IWQP4Net: An Efficient Convolution Neural Network for Irrigation Water Quality Prediction
Round 1
Reviewer 1 Report
Title: IWQP4Net: An Efficient Convolution Neural Network for Irrigation Water Quality Prediction
ID: water-2322904
The manuscript focuses on the prediction of pH in irrigation waters, which is a practical and interesting topic. The authors employed an ML technique, CNN, to predict the water quality for the following day, using some daily water parameters in different stations. They also compared the results with other models to assess their results.
Although the manuscript is well structured, in my opinion, there are some gaps in the description of the input database, stations, hyperparameter tuning, run time, etc.
Some other major comments that need to be addressed are as follows:
Title:
- What is the logic behind “4Net” in “IWQP4Net”?
Abstract:
- 3 lines of the 12 lines are dedicated to the introduction and the importance of the subject which is quite long.
- Another 3 lines have been assigned to details of the model architecture, which is not only quite long but also no need to mention these details in the abstract.
Introduction:
- Reference 18 employed the CNN method to forecast water quality. What is the superiority of the present model over that?
- Contributions 1 and 2 are almost the same: developing a CNN model for IWQ prediction.
Section 2. Methods and Materials:
- In section 2.1 it is mentioned that daily samples are taken from 36 locations; while in section 2.2 it is mentioned regression will have 37 outputs. Is it only a typo? Also, in 2.1, it is mentioned 12 parameters; while in Figure 1, the number of features is 11.
- From 12 input features, only temperature, specific conductance, and the volume of dissolved oxygen have been addressed. What are the others? Do those parameters used as images or input database includes values and numbers?
- It is mentioned that the input data is made up of daily samples but has not been specified how many days are considered. How much data is used for training? Is it tuned?
- In this section, it is supposed to be indicated the software and library used for the CNN model. Also, hyperparameters and the tuning process should be presented in detail.
- Explain more about the number of the training set and testing set.
Section 4. Experimental results and discussion
- Fig 4 depicts that SVR and kNN perform better than IWQP4Net both in training and testing.
For example, the slope of the regression line in testing for IWQP4Net is 0.098 while for SVR and kNN are 0.356 and 0.302, respectively.
Fig 5Histograms of residuals do not confirm the superiority of CNN in my point of view.
5. Conclusion and future works
The future work stated in the conclusion is a general statement which already has been applied to CNN techniques in different aspects of knowledge. However, the proposed IWQP4Net model cannot be used for other different applications and databases without any tuning or adjusting hyperparameters.
Author Response
Dear Editor,
Thank you for allowing a resubmission of our manuscript, with an opportunity to address the reviewers’ comments.
We are uploading (a) our point-by-point response to the comments (below) (response to reviewers), and (b) an updated manuscript with yellow highlighting indicating changes,
Response to Reviewers comments:
We are very much thankful to the reviewers for their deep and thorough review. We have revised our present research paper in the light of their useful suggestions and comments. We hope our revision has improved the paper to a level of their satisfaction. Number wise answers to their specific comments/suggestions/queries are as follows.
Reverser 1
The manuscript focuses on the prediction of pH in irrigation waters, which is a practical and interesting topic. The authors employed an ML technique, CNN, to predict the water quality for the following day, using some daily water parameters in different stations. They also compared the results with other models to assess their results.
Although the manuscript is well structured, in my opinion, there are some gaps in the description of the input database, stations, hyperparameter tuning, run time, etc.
Some other major comments that need to be addressed are as follows:
Author response: Thank you very much for your opinion in our work
Title:
- What is the logic behind “4Net” in “IWQP4Net”?
Author response: The acronym 'IWQP' refers to 'Irrigation Water Quality Prediction' and '4Net' refers to 4 layers with learnable (weighted) parameters in the prediction model. This is in agreement with state-of-the-art LeNet (LeCun et al. 1998), AlexNet (Krizhevsky et al., 2012), and VGG (Simonyan and Zisserman 2015) networks.
Abstract:
- 3 lines of the 12 lines are dedicated to the introduction and the importance of the subject which is quite long. Another 3 lines have been assigned to details of the model architecture, which is not only quite long but also no need to mention these details in the abstract.
Author response: Thank you for your point, the abstract is revised to reduce the importance of the subject and details of the proposed model as suggested
Introduction:
- Reference 18 employed the CNN method to forecast water quality. What is the superiority of the present model over that?
Author response: The model in [18] used a hybrid CNN-LSTM to predict water pH quality for single station. It comprises 6 trainable layers with approximately twice the number of neurons in each layer than the proposed IWQP4Net. Hence, the model in [18] is computationally expensive for simple regression while the proposed model is computationally inexpensive for multiple regression. A direct comparison of the model in [18] and IWQ4Net models is not possible as both are solving different problems using different datasets.
- Contributions 1 and 2 are almost the same: developing a CNN model for IWQ prediction.
Author response: Thank you for your comment, contribution 1 and 2 are combined and a 3rd contribution is added
Section 2. Methods and Materials:
- In section 2.1 it is mentioned that daily samples are taken from 36 locations; while in section 2.2 it is mentioned regression will have 37 outputs. Is it only a typo? Also, in 2.1, it is mentioned 12 parameters; while in Figure 1, the number of features is 11.
Author response: Thank you for your comment, the dataset comprises water quality values for 37 water stations represented using 11 features. We have revised the corresponding sections to correct the information.
- From 12 input features, only temperature, specific conductance, and the volume of dissolved oxygen have been addressed. What are the others? Do those parameters used as images or input database includes values and numbers?
Author response: Thank you for your comment, the dataset used in this work is extracted from derived from the United States Geological Survey (USGS Water Data for the Nation Help, available at: https://help.waterdata.usgs.gov/), as cited in the paper. The dataset is available as number without any metadata for the subset of the features selected from the complete water data available on the original source. Hence, we have not discussed the remaining features in this papers.
- It is mentioned that the input data is made up of daily samples but has not been specified how many days are considered. How much data is used for training? Is it tuned?
Author response: Thank you for your comment, the exact period of data collection is added to section 2.1.
- In this section, it is supposed to be indicated the software and library used for the CNN model. Also, hyperparameters and the tuning process should be presented in detail.
Author response: Thank you for your comment, the software details for CNN model implementation is assed to section 4 line 192. This paper does not use any automatic architecture search for model optimization. The subjective analysis used for tuning the hyperparameters is mentioned in section 4 lines 198–209.
- Explain more about the number of the training set and testing set.
Author response: Thank you for the comment, training and testing sets are explained in detail.
Section 4. Experimental results and discussion
- Fig 4 depicts that SVR and kNN perform better than IWQP4Net both in training and testing.
For example, the slope of the regression line in testing for IWQP4Net is 0.098 while for SVR and kNN are 0.356 and 0.302, respectively.
Fig 5 Histograms of residuals do not confirm the superiority of CNN in my point of view.
Author response: Thank you for the comment, Fig 4 depicts one of the simulated cases for each individual model. The SVR depicts the best performance for an iteration, average performance of the IWQP4Net is better than all competing models as reported in Table 5. Fig 4 is revised to depict the best possible cases for each individual model which is in unison with results shown in Table 5.
Fig 5 depicts the distribution of error prediction (residual) error for each model. The more examples predict an error closer to 0 for IWQP4Net than other models for both training and testing data. Hence, the proposed CNN is qualitatively better than other competing models.
- Conclusion and future works
The future work stated in the conclusion is a general statement which already has been applied to CNN techniques in different aspects of knowledge. However, the proposed IWQP4Net model cannot be used for other different applications and databases without any tuning or adjusting hyperparameters.
Author response: Thank you for your comment, future work part has been fixed as suggested.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Dear Editor,
Water
In this paper, an IWQP 253 approach is developed by using an efficient CNN structure. The subject addressed is interesting and within the scope of the Water. Nevertheless, some major revisions have been found:
- The statistical criteria of the inputs and output dataset should be presented in one table.
- How do you are determined the structure of IWQP4Net? It is suggested to structure of IWQP4Net is determined by a grid search method or optimization method such as a particle swarm optimization algorithm. Selecting the optimal structure of IWQP4Net may lead to more accuracy of them.
- In this study, pH was used as a water quality criterion. However, there are other criteria such as EC, TDS, COD, BOD, etc. It is better in figures and other sections of the manuscripts, the phrase "pH" is mentioned instead phrase of "water quality".
- The accuracy of modeling water quality is not presented clearly. It is suggested to relative criteria such as mean absolute percentage error or relative root mean square error are employed to solve this problem. Also, RMSE is the root of MSE and one of RMSE or MSE should be presented.
- It seems that the accuracy of algorithms is low, this issue should be solved. Models that are not accurate cannot help water resources management. I recommend using the biggest dataset, feature selection method, using input data with lagging time, and data preprocessing methods such as wavelet transform.
- Please compare the results of the present study with previous studies.
Considering the mentioned points, this study in the current version needs major revisions.
With kind regards,
Author Response
Dear Editor,
Thank you for allowing a resubmission of our manuscript, with an opportunity to address the reviewers’ comments.
We are uploading (a) our point-by-point response to the comments (below) (response to reviewers), and (b) an updated manuscript with yellow highlighting indicating changes,
Response to Reviewers comments:
We are very much thankful to the reviewers for their deep and thorough review. We have revised our present research paper in the light of their useful suggestions and comments. We hope our revision has improved the paper to a level of their satisfaction. Number wise answers to their specific comments/suggestions/queries are as follows.
Reverser 2
In this paper, an IWQP 253 approach is developed by using an efficient CNN structure. The subject addressed is interesting and within the scope of the Water. Nevertheless, some major revisions have been found:
- The statistical criteria of the inputs and output dataset should be presented in one table.
Author response: Thank you for the comment, the summary of input and output feature statistics is added in table 1.
- How do you are determined the structure of IWQP4Net? It is suggested to structure of IWQP4Net is determined by a grid search method or optimization method such as a particle swarm optimization algorithm. Selecting the optimal structure of IWQP4Net may lead to more accuracy of them.
Author response: Thank you for the comment, the structure for IWQP4Net is determined by iterative search over varied range (grid) of hyper-parameters. The subjective criteria used to finalize the optimum model architecture is mentioned in section 4 lines 198–209. In future, optimization method may be explored to improve and compare the model performance.
- In this study, pH was used as a water quality criterion. However, there are other criteria such as EC, TDS, COD, BOD, etc. It is better in figures and other sections of the manuscripts, the phrase "pH" is mentioned instead phrase of "water quality".
Author response: Thank you for the comment, the dataset used in this work has only pH as a water quality index. The dataset with multiple water stations using other water quality criteria is not available. The phrase "water quality" is replaced by "pH" at appropriate places in the paper.
- The accuracy of modeling water quality is not presented clearly. It is suggested to relative criteria such as mean absolute percentage error or relative root mean square error are employed to solve this problem. Also, RMSE is the root of MSE and one of RMSE or MSE should be presented.
Author response: Thank you for the comment, the result section is revised to discuss the mean absolute percentage error. The result duplication caused by RMSE and MSE is resolved by removing MSE.
- It seems that the accuracy of algorithms is low, this issue should be solved. Models that are not accurate cannot help water resources management. I recommend using the biggest dataset, feature selection method, using input data with lagging time, and data preprocessing methods such as wavelet transform.
Author response: Thank you for your comments, the current work pioneers the use of CNN for water quality prediction using multiple regression. As per the authors' knowledge no other dataset with such specification is available. This work reports the preliminary analysis using CNN model. In future the accuracy can be increases by data preprocessing and feature selection techniques, as suggested.
- Please compare the results of the present study with previous studies.
Author response: Thank you for the comment, as per the authors' knowledge earlier reported works on water quality prediction used simple regression i.e. dataset with single water station output. In this work, multiple regression uses dataset with output of 37 water stations and hence, a direct comparison with earlier reported works is not possible. A comparison of most commonly used machine learning models as suggested in earlier works is presented in the current scenario.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Most of the comments are addressed properly.
Just all the input parameters need to be introduced but only a couple of them have been named so far.
Author Response
Most of the comments are addressed properly.
Author response: Thank you very much for accepting our work
Just all the input parameters need to be introduced but only a couple of them have been named so far.
Author response: Thank you for your comment, Unfortunately, the information in regards input parameters is not available anywhere. We, the authors checked other papers, author's website and even official US data website
Reviewer 2 Report
Dear Editor,
Water
This manuscript is previously carefully evaluated. The current version of the manuscript is acceptable. .
Author Response
This manuscript is previously carefully evaluated. The current version of the manuscript is acceptable. .
Author response: Thank you very much for accepting our work
Round 3
Reviewer 1 Report
Although it is strange to conduct a model with unknown input data, the manuscript can be published in the present form.
