Development of Machine Learning Flood Model Using Artificial Neural Network (ANN) at Var River
Round 1
Reviewer 1 Report
The article is about comparative model base on ANN method, MIKE 11 model and others methods. The models works with daily step and are applied to mountain basin. For evaluation are used standards criteria.
Recommendations:
1)Firstly at the article is used over 60 different citation, which is rather lot if the most of cited literature is used in Introduction and does not have any direct impact on further text.
2) In the article I am missing figures of results for others methods. I think that authors should considered theirs placing in the text.
3) Also more detailed info about basin such a land cover, open large reservoir and lake (if some are located) should be provided.
Small problems:
1) Authors have to check the units and their formats in the text. In the article is too much mistakes in this matter. Example :
The study area used for this investigation is a 44.93 km2 sub river Vésubie which is located
southeast part of France at Latitude: 44°04′12′′ N, Longitude: 7°15′19′′ E with elevation above sea level: 964 m = 3162 ft. River Vésubie is one of the segments of the River Var (such as Tinee, Upper Var, Esteron, Vésubie, and Lower var) covering an area of 393.35 km2. With a 37.19 % average slope, river Vésubie has min 152 m and a maximum of 3001 m elevations.
Problems:
1) I am not sure if only one meteorological station is enough for sufficient descriptions for chosen parameters in basin as potential evapotranspiration and temperature if height gradient in the basin is considered.
2) In the article I am missing information about set parameters for ANN (number hidden neurons in layers, number hidden layers,…..) and for others methods. Usually during training ANN is data set divided to training, validation and testing data sets. So I would like to ask if this divided was provided by random or different way.
3) Training period is really short and models have to work with snow during some months. During winter and snow melting period and rest of year are different dependence between rainfall and discharge. So ANN methods are usually trained separately for this period. Did you consider this approach?
4) I am also missing some parameters which could describe previous rainfall or saturation, which has usually rather strong impact on run off and could lead to better results for others methods than NAM model, which does not used this parameters in this application.
5) I am also missing explanation why NAM model has such strange behaviour during validation period (fig 10b from 2011-01 to 2011-09 and from 2014-01 to 2014-09).
6) Form results showed in fig 10a is reasonable thought that ANN methods are trained on average behaviour and lost its performance during drought period (too much inputs with null rainfall value). Therefore I am not sure if this is desire behaviour.
5) I am also missing explanation why Decision three has such poor performance during validation period when during training period has correlation value almost 1.
Result:
In my opinion the results of ANN model could be better if authors used some others parameters (such as saturation, previous rainfall) or use some training data (such a K-fold cross validation method), but I am mostly missing explanation of results and behaviour of models. Also I am not sure what main results of article are. Is it testing ANN models tool for simply simulation of run off process or evaluation of simple NAM model performance. Authors should better state what main idea of the article is. Also in the conclusion chapter the authors should explained or discuss (provide their idea) for some results which was achieved.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The Authors present a method that the performance of the Artificial Neural Network model to predict discharge(output) by using river flow characterization. Specifically, Feed Forward Neural Network(FFNN, a type of ANN), random forest (RF), decision tree (DT), and classic numerical model (Hydrodynamic Models in MIKE) were developed and compared by several performance indicators (R, MAE,RMSE). Despite the work is relatively complete, some improvements should be done in order to make the manuscript eligible for publication. Some comments are followed (grouped by section).
v Introduction
“There are plenty of Artificial Neural Network models available and research scientists are still figuring out which models should be employed or best fit for a specific problem [61-67].”
The sentence should be expanded. Which ANN models are available for your research problem? All relevant ANN models should be shown in this part. And clarify why you choose the FFNN model.
v Materials and Methods
The parameters of all the models used in this study were not shown appropriately, especially for ANN model. The comparison of algorithms is based on the parameterization. If the parameters are not proper, the performance of the algorithm will vary dramatically.
v Results and Discussion
1) In the description of Fig. 10, “It is also observed that ANN has better validation efficiency as compared to calibration in comparison to the ANN model.” What does this sentence mean? ANN is to compare with ANN?
2) “In a comparison of all the performed models, Decision trees and Random forests have the best prediction, because of their low overfitting and easy interpretability. ” This is inconsistent with the information of table 3. According to table 3, DT and RF have good performance on training data but perform badly on testing data. Please clarify this point.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
I believe that authors answered almost all my questions to my satisfaction.
Minor problems:
Authors should at least write some percentage representation to land cover in line 112.
Only some units are missing °C and authors should hold unitary in m3/s. Please check units in lines 105-114 and 144-156.
Recommendation:
As you write in responded letter. Model which used inner parameters should be started with enough time frame before validation period or period in which is not inner parameters found (stabilized) should not be evaluated, because in this period are expected poor result, which can strongly influence criteria results value. So in next research you should continue with trained NAM model to validation period and not start NAM model from beginning. You also did not start training ANN model from beginning for validation period.
Reviewer 2 Report
I have no further comments. The authors have provided the substantial revision on the manuscript.
