Rainfall Forecast Model Based on the TabNet Model

Round 1

Reviewer 1 Report

Dear Authors,

I have revised your manuscript again. I still have some concerns, which need to be addressed before considering for final publication.

General comments:

1. The line numbering is missing.
2. References should be numbered in order of appearance. Especially check the second paragraph on page one, the text in red.
3. Use the current version of the template - https://www.mdpi.com/files/word-templates/water-template.dot

Specific comments:

1. Subsection "3.2. Evaluation metric". Since you decided to use the MAE evaluation metric, in my opinion it is worth adding another measure of error prediction which is MAPE (Mean Absolute Percentage Error) because it is expressed as a percentage. For convenience, refer to the following papers: https://doi.org/10.3390/rs12111744, https://doi.org/10.3390/app9142773, https://doi.org/10.3390/rs12061024.
2. Unfortunately, my opinion is still actual. Section "3. Results and Discussion" should be divided into two separate sections "Results" and "Discussion". In addition, the discussion of results should be expanded to include analysis of results from other papers. Try to compare the results obtained by the TabNet method with the results obtained by other methods. This comparison will be very interesting. There are many methods and papers on the subject of rainfall forecasting. New references should also be added.
3. Section "References" still needs to be improved. Style of the references is not in accordance with the requirements of Water. I suggest using bibliography software package like Mendeley, Zotero, EndNote etc.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This study proposed a machine learning-based approach for rainfall predictions using 5 years of meteorological data. Overall the manuscript looks like a DL-based report instead of a research paper. Again, the potential contribution is not significant. My major comments and questions are as follows:

• The introduction section is poor. What were the previously established ML/Dl techniques available to predict precipitation? What is the reason for choosing your DL technique over other established ML/DL techniques? There are tons of ML-based forecasting models such as Quantile regression forest, SVM, boosted tree, decision tree, neral network etc., which were successfully used to estimate precipitation. You have to introduce these recently established nonparametric/machine learning techniques which successfully estimate precipitation. Then you should propose your framework and provide the uniqueness of the data blending algorithm and its potential impacts, over other recently established state-of-the-art ML techniques in hydrologic applications? Please introduce these works and their potential impact. The authors should explain this aspect in the introduction section. Otherwise, the readers cannot see the importance of your proposed methods over other techniques. The authors need to provide more literature reviews in the introduction section. Please follow the recent papers    mostly focusing on ML-based rainfall forecast:

1. Zhang, Ling, et al. "Merging multiple satellite-based precipitation products and gauge observations using a novel double machine learning approach." Journal of Hydrology 594 (2021): 125969.
2. Derin, Yagmur, et al. "Modeling Level 2 Passive Microwave Precipitation Retrieval Error Over Complex Terrain Using a Nonparametric Statistical Technique." IEEE Transactions on Geoscience and Remote Sensing (2020).

1. Zhou, Yuanyuan, et al. "Assimilation of Multi-Source Precipitation Data over Southeast China Using a Nonparametric Framework." Remote Sensing 13.6 (2021): 1057.
2. Yin, Jiabo, et al. "Blending multi-satellite, atmospheric reanalysis and gauge precipitation products to facilitate hydrological modelling." Journal of Hydrology 593 (2021): 125878.
3. Zhou, Yuanyuan, et al. "Assimilation of Multi-Source Precipitation Data over Southeast China Using a Nonparametric Framework." Remote Sensing 13.6 (2021): 1057.
4. Bhuiyan, et al.  Machine Learning-Based Error Modeling to Improve GPM IMERG Precipitation Product over the Brahmaputra River Basin. Forecasting 2020, 2, 248-266.

• You need to present the study area and data section  after the introduction section (here is the order: Introduction, Study area/Data, methodology, model evaluations, results, discussions, conclusions)
• Can you please provide a table for validation/training/testing samples along with other information? How much dataset used for validation and testing? Please justify?
• You need to verify rainfall prediction results in terms of season. Also, show some plots

1. Provide time series for original rainfall and predicted rainfall
2. Random error (RMSE) and Systematic error (MRE)

• What is the impact of the model for extreme rainfall predcition? Please add a discussion section and compare your results with previously established results?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Overall, the authors significantly improved the manuscript by addressing most of the comments. I recommended the manuscript for publication. Congratulations to the authors!

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

In my opinion, this manuscript should be rejected and reconsidered after an in-depth reorganization. In details, Authors only focused on the probability of occurrence for rainfall but, into a context of rainfall prediction, intensity of rainfall should be also investigated, mainly for real-time frameworks. 

Reviewer 2 Report

This study proposes a precipitation forecasting model based on the Tabular Learning (TabNet) neural network utilizing self-supervised learning along with feature engineering methods to improve the performance of the model.  I always welcome studies with the aim of improving the Spatio-temporal characterization of precipitation estimates.  In my view, it does not make a new contribution in terms of precipitation forecast. The novelty is a key criterion for our selection of manuscripts to be considered for publication in the Water Journal. I have some recommendations for future submission.

• The authors mostly focused on Tab Net architecture construction instead of significant science/engineering applications. In your introduction section, you mainly explained your proposed method. First, you should discuss the significant uncertainties in satellite datasets throughout warm and cold seasons, over the study regions. Then you should introduce other Machine learning-based rainfall forecasting models where they utilize satellite precipitation products along with auxiliary variables such as soil moisture, temperature elevation, etc. But you mentioned very limited study. You just mention one study which is the random forest forecasting model. There are tons of ML-based forecasting models such as Quantile regression forest, SVM, boosted tree, etc., which you need to introduce. I think the authors need to propose a detailed comprehensive introduction section. What is the uniqueness of the proposed technique and its potential impacts, over other established techniques? The authors should explain with a couple of new paragraphs on this aspect in the introduction section. Also, you need to provide more literature review in the introduction section associated with the research gap/limitation

At the global scale, precipitation estimation primarily relies on satellite-based observations. Why you did not consider any of them. A better explanation of why the authors did not consider satellite precipitation products such as  TRMM 3B43v7, PERSIANN-CDR, and IMERG were required? For instance, why you did not include Global Satellite Mapping of Precipitation (GSMaP), and Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks (PERSIANN)? Recent studies showed that extensive evaluation of all these global-scale high-resolution satellite-based rainfall (SBR) products were evaluated and explained the merit of those precipitation products in accurate precipitation estimates. The authors should explain this aspect in the introduction section. Please include GSMaP and CMORPH in your analysis.

• Can you provide a study area map where you can indicate your selected catchments? As the model was applied for different climatic conditions, can you provide climatic information for the selected study areas? You can show Köppen–Geiger climatic zones on the map.

• I suggest the use of the modified Kling-Gupta efficiency. This index decomposes the total performance of the precipitation products into linear correlation (r), bias (beta), and variability ratio (gamma). Also, the performances obtained from this index can be compared among precipitation products.

• Did you take into consideration the reporting time difference between your features and observation rain? Please explain. You should provide a new section dataset and study area and explain.

• Discussion: The discussion must be expanded, and more references added. In the introduction, you mention articles that can be used to discuss and compare your results.

Here are the few ML-based rain forecast model based on quantile regression forests (QRF), Random forest( RF), SVM, neural net  :

Bhuiyan, et al. "A nonparametric statistical technique for modeling overland TMI (2A12) rainfall retrieval error." IEEE Geoscience and Remote Sensing Letters 14.11 (2017): 1898-1902.

Zhang, Ling, et al. "Merging multiple satellite-based precipitation products and gauge observations using a novel double machine learning approach." Journal of Hydrology 594 (2021): 125969.

Chiang, et al. "Precipitation assimilation from gauge and satellite products by a Bayesian method with Gamma distribution." International Journal of Remote Sensing 42.3 (2021): 1017-1034.

 Banadkooki, et al. "Precipitation forecasting using multilayer neural network and support vector machine optimization based on flow regime algorithm taking into account uncertainties of soft computing models." Sustainability 11.23 (2019): 6681.

Banadkooki, et al. "Precipitation forecasting using multilayer neural network and support vector machine optimization based on flow regime algorithm taking into account uncertainties of soft computing models." Sustainability 11.23 (2019): 6681.

Kolluru, et al. "Secondary precipitation estimate merging using machine learning: development and evaluation over Krishna river basin, India." Remote Sensing 12.18 (2020): 3013.

Reviewer 3 Report

Authors have properly addressed my concerns. I'm good with the current paper!

Reviewer 4 Report

Dear Authors,

I revised the manuscript "Rainfall forecast model based on the TabNet Model" submitted to Water journal. The manuscript is very interesting. However, I have some concerns, which need to be addressed before considering for final publication.

General comment

References should be numbered in order of appearance.

Specific comments

Table 1. In the "Value" column, use English names instead of Chinese names. Check again the numeric values contained in the "Value" column. In my opinion, the last column "Unit" should be split into two separate columns "Unit" and "Precision".

In my opinion, section "3. Results and Discussion" should be divided into two separate sections "Results" and "Discussion". In addition, the discussion of results should be expanded to include analysis of results from other papers. Try to compare the results obtained by the TabNet method with the results obtained by other methods. This comparison will be very interesting. There are many methods and papers on the subject of rainfall forecasting. New references should also be added.

Section "References". Style of the references is not in accordance with the requirements of Water. I suggest using bibliography software package like Mendeley, Zotero, EndNote etc.