Comments on the manuscript “Reconstruction of 3D precipitation measurements from FY-3G MWRI-RM imaging and sounding channels” by Yang et al.

Overview

The authors present a neural network approach to reconstruct radar reflectivity profiles from passive microwave observations. The network is trained using a month (October 23 to November 31, 2023) of spatially and temporally collocated radar and microwave radiometer data from the FengYun 3G (FY-3G) satellite, i.e., the 35 GHz Precipitation Measurement Radar (PMR) and the 10.65 to 190.31 GHz Micro-Wave Radiation Imager for the Rainfall Mission (MWRI-RM). The authors estimated the effects of temperature sounding channels at 50-53 and 118 GHz and polarization difference on the prediction by comparing three networks trained with different input channel combinations. The experiments show that the 50-53 and 118 GHz channels improve the reconstruction of radar reflectivity profiles, especially over land. The observation-based link between active and passive microwaves on global scales provides new opportunities to extract information from passive microwave observations as they do not rely on forward model assumptions and could be beneficial for global assimilation of passive microwaves.

General comments

The work is well structured, and the case studies prove the model’s ability to reconstruct radar reflectivity profiles. Below are a few general comments on the neural network training and evaluation that require additional work but need to be addressed in a revised version of the manuscript.

There are various problems during the data preparation, model architecture, and training procedure that likely cause overfitting and limited generalization capabilities of all three trained models. Such a lack of generalization would make it difficult to relate their prediction skill to the information content of the model input. Despite the results being relatively close to the expectation, I would highly recommend adjusting several methodological aspects.

The authors perform a random split into training and test data sets with a ratio of 80/20. However, the input data is autocorrelated in space and time. In the current setup, a test scene can be as close as a few ten kilometers to the center of a training scene. This might lead to the model learning the reflectivity profile of single mesoscale precipitation systems rather than learning the generalized radiative TB feature’s relation to reflectivity profiles. I suggest that the authors split their data into, e.g., weekly chunks. Also, the oversampling of precipitation scenes before splitting the data into training and test sets might lead to the occurrence of the same precipitation scene in training and test sets. However, this should be avoided. A way to replicate but still challenge the model would be data augmentation by rotating or flipping the TB field around the central footprint. Additionally, an independent validation set is needed to compare the three models. This was done only during case studies but without any quantification. Currently, the scores are computed only on the test data, which is problematic not only for the reasons stated above but also as its loss was used to determine the optimal parameters during early stopping. Finally, the number of model parameters exceeds the number of samples by a factor of five, although it should not exceed a factor of 0.1. The first fully connected layer contains more than 5 million parameters that would be sufficient for the model to learn a range of reflectivity profiles encountered during training.

I would like to encourage the authors to publish their prepared samples used to train, test, and evaluate the models as well as the three fully trained models to make their results reproducible.

Specific comments

Figure 1: This figure contains many errors. I suggest removing it entirely, especially since TRMM and GPM CO are not part of the manuscript. The MWRI-RM channels should be presented in a table, ideally linked to the three experiments following later. Correct the MWRI-RM channel definitions using Table 4 in https://doi.org/10.34133/remotesensing.0097

Line 40: MWRI-RM observes only V-pol at 166 GHz, while GMI observes both polarizations at this frequency. How would the addition of polarization information at this frequency improve the reconstruction of radar reflectivity profiles, especially under higher scattering by snow?

Line 45: How many independent height levels of liquid water content can be retrieved from dual oxygen absorption sounding channels? This information would be helpful to understand the physical limitations of reconstructing radar profiles with >100 vertical degrees of freedom.

Line 52: Provide a table with the accuracy for each channel. How does the channel accuracy affect the reconstructed radar profiles? Would it be helpful to inform the model on channel accuracy?

Line 65: This sentence is unclear to me. Which direct relationships between active and passive microwave observations need to be identified? Explain why a direct relationship between TB and radar reflectivity is needed. In general, forward models can simulate both the active and passive signal. Is it due to assumptions in the forward model or due to differences in weather models and reality? This is a key question of this work that needs to be explained.

Line 73: It would be helpful to have a table of previous work and the respective methods. How does the method used in this work differ from the Res-UNet in Brüning et al. (2024), and how are the differences motivated? What is the accuracy of other work that replicates radar reflectivity profiles?

Line 76: What is meant by “limited to specific precipitation scenarios due to limited spectral channels of GMI”? Are only specific scenarios used to train the model in that work?

Line 85: It might be confusing to mention 3D structures when the output of your model is 1D. The network does not know about the 3D nature of radar reflectivity fields and is not forced to be spatially consistent in the radar reflectivity space. Only the consecutive application of the model along and across track dimensions leads to a 3D field. Would an improved model performance be expected if the output is a 3D field? Why was this approach not chosen for this work?

Line 86: Why are the non-precipitating scenes not split into land and ocean as well?

Line 106: Provide a table that lists the footprint dimensions of the different channels.

Figure 2b: Mention how the footprints are scaled in size and number instead of adding “not to scale.” What does “*Npoints” mean? Indicate the footprints used for training here and combine them with Fig. 3.

Line 115: Are the Level 1 products of PMR corrected for two-way attenuation by liquid water and water vapor? If not, how does this affect the reconstruction and comparison with PMR observations under varying incidence angles? How does multiple scattering affect the reconstruction in the presence of high-density ice particles typical for extreme precipitation events?

Line 116: Why are case studies using data from July 2023 if the data is released after October 23, 2023? Clarify this. I assume the model was trained on the October/November data. If that is true, do you expect the same performance for other months of the year? What is meant by “used here for the preliminary research”?

Line 126: Mention the PMR scan interval of 0.7° and provide the number of cross-track scans that match the 2° incidence angle criterion.

Line 130: A threshold of 8 km cuts off the reflectivity profiles during the case studies. Also, the additional 183.31+/-2 GHz channel of MWRI-RM compared to GMI peaks around this height, depending on humidity levels. By how much does the threshold of 8 km change the ratio of clear-sky and cloudy pixels compared to a 9 or 10 km threshold?

Line 133: How do 136 range bins from 1.1 to 8 km match with the range resolution of 250 m of PMR? Also, note that in Table 1, the output is 138. Why are both different? How was the noise, sidelobe clutter, and ground clutter filtered, and is the filter considered surface type?

Line 138: What does NaN mean in the context of the loss function? How is terrain treated, i.e., subsurface PMR pixels? It would be important information for the reconstruction to know the lowest boundary of the precipitation field. In Fig. 7, it seems that mountains extend up to 2 km, and the models add precipitation below the surface (see. Fig. 7i right above the letter A)

Line 143: Could it happen that two PMR columns are assigned to the same MWRI-RM footprint due to their different spatial resolution? Are those duplicates filtered out?

Figure 3: This should be added to Fig. 2 for better understanding. Indicate the size of a 15x15 patch in kilometers.

Line 152: How is a precipitation event defined? Provide information on the filter method or threshold.

Line 155: How many samples were available prior to the oversampling? Based on this description and the diagram in Fig. 4, it seems that oversampling was performed prior to the data splitting. Clarify this. See also the general comment regarding random data splitting.

Line 157: Describe the standardization method.

Line 157: Does logarithmic transformation mean Ze was transformed to dBZ? Clarify this.

Figure 4: Explain the meaning of “values adjustment” in the PMR branch.

Table 1: The number of parameters is very large and exceeds the number of samples. The first dense layer contains 28800 * 200 + 200 (5.76 M parameters). This is much more than the number of samples (about 1 M) and will likely lead to poor generalization of all three trained models. Pooling layers could help to reduce dimensionality before passing to dense layers. Also, note that the number of parameters is not correctly shown in the table because the bias terms are missing for Conv (32, 64, 128) and FC (200, 138). It might be more meaningful to show the activation shape, activation size, and total number of parameters for each layer. Mention the kernel size in the text. Also, see the comment on line 133 regarding the confusion about the vertical resolution of the Ze profile (136 or 138 bins).

Line 170: Use shorter names for the experiments (e.g., Ex14, Ex26, etc.).

Line 171: Mention the batch size, number of epochs / early stopping, learning rate, and other details on training (restore best weights, etc.). Add figures of the training and test loss for each epoch until training is stopped for each of the three models to the appendix.

Line 173: Clarify why the published code has a different loss than MSE (sum of squared errors: tf.reduce_sum(tf.square(y_true_filtered - y_pred_filtered))) in line 58 of training_cnn.py.

Line 174: It would be good to separate the description of experiments from the model architecture chapter. The paragraphs of each experiment repeat large parts of the introduction and do not belong to the methodology section.

Line 193: Explain how the model output advances weather forecasting techniques and links to hydrometeors simulated by weather models.

Line 215: See comment on line 152.

Line 225: See comment on line 193.

Line 231: It would be good to split non-precipitation in land and ocean as well; see comment on line 86. How does the model perform in coastal regions where parts of the passive channel are affected by land?

Table 2: Reduce the precision of metrics to those digits that are significant. The names of the experiments are wrong; all are named “baseline.” The F1 score is the same for land and ocean for all three models. Is that correct?

Line 234: In general, the melting layer is below 3 km in mid-latitudes, which are also covered by the satellite. Why does the RMSE show a peak at 4-5 km only and not at lower heights?

Line 236: How much of the uncertainty over land can be attributed to the lack of topography information?

Line 244: The scattering signal of precipitation at the 50-53 and 118 GHz channels is very small.

Line 254: How does the F1 score vary between land and ocean? See comment on Table 2.

Line 259: The month of July is not mentioned in the data section.

Line 260: Explain why these events are “challenging” for the model. Those are strong precipitation events with clear passive microwave signatures.

Line 280: The discussion following this line is very vague and subjective. I recommend quantifying the comparison between PMR and the reconstructed profiles and providing a difference and scatter plot between the observed profile and the predictions. The discussion on the melting layer could be supported by comparing it with reanalysis data.

Figure 5: Use the same x-axis among the columns to make them comparable, ideally starting at 0 for RMSE and STD. How was the yellow area in a and d calculated, considering that the melting layer varies with latitude? In panels g and I, the polarization model performs worse than the model without polarization close to the surface. Why?

Line 294: What is meant by “vertical resolution,” and how was it determined?

Line 295: Add a reference to the precipitation amount.

Line 296: Mark the regions mentioned in the text inside the map.

Line 303: Provide a quantitative comparison between observation and prediction.

Line 321: Add ground-based radar data to the data section.

Line 321: What does “consistent” mean? It is not obvious from Fig. 9 and Fig. 8b. Plot both data in one figure with the same geographic extent and color bar. Ideally, a simple scatter plot should be made to show that the model and observation agree.

Line 353: What is meant by representative training dataset? Why is the data used here not representative?

Line 354: Clarify if resolution means vertical or horizontal.

Line 363: Mention snow scattering as well.

Line 369: What exactly is the advantage of GANs and diffusion models compared to the method selected here? This needs more explanation, potentially linked with the introduction where previous work is presented.

Line 372: “fully replicate” sounds as if it is, in principle, possible to create a radar reflectivity profile from passive microwave observations equivalent to a radar observation with the right AI method. I suggest to rewrite this. It is impossible to retrieve unique information on each height bin from the limited information of passive microwave observations. However, it would be interesting to see which channels provide information on the height level of the reflectivity profile. Could this be seen when computing the Jacobian of the model with 35 input channels?

Technical corrections

Line 7: What does “VPR” stand for? I suggest using “reflectivity profiles.” and not “VPR.”

Line 15: The start of the sentence should be a capitalized letter.

Line 18: The acronym “AWR” is rarely used. Use “radar” instead.

Line 22: Mentioning nations is distracting. Remove them throughout the manuscript.

Line 25: See comment on line 22.

Line 28: Remove the word “wideband.”

Line 32: Rephrase the beginning of the sentence to more neutral wording like “To further advance global precipitation observations.”

Line 49: Remove the word “strong.”

Line 56: The year in the reference is missing.

Line 61: See comment on line 18.

Line 73: Does “precipitation measurement radar” refer to PMR or any radar?

Line 90: The “HIW” acronym is not needed.

Line 93: See comment on line 22.

Line 96: DPR is written instead of PMR.

Line 104: Replace “frequency points” with “channels.”

Line 112: Move this to the data availability section.

Line 128: Remove the word “detrimental.”

Line 165: Remove the word “advanced.”

Line 168: Remove the word “enriched.”

Line 249: Replace “precipitation reflectivity” with “radar reflectivity.”

Line 294: Replace the word “fuzziness.”

This study proposed a method to reconstruct Ku band reflectivity profiles of PMR based on observations of MWRI-RM. I have some questions about this paper.

• What is the major difference of reconstruction method between this paper and your previous study (Yang, Y., Han, W., Sun, H., Xie, H., and Gao, Z.: Reconstruction of 3D DPR Observations Using GMI Radiances, Geophysical Research Letters, 51, e2023GL106 846, https://doi.org/10.1029/2023GL106846, 2024.), Except for the data, it seems that the methods used in the two studies are similar.
• What is the spatial resolution of the reconstructed reflectivity profiles?
• What exactly are the channels used in each experiment? I think it would be better to show a list of channels used in each experiment.
• In “Full Channel Experiment”, all 26 channels of MWRI-RM were used in the training model, and in “Polarization Difference Enhanced Experiment”, 9 additional Tb polarization difference data were added in the training model, however, the channels used to calculate the polarization difference have already been used in the “Full Channel Experiment”, there is no additional information was added in the “Polarization Difference Enhanced Experiment” compared to “Full Channel Experiment” from the perspective of amount of information. It is strange that the accuracy of the reconstructed reflectivity profiles in the “Polarization Difference Enhanced Experiment” is better than that in the “Full Channel Experiment”.