 | 460 | 44 | 1,525 | 40 | 53 |
- HTML: 1,021
- PDF: 460
- XML: 44
- Total: 1,525
- BibTeX: 40
- EndNote: 53
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Minghua Liu
Yunfan Yang
Haofei Sun
Ruoying Yin
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(8585 KB) - Metadata XML
The paper used machine learning and microwave channels to retrieve temperature and humidity profiles. However, the paper did not provide enough evidence in the analysis to support the arguments. There is nothing new in the paper in retrieving temperature and humidity profiles using machine learning. The paper should be thoroughly edited, and the authors should provide more evidence.