Neural Network Approach to Forecast Hourly Intense Rainfall Using GNSS Precipitable Water Vapor and Meteorological Sensors

Round 1

Reviewer 1 Report

GENERAL REMARKS:

This study nicely explains a novel methodology for a possible short-term forecast of intense rainfall based on explained neural network system, with combining the available Global Navigation and Positioning System data, meteorological measurements and spaceborne observations. This method offers useful applications, in spite of well stated limitations such as not having longer time series (only 5 years) and data from different locations (having only 1 station) for the better assessment of this kind of short-term forecast of heavy rainfall events, also presented by authors. The manuscript complements supplementary figures and tables. However, there are just a few minor corrections and comments listed below that authors could address and correct before its final submission.

DETAILED REMARKS:

#1:

Line 157:

“2.3. SEVERI data.”

Please correct: “SEVERI” to “SEVIRI”.

#2:

Line 170-171:

Please correct for the correct tense: “can be visualize in Figure 2”.

Correct to: “can be visualized” or “is visualized”.

#3:

Line 174:

“the GAMIT/GLOBK software”

Please specify here a full name of the software acronyms and give info what it means for those who are not familiar of it.

#4:

Line 177:

Also, give a full name or short description of  “the ITRF14 reference frame”.

#5:

Line 312:

Please correct: “a slightly decrease” to “ a slight decrease”.

#6:

Line 346:

Please correct: “does not produces” to “does not produce”.

#7:

Line 398:

“It can observed the seasonal behavior of the different rainfall patterns, ..”

A beginning of this sentence is not clear. Please clarify or correct for a correct tense (e.g., a grammar issue: “We can observe the seasonal behavior of …”; Or “It [clarify what is it?] can observe the seasonal behavior of …”.

Author Response

We would like to thank the reviewer for the comments and suggestions. Please consider the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

This is a contribution on an important topic with interesting results. The reviewer however has the following remarks to make:

1.       The authors speak of using GNSS, but which GNSS did they use and was use made of multi-GNSS and if so how was this handled, see (Teunissen, Montenbruck (2017): Handbook of GNSS)?

2.       The authors study water vapour at a very localized region. Why is it then that they need precise orbits and global stations?

Author Response

We would like to thank the reviewer for the comments and suggestions. Please consider the attached file.

Author Response File: Author Response.pdf

Reviewer 3 Report

Review Summary and concerns:

This manuscript addresses a very interesting topic, the hourly intense rainfall forecasting, performed by using the neural network technique applied to the 5-year observation of GNSS precipitable water vapor and meteorological sensors. It is well written in all the sections. The experiment is well designed, described and explained, in particular, the network parameter configuration, optimization, and evaluation. With respect to the ground truth and the prior studies, the results justify the performance of neural network technique in intense rain forecasting is high correctness and low false positive rate. This work leads to a very high standard paper that absolutely deserves to be published.

I have very few comments:

a) The neural network is designed with the parameters of 1 hidden layer, less than 10 neurons, and so on, while the training set size is 2/3 of 5-year observations (~20000). The training set always plays a significant role in the performance of classification, however too large size set might result in the overtraining problem, for instance see [*]. How is this work designed to avoid this problem?

[*] Foody, G. M., McCulloch, M. B., & Yates, W. B. (1995). The effect of training set size and composition on artificial neural network classification. International Journal of Remote Sensing, 16(9), 1707-1723.

b) Overfitting is a very important problem in neural networks as well. The authors have designed several groups of comparative experiments in order to obtain the optimal network parameters. Do the experiments meet the overfitting case? If not, how is the experiment designed to prevent this problem? I would like to be thankful if it could be described.

c) #256 In Table 3, the values of defined three classes for Rain classification should be the same as described in line #245.

Author Response

We would like to thank the reviewer for the comments and suggestions. Please consider the attached file.

Author Response File: Author Response.pdf