the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Jul 2026
CBAM-U-Net-Based Retrieval of Radar Composite Reflectivity from FY-4A Satellite Observations over Complex Terrain in Sichuan, China
Abstract. To address radar coverage blind spots in complex terrain, this study proposes an end-to-end deep learning framework to retrieve Radar Composite Reflectivity (RCRF) from FY-4A satellite multi-channel observations. We introduce CBAM-UNet, embedding a lightweight Convolutional Block Attention Module into a U-Net backbone. This dual-dimensional mechanism adaptively filters critical infrared spectral bands and precisely localizes intense convective cores. Evaluated on a comprehensively matched satellite-radar dataset (14,023 samples) from Sichuan Province (May–November 2023), CBAM-U-Net significantly outperforms mainstream CNN and Transformer baselines in retrieval accuracy (RMSE = 6.8290 dBZ, R2 = 0.6277) and structural fidelity (SSIM = 0.7894). Crucially, within the challenging severe echo regime (45–70 dBZ), the model achieves optimal Probability of Detection (POD = 0.5296) and Critical Success Index (CSI = 0.4384). Furthermore, crosssensor evaluations using FY-4B data demonstrate its robust zero-shot generalization against observational domain shifts. This research highlights the efficacy of integrating satellite multispectral features with attention-augmented networks to compensate for radar blind spots, providing reliable support for severe convective weather monitoring.
- Preprint
(12765 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-2395', Clément Guilloteau, 31 Jul 2026
- This is far from the first article presenting the use of CNN for retrieving precipitation intensity (or equivalent radar reflectivity) from geostationary multispectral infrared images. The authors need to justify the novelty of their work and to highlight what distinguishes it from previous similar works. The evaluation part is rudimentary, it consists in a listing of standard performance metrics, repeated three times for three different periods. This contribution is stereotyped and incremental rather than novel.
- The abstract and introduction “promise” things that are not really delivered. It is claimed that the objective is to retrieve the "3D intra-cloud precipitation structure" but only the composite reflectivity is retrieved, the vertical variability is not resolved. The algorithm is supposedly intended for filling gaps in radar coverage but the manuscript does not go all the way to demonstrate its utilization as such, no case of "filled" radar images is showed.
- Several statements regarding the instruments capabilities or the cloud physics are inaccurate or are oversimplifications, or elude important aspects of the retrieval physics (see line-specific comments attached).
- Some methodological choices need to be further justified and their implication must be discussed. For example the choice of using a max-pooling operator when aggregating the radar composite reflectivity from 0.01 deg to 0.04 deg.
Citation: https://doi.org/10.5194/egusphere-2026-2395-RC1 -
AC1: 'Reply on RC1', Wen Kang, 26 Aug 2026
Dear Reviewer,
We sincerely appreciate the constructive comments and valuable suggestions from you and the reviewers, which have greatly helped us improve the quality of this manuscript.
We have carefully addressed all the raised comments point‑by‑point**. A detailed, item‑by‑item response to each reviewer’s question, together with the corresponding revisions made in the revised manuscript, is provided in the attached supplementary PDF file.
Thank you again for your time and efforts. We are looking forward to your further feedback.
-
RC2: 'additional Comment on egusphere-2026-2395', Clément Guilloteau, 11 Aug 2026
It is not clear how the dataset is partitioned into training, validation, and testing subsets. The only information provided is that “the dataset was partitioned into training, validation, and testing subsets following a standard 7:2:1 ratio.” This raises concerns about the independence of the three subsets. If samples in the different subsets are in close spatial and/or temporal proximity, or potentially originate from the same storm systems, the subsets cannot be considered statistically independent. The authors should clarify how the partitioning was performed and, in particular, whether measures were taken to prevent samples associated with the same or closely related storm systems from being distributed across different subsets. Otherwise, the reported model performance may be affected by information leakage and may not accurately reflect the model’s ability to generalize to independent cases.
Citation: https://doi.org/10.5194/egusphere-2026-2395-RC2 -
AC2: 'Reply on RC2', Wen Kang, 26 Aug 2026
Dear reviewer, I have already responded to the question you raised in General Comments 5 of the PDF file.
Citation: https://doi.org/10.5194/egusphere-2026-2395-AC2
-
AC2: 'Reply on RC2', Wen Kang, 26 Aug 2026
-
RC3: 'Comment on egusphere-2026-2395', Anonymous Referee #2, 14 Aug 2026
Overall:
The stated objective of this manuscript is well within the scope of AMT. However, there are a few major issues. Throughout the article, claims are made regarding the operational utility of this model filling in radar blind spots. However, none of the evaluations shown appear to evaluate the performance of this model in blind spots. Furthermore, The previous literature involving the estimation of ground-based radar reflectivity from passive imager observations is also not discussed. Much of the language in this paper is overly positive regarding the performance of the proposed model. Throughout the text, many claims are made about the performance of this model that do not reasonably follow from the experiments performed.General Comments:
1. The manuscript focuses on the incorporation of CBAM into a traditional U-Net architecture and compares with other off-the-shelf architectures. However, there is not much discussion of hyperparameter details of their chosen architecture and there is no discussion of the hyperparameters chosen for the competing models. If the architecture of these models is to remain the focus of this manuscript, these details (e.g. number of convolutional filters, block design, batch size, learning rate, optimizer, etc...) need to be listed and discussed. Ideally there would be some common point of comparison between these models controlling for sample throughput, number of trainable parameters, or compute usage. Currently, readers cannot conclude much from the given results since hyperparameters could vary greatly between each of these models.
2. The manuscript is not well-referenced and largely ignores previous literature in the field that explores the use of geostationary imager observations to emulate composite reflectivity from ground-based measurements. I suggest that this article includes a detailing of previous methods used, a discussion of their deficiencies, and how the proposed CBAM-U-Net differs from them.
3. Overall, the language in this paper is overly positive regarding the performance of the presented model. I recommend heavy revision of the text throughout and tempering much of the language used. For example... Line 577: “[...] the proposed model exhibited a superior capability in delineating sever convective cores [...]” seems too strong to me given the very similar performance to other U-Nets. Much of the text in this manuscript does not reflect the level of evidence provided in the experiments shown. There are several factual mistakes in the introduction and background.
4. One claimed contribution of this paper is “addressing the operational challenge of radar coverage blind spots over complex topography. However, none of the shown experiments specifically address this. Generalization is only determined by running the model on a different satellite sensor. No experiments evaluate the estimation of precipitation structures in radar blind spots.
Specific Comments:
Line 8: What is meant by “optimal” here? Is this with respect to the other models shown?
Line 30: “Although satellites primarily detect cloud-top thermal radiation and microphysical properties [...]” Please be more specific about which kinds of sensors are described by this statement. I would agree this is correct for passive IR imaging instruments. However, this is not true for satellite cloud lidars and radars. You might also clarify that you mean ground-based radars in the latter part of this sentence.Line 34: “aligns spatiotemporally [...]” This is not necessarily true for all IR channels (e.g., water vapor in the 5-7 µm range, CO2 at ~13-14 µm, and ozone at ~9.6 µm.
Line 38: Please describe these conventional models that are being contrasted here. What “physical retrieval techniques” exist to solve this task?
Line 50: I do not agree that this is a characteristic of U-Nets. A counterfactual to this would be any diffusion model that uses a U-Net architecture. My expectation would be that the loss function used would have a much stronger impact on the “smoothness” of the predictions.
Line 64: Please provide the citation for CBAM.
Line 70 and throughout: I don’t agree that what has been developed here qualifies as an “expert system.” Most definitions I see involve the emulating human experts.
Line 80: It is not obvious to me how “immense value for real-time disaster early warning operations” follows from “robust cross-sensor zero-shot generalization.” Please add some explanation to this. This is also not an especially large dataset. If it were extremely large, we would expect to see much better performance from the transformers than we do here.
Line 133: “To further augment the model’s feature representation, two BTD formulations (BTD1 and BTD2) are engineered.” A difference between two channels is something that a neural network can learn extremely easily in just a single fully-connected layer. Therefore, it’s not obvious to me that these differences would add any value provided that the training data set is large. A following sentence “This physically informed feature engineering approach provides a solid foundation for [...]” overstates the importance of adding these differences to such a large neural network, in my opinion.
Lines 140 – 163: The ground truth for the model appears to be a 2-D array of column-wise maximum reflectivity. This seems to be at odds with the stated goal of estimated “3D ‘intra-cloud precipitation structure’” in Line 123.
Lines 164 – 174: How are the training and testing set separated? There is not enough information here to confirm that there is no leakage between these sets.
Lines 164 – 174: How is parallax handled in the matching of the radar and satellite data? Elevated cloud-tops and precipitation structures could be mismatched if this area is at a high enough viewing angle from FY-4A.
Lines 175 – 200: I do not feel that this section is necessary to be included in the article unless it is not present in existing publicly available FY-4A/AGRI literature. While the calculation of latitude and longitude can be a somewhat involved process, as this text shows, this does not contribute much to the stated objective of the manuscript. I feel that a citation to a FY-4A user’s manual (or similar) would be sufficient.
Line 218: Wouldn’t the resolution of the sensor vary as a function of viewing angle? What is the approximate viewing angle at this location?
Line 220: Why wouldn’t an average or interpolation be sufficient? This choice of max-pooling needs to be better motivated and needs to be made sufficiently clear throughout the manuscript that this filter is applied to the ground-truth data and changes the interpretation of the resulting model.
Line 244: “It is important to [...]” Since Batch Normalization does not appear to be used in this work, the following two sentences can be removed, in my opinion.
Line 259: “[...] U-Net models frequently struggle to reconcile macroscopic global structures with localized peak details” This is not necessarily a characteristic of U-Nets in general. Diffusion models are often U-Nets and do not have this issue.
Line 255-275: This section needs to be better referenced. Which other attention mechanisms have been introduced? Why are the suboptimal compared to CBAM? What is the citation for CBAM?
Line 273: How are these datasets separately? Ideally neighboring times are not included in the training and testing sets. There should be a time buffer in between training and testing to ensure no leakage.
Line 274: What is meant by “enhancement effects?”
Figure 4: How are the odd-numbered dimensions handled by the 2x2 max-pooling operation? E.g., for a 301x214 image, how does the max-pool result in 150x107?
Line 316: “dynamically emphasizes the structurally significant regions of precipitation systems” How was it determined that the regions emphasized were structurally significant? How do we know that clear-sky regions are not also emphasized in the attention maps?
Line 372: “This observation [...]” There are many reasons why a U-Net could outperform a transformer here. One is the inductive bias in convolutional layers not present in transformers allowing them to perform better on smaller datasets. I see no reason why one would conclude that it is specifically the encoder-decoder design that creates this difference observed here.
References:
- Several of the listed references do not have DOIs listed, which would ease finding the cited literature.
- Duan et al. 2021 appears to be duplicated.
- Sun et al. 2021 appears to be duplicated
- The year may be incorrect for article by Kou, L et al.Citation: https://doi.org/10.5194/egusphere-2026-2395-RC3 -
AC3: 'Reply on RC3', Wen Kang, 26 Aug 2026
Dear Reviewer,
We sincerely appreciate the constructive comments and valuable suggestions from you and the reviewers, which have greatly helped us improve the quality of this manuscript.
We have carefully addressed all the raised comments point‑by‑point. A detailed, item‑by‑item response to each reviewer’s question, together with the corresponding revisions made in the revised manuscript, is provided in the attached supplementary PDF file.
Thank you again for your time and efforts. We are looking forward to your further feedback.
-
AC3: 'Reply on RC3', Wen Kang, 26 Aug 2026
-
AC4: 'Final response by authors', Wen Kang, 14 Sep 2026
We thank the editor and reviewers for their valuable comments and suggestions. We have carefully addressed all raised points in the revised manuscript. The revised manuscript, tracked-changes version, and point-by-point response file have been uploaded. We believe the manuscript has been substantially improved, and we now submit our final response for further editorial consideration.
Citation: https://doi.org/10.5194/egusphere-2026-2395-AC4
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 172 | 60 | 22 | 254 | 21 | 20 |
- HTML: 172
- PDF: 60
- XML: 22
- Total: 254
- BibTeX: 21
- EndNote: 20
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
