Rainfall Forecast Using Machine Learning with High Spatiotemporal Satellite Imagery Every 10 Minutes
, Ilham Jamaluddin 3, Tang-Huang Lin 4,5,*, Hary Aprianto Wijaya Siahaan 2,6 and Ying-Nong Chen 3,4Round 1
Reviewer 1 Report
Review of “Rainfall Forecasting Using Machine Learning with High-Spatiotemporal-Resolution Satellite Imagery Every 10 Minutes” by Febryanto Simanjuntak, Ilham Jamaluddin, TangHuang Lin, Hary Aprianto Wijaya Siahaan, and Ying-Nong Chen.
The study presents a machine learning methodology to forecast rainfall and rain rate in Indonesia. The study utilized data from Himawari-8 and IMERG to train and initialize the machine learning model described. Model accuracy in forecasting the rain rate at different temporal resolutions was estimated using the data obtained from IMERG and 41 meteorological stations located across the study area.
Overall, a very well-written manuscript with some issues. The figures were fine and legible. The authors did a good job of explaining their methodology in detail. I have some major concerns which are listed below. Overall, the manuscript feels more like a case-study analysis without proper justification of case selection or limited data utilization. I feel that more can be done with the results obtained/presented and more cases are needed to test the robustness of the methods described. The manuscript has the potential for publication but needs major revisions.
Major comments:
Comment #1: Authors should provide justification for why only a single case was tested or utilized in the study. I feel this is the biggest weakness of the study presented. This is required to address the limitations of the model, and the dependence of the relative errors on the extremeness of the rainfall event.
Comment #2: Authors identified different islands/regions in Figure 1 but did not try to test the sensitivity of the geography on the forecast accuracy. Given the analysis performed, this seems possible and would increase the robustness of the model. Authors should at least test the model accuracy for the Regions mentioned in Lines 145-155. The authors did a qualitative check regarding this near Lines 370-374, but more such analysis is required.
Comment #3: Discussion around Figures 5 and 6 seems heavily based on visual inspection or rather qualitative. Later in Sec 3.2, MAE and RMSE were reported. Did the authors re-sampled their model forecast data to match the spatial resolution as IMERG in evaluating the error? In Table 2, K is mentioned as the simulated data but no information on its spatial resolution. It would be interesting to see the histogram of actual error (i.e., Ki-Oi) to see the range of errors and their occurrence frequencies. MAE does not tell if the model is biased towards over- or under-estimating, but the histogram of actual error would give that information. This is needed, because, from Figure 6 it is obvious that the model failed in predicting the highest rainfall rate that is observed in the red circles in Figure 6. Since the gridded data for rainfall rate is available, it would benefit the manuscript if a time series analysis can be conducted (for both individual regions and overall) to see how the model behaves with an increase in forecast time.
Minor comments:
Comment #1: Please mark Regions A, B, and C in Figure 1.
Comment #2: Line 218: wrong rain rate units
Comment #3: Line 258: “forget get layer”?
Comment #4: Figure 4: mark the land boundaries.
Comment #5: Line 451: Figure 2 or Figure 1?
Comment #6: Table 3: What is RnR?
Comment #7: I feel Figure 5 is redundant as the information about Rain or no-Rain is available from Figure 6.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper developed a novel method to utilize Himawari-8 and GPM-IMERG satellite data for high spatial and temporal resolution precipitation prediction. The application of LSTM and RF for precipitation generation is interesting, but those flaws should be fixed before a proper review. My comments are as follows.
For the evaluation of the precipitation forecast, comparisons can be made with other rainfall products (e.g., CHIRPS, MSWEP) or higher temporal resolution (e.g., 10 minutes) automatic weather stations. Moderate English improvements are required before the paper can be considered for publication. What follows below is a sample, and the list is not exhaustive.
Line 20: In my opinion, ”the results of the numerical weather model used by the Indonesia Agency for Meteorology, Climatology, and Geophysics are only able to predict the rainfall with a temporal resolution of 1–3 hours and cannot yet address the need for rainfall information with a high spatial and temporal resolution.” Is not needed. If the method authors promoted relies on specific regions?
Introduction can review more recently published papers, especially novel technologies they have provided in precipitation product generation based on machine learning. “remote sensing” journal provided and reviewed a lot of similar articles in the field of combining different machine learning methods (LSTM or RF) to produce more accurate precipitation data with multiple remote sensing data.
Line 104: If GPM is necessary to be included?
Line 252: In Figure 2, what do the different colors represent for? What does T and their subscript indexes represent for? More details should be added in the caption below the figure.
Line 268: How many bathes used here?
Line 284: Figure 3 is not clear enough.
Line 338: Figure 4 location is not proper.
Line 410: The density of rainfall has not been well presented by the forecast.
Line 414: A spatial distribution evaluation of the predicted rain rate can be made with the IMERG as a ground truth, for example, a map of RMSE or correlation coefficient can illustrate the performance of the forecast more intuitively.
Line 425: Discussion can be improved. Results can be made comparisons with recently published papers or other rainfall products (e.g., CHIRPS, MSWE, etc.).
Line 428: Reference format should be Author date.
Line 433: Reference format should be Author date.
Line 437: The same as above.
Line 445: Not correct.
Line 456: Table 4 is not needed.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have done a good job addressing the comments. I would like to thank the authors for considering many of the suggestions from round 1 and incorporating them into the manuscript.
The manuscript reads fine now and is publication-ready after addressing the very minor comments listed below. Also, please make sure to upload a clean copy with track-changes removed, as I had difficulty in figuring out text/figures that are still in the manuscript, example: Figures 2,3, and 6 created huge white spaces on pages 8, 11, and 17 respectively with some text still visible. This makes it difficult to know whether to read through it.
Minor comments:
1) Figure 1 looks really busy. There are a lot of letters in different colors, but they are not described in the figure caption or in the text. I assume the authors meant to show the Regions used in the study and have duplicate markers. Please correct this.
2) Give units for the color field when a color bar is used.
3) Figure numbers might be wrong. I see two Figure 5s.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
In my opinion, the present form is suitable for the publication.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
