the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 18 Aug 2026
Introducing a Discrete Multi-beam CubeSat Radar for Global Precipitation Observations: Conceptual Demonstration and AI-powered 3D Reflectivity Reconstruction
Abstract. Current spaceborne precipitation radars, including the Global Precipitation Measurement Dual-frequency Precipitation Radar and the FengYun-3G Precipitation Measurement Radar (FY-3G PMR), provide unique three-dimensional (3D) observations of global precipitation systems. However, limited spatiotemporal sampling and relatively long revisit intervals constrain the monitoring of rapidly evolving storms. This study proposes a discrete multi-beam precipitation radar concept for small-satellite constellations and develops a radar-constrained active–passive inpainting framework for reconstructing unsampled inter-beam observations. Collocated FY-3G PMR reflectivity and Microwave Radiation Imager for the Rainfall Mission (MWRI-RM) brightness temperatures observed from the same platform are used to emulate sparse radar sampling. The proposed model integrates sparse radar reflectivity and 26 passive microwave channels within a squeeze-and-excitation U-Net. Independent experiments are conducted for the PMR Ku- and Ka-band observations under different beam-spacing configurations. Reconstruction performance is evaluated exclusively within the masked inter-beam regions, and a boundary-smoothness constraint promotes continuity between reconstructed and observed beams. Channel-ablation experiments are further used to diagnose model sensitivity to individual passive microwave channels. The results demonstrate that sparse radar profiles and passive microwave observations can jointly recover coherent 3D reflectivity structures over a range of sampling densities. Across both radar frequencies, the overall mean absolute error and error standard deviation remained below approximately 2.0 and 3.0 dB, respectively, across configurations in which one to five beams were skipped between adjacent retained beams. The dependence of reconstruction accuracy on beam spacing provides quantitative guidance for balancing beam number, cross-track sampling extent, and constellation cost, while the channel-sensitivity results inform the design of future active–passive payloads. The proposed framework also provides a transferable basis for retrieval and reconstruction algorithms for future sparse-sampling radar missions.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Atmospheric Measurement Techniques.
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.- Preprint
(24058 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-4804', Kamil Mroz, 07 Oct 2026
-
RC2: 'Comment on egusphere-2026-4804', Anonymous Referee #2, 16 Oct 2026
Review of: “Introducing a Discrete Multi-Beam CubeSat Radar for Global Precipitation Observations: Conceptual Demonstration and AI-Powered 3D Reflectivity Reconstruction” - Liu et al.
Prepared for: Atmospheric Measurement Techniques
Summary: This work introduces and evaluates a deep learning method for infilling synthetic radar observations from a hypothetical spaceborne instrument with multiple fixed beams. It considers an exciting alternative spaceborne observation paradigm, where less capable but much cheaper instruments are supplemented by AI to achieve results similar to more expensive conventional observing systems.
Recommendation: Accept pending revisions. Overall, the study is well motivated and an interesting contribution. The model appears to work (though it is not exactly clear how well), and designing future instruments that rely on AI-based retrievals to augment their capabilities is an exciting new area of research. There are several major issues, below, that need to be addressed before I would recommend publication.
Major Comments:
- This study lacks a clear experimental baseline or control. The study primarily focuses on the development and evaluation of a deep learning algorithm, but the evaluation only reports the model error in absolute terms. How can we assess whether an MAE of 2dbz represents acceptable or good performance? What is a level of error that is acceptable before this solution could be considered a legitimate alternative to a scanning radar? Typically, I would at least expect a comparison to a baseline model for infilling. In the conclusions section it is mentioned that the MAE is on the scale of ground-based radar calibration uncertainty, but that does not seem like a reasonable or sufficient comparison. Furthermore, some of the individual errors shown in Figures 10 and 11 far exceed this level, which is relevant to (2), below.
- MAE (and RMSE) alone are insufficient to determine whether this is an acceptable alternative to a scanning radar. These metrics average performance over pixels that are trivial to predict and those that are difficult. Figures 10-11 show individual errors on the order of 10-20dbz which could be critical when using a system like this operationally, since this could be the difference between a downpour and moderate drizzle for example. Some of the most important observations that would need to be predicted accurately by this algorithm are minorities (e.g. location and strength of precipitation in convective cores). A metric that takes into account spatial structure (e.g. structural similarity or spatial power spectral density) or focuses on severe weather and/or localizing storm structures would elevate this study.
- Please be specific about testing and validation splits (what orbits, how were the 20% of validation samples selected). Also, was the validation set used only to monitor for overfitting or was it used for model design? Even if no model tuning was done and design choices were made solely based on past literature, that should be discussed.
- Code and data are not shared in accordance with AMTs code and data policy: https://www.atmospheric-measurement-techniques.net/policies/data_policy.html. Please provide a proper citation and DOI for the radar dataset used and share code via a repository with a DOI rather than by direct request to the authors.
Minor Comments:
- While I think it is a sensible thing to try, the use of squeeze and excitation blocks is a significant deviation from the conventional UNet architecture and is not experimentally motivated. Arguments are made that they allow feature calibration and help process heterogeneous input channels, but their use is not quantitatively supported.
- L111-116: You should add to the description of FY-3G PMR that it is phased array scanning.
- L140: note the 1 in the superscript in Equation 1 is referring to the channel dimension (I think)
- L142: change “complete” to “incomplete”?
- Eq3: Note in the description of Eq. 3 that $f_{\theta}^{(b,n)}$ represents the ML model
- Line 155: Why is precipitation required for all beams, wont it be very common to see precipitation in only a subset?
- Table 2: Consider adding the spacing between beams in physical units to give a sense of scale. Maybe just the distance between the center beam and the closest off-center beam at the ground.
- L214: B4 is not really a bottleneck because of the UNet skip connections.
- L214: Batchnorm is often detrimental or of little use for regression type problems. Was this choice motivated experimentally?
- 5: “SE” was never defined
- Eq 11: “V” was never defined
- Eq 12: Isn’t \sum G_n always non-zero? Are cases with no valid radar observations included in the dataset?
- Eq 17: Again, there is an unusual design choice that is not experimentally motivated: the L_smooth loss term. Is this because a specific problem was observed with the boundary pixels? Did you experimentally verify that this term improves results?
- Line 299: I believe there also tends to be more precipitation at those levels.
- Section 4.1.2: I find it odd that the authors did single-channel dropouts when assessing their importance despite being aware of the limited usefulness of this experiment due to correlations between channels. It also seems that, given this study is about exploring how deep learning could allow launching a less expensive and complicated instrument, a more useful experiment would be dropping out large groups of microwave observations or dropping out the microwave observations altogether.
- Figures 5 and 7: you might report the value for Zm separately or use a broken axis because it is limiting my ability to see differences between the microwave channels. Also, this axis is not periodic, so a bar chart is more appropriate.
- Table 3: the subscript “sparse” is written in math mode.
- 428: typical calibration uncertainty for a ground-based radar might provide some context, but is not an appropriate benchmark for the ML model skill for multiple reasons: it is a different source of error, the input to the ML model already contains this error, ground-based radars are different instruments, the distributions of absolute model error and calibration uncertainty are different…
Citation: https://doi.org/10.5194/egusphere-2026-4804-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 163 | 78 | 56 | 297 | 57 | 56 |
- HTML: 163
- PDF: 78
- XML: 56
- Total: 297
- BibTeX: 57
- EndNote: 56
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The paper proposes a discrete multi-beam precipitation radar concept for a small-satellite constellation, where a limited number of fixed radar beams provide direct 3-D precipitation observations and the gaps are reconstructed using AI and passive microwave observations. The authors use real FY-3G PMR Ku- and Ka-band observations together with 26 MWRI-RM channels to emulate sparse radar sampling. The study therefore provides a useful proof of concept for trading radar sampling density against AI-based reconstruction, with potential implications for reducing payload and/or constellation requirements.
The main strengths are:
The scientific concept is clear and interesting.
The use of real collocated active/passive observations makes the demonstration convincing.
The experiments provide a clear test of how reconstruction quality changes with increasing beam separation.
Evaluation is performed only in the reconstructed region, avoiding statistics being dominated by observed pixels.
The authors clearly acknowledge several important limitations of the current experiment.
Main comments
1. Scope of the demonstration
The experiment does not yet fully demonstrate the proposed instrument concept. It mainly demonstrates AI reconstruction from artificially sparse sampling of an existing radar swath. Real instrument effects such as antenna patterns, incidence angles, beam filling, clutter and sensitivity are not represented. I suggest making this distinction clearer throughout the manuscript and possibly moderating the title.
2. How much does AI improve the reconstruction?
The reported ~2–3 dB errors are impressive, but the experiment is particularly favourable because the missing beams are surrounded by observations from the same radar swath. The key question is therefore how much better the neural network performs than simpler approaches.
I strongly recommend comparison with at least:
This would demonstrate the actual added value of the proposed AI approach and the passive microwave observations.
3. Channel-importance analysis
The channel-ablation analysis is useful, but its interpretation should be more cautious. Setting one channel to its mean value during inference creates an input that is not physical and is outside the training distribution. Because the microwave channels are also strongly correlated, this is not necessarily a measure of the information content of an individual channel.
A more physically meaningful experiment would be to remove and retrain the network without complete frequency bands, e.g. without the 19 GHz channels. The authors should also clarify that the current experiment provides a relative sensitivity of the trained network to channel removal, rather than an absolute measure of channel information content.
The results could also be discussed in terms of the different spatial scales and hydrometeor sensitivities of the microwave frequencies. For example, higher frequencies (>~89 GHz) provide information on small ice particles that may be below radar sensitivity, while lower-frequency observations have much larger footprints and provide mainly coarse-scale information.
4. Representativeness of the test dataset
The independent test set consists of only two orbits and may not be sufficient to demonstrate performance across a representative range of precipitation systems.
More importantly, requiring precipitation to be present across all 25 central beams introduces a strong sampling bias towards large, horizontally extensive systems. These are also likely to be the easiest cases for interpolation. The performance for convective cells, small-scale precipitation structures and strong gradients is much more relevant to the proposed concept and should be investigated or discussed more explicitly.
5. Potential NWP application
The potential relevance to numerical weather prediction would be worth discussing. Data assimilation systems generally need to account for spatial correlations between observations, and observations are therefore often thinned before assimilation.
A low-cost multi-beam radar with sufficient separation between beams could potentially provide a useful compromise between spatial coverage and observation independence. A constellation could then provide improved temporal and spatial sampling while retaining a relatively sparse set of observations suitable for assimilation. This could provide an additional motivation for the proposed mission concept.
Minor comments
1. Input representation
Please explain more clearly the rationale for the chosen U-Net input representation and how the 3-D radar field and 2-D TB fields are mapped into the input dimensions. Figure 3a illustrates this particularly well and could be explicitly referenced here.
2. Measurement uncertainty
The reference radar measurements themselves contain random measurement uncertainty, of order ~1 dB. Therefore, the reconstruction error cannot realistically be expected to fall substantially below this level. Assuming independent errors,
MSEtotal≅MSEreconstruction + MSEmeasurement
The authors should consider evaluating the reconstruction above the sensitivity threshold, e.g. (Z > Zsens+3) dB, where measurement uncertainty has a smaller impact. The increase in RMSE at high altitudes/low reflectivities may otherwise be partly related to measurement uncertainty.
3. Visualisation
It would be more informative to show the original field, reconstructed field, and their difference. The difference field would help identify errors around strong gradients, convective cores, precipitation boundaries and the melting layer, complementing the global statistics.