Espresso: A Global Deep Learning Model to Estimate Precipitation from Satellite Observations

Round 1

Reviewer 1 Report

1. The adopted model is very deep, so it is supposed to analyze the feature well and estimate the precipitation accurately.

2. The model is adopted as it is; no hyper-parameters are tuned according to objective except a few.

3. A constant overestimation is noticed in several figures (Fig. 5, 10, and 11)

4. The accuracy for the low precipitation (< 5mm) is extremely good but suffers much for higher precipitation (> 5mm).

5. A few features are given in the wrong way to the model (Fig. 3) such as latitude, longitude, sine, and cosine of the day (to capture the cyclic components), which are given in a 2D shape, which does not make any sense. Since they do not change over space, they will eventually be ignored by a model. Instead, single values could have been given to the final extracted features, like a time series.

Some improvements in English language necessary. 

Author Response

Dear Reviewer,

Thank you for your thorough review. You will find below a point-by-point answer to your comments, including references to the modifications in the revised manuscript.

1. The adopted model is indeed deep, allowing it to capture intricate features for accurate precipitation estimation. However, we appreciate your suggestion and provided additional details about the model's architecture in paragraph Methodology/Model lines 198-210.
2. We acknowledge your concern about the hyper-parameter tuning. While the model was primarily adopted as-is, in order to use a proven and reliable architecture, we did perform a limited hyper-parameter tuning to ensure optimal performance. We included these details in the revised manuscript to clarify this point (Appendix B, lines 574-607).
3. The consistent overestimation observed in Figures 5, 10, and 11 is an important observation. We clarified this issue (lines 341-343 and 478-481) and provided insights into the possible sources of overestimation. We also discussed potential solutions in the section Discussion (lines 511-516).
4. Your observation regarding the model's performance for different precipitation levels is noted. We included a more detailed analysis of the model's accuracy for both low and high precipitation levels and discussed potential reasons for the observed performance variation on lines 352-354 and 359-364
5. Your insight into the input features' presentation, specifically the latitude, longitude, sine, and cosine of the day, is highly valuable. With regard to the sine and cosine of the day, we can indeed incorporate this information into the output of the final encoder layer before the first decoder layer. These values would be specific to each sample in the dataset. We have included your suggestion in the discussion section (lines 523-527) for future work. However, to the best of our knowledge, there is no theoretical reason for the model to disregard a constant presented as a 2D shape in the encoder input layer, and our approach appears valid. It would be appreciated if you could provide a reference on this matter. Regarding the latitude and longitude, these variables do vary in space, and when applying the neural network to the entire globe, it might not be feasible to implement your suggestion in this scenario. Furthermore, the inclusion of these input features was not retained in the subsequent experiments and does not compromise the validity of the subsequent comparisons.

Thank you again for your review which contributed to the refinement of our manuscript.

Sincerely,

Léa Berthomier

Reviewer 2 Report

Manuscript titled ‘Espresso : a Global Deep Learning Model to Estimate Precipitations from Satellite Observations’ by Léa Berthomier discussed about their operational deep learning based model Espresso, designed for estimating precipitation from satellite observations on a global scale. Author used geostationary satellite data as input and generating instantaneous rainfall rates, calibrated using data from the Global Precipitation Measurement Core Observatory for this deep learning model Espresso. Overall, manuscript is interesting, innovative and worth reading. I recommend this research article to publish in Meteorology after few minor changes.

Minor comments:

1. Figures numbering should be included for Figs. 5, 10 and 11, and elaborate Figs captions accordingly.

2. In lines 143-144: “To develop a global, real-time product, input data is based on geostationary (GEO) data. During the experimental phase, data from five geostationary satellites was selected.” correct sentences and grammars.

3. In lines 445-447: “Figure 10 also highlights Espresso’s excellent spatial resolution (5km), which is comparable to GHE, PDIR-NOW, and HSAF, and notably superior to IMERG and GSMAP (10 km).”

and 459-461: “While GHE, HSAF, and QPE were successful in pinpointing the storm’s location, they consistently underestimated its intensity. On the other hand, both Espresso and PERSIANN exhibited remarkable proficiency in storm localization and intensity estimation.”,

where are HSAF and PERSIANN in Figs. 10 and 11?

4. In Figure 2 latitude and longitude co-ordinates and figs numbering should be added to see their clear location. Captions also are needed to elaborate.

English is fine.

Author Response

Dear Reviewer,

Thank you for your thorough review. You will find below a point-by-point answer to your comments, including references to the modifications in the revised manuscript.

1. We added the figures numbering and modified the captions accordingly.
2. We rephrased the lines 143-144 in the revised manuscript.
3. Your observation about the missing references to HSAF and PERSIANN in Figs. 10 and 11 is valid. There was a confusion between HSAF/P-IN-SEVIRI and PDIR-NOW/PERSIANN. We modified the text accordingly to leave only the reference to the products names (P-IN-SEVIRI and PDIR-NOW), lines 458, 471 and 473.
4. We incorporated latitude and longitude coordinates in Figure 2, added figures numbering, and detailled the captions.

Thank you again for your review which contributed to the refinement of our manuscript.

Sincerely,

Léa Berthomier

Round 2

Reviewer 1 Report

The authors have mostly answered my concerns satisfactorily, except for my 5th concern. I feel the paper is suitable for publication. 

Minor language errors are detected.