Next Article in Journal
A Comparative Analysis of Remotely Sensed and High-Fidelity ArcSWAT Evapotranspiration Estimates Across Various Timescales in the Upper Anthemountas Basin, Greece
Previous Article in Journal
A Paleo-Perspective of 21st Century Drought in the Hron River (Slovakia)
 
 
Article
Peer-Review Record

Just a Single-Layer CNN for Stochastic Modeling: A Discriminator-Free Approach

Hydrology 2025, 12(7), 170; https://doi.org/10.3390/hydrology12070170
by Evangelos Rozos
Reviewer 1: Anonymous
Reviewer 2:
Hydrology 2025, 12(7), 170; https://doi.org/10.3390/hydrology12070170
Submission received: 17 June 2025 / Accepted: 24 June 2025 / Published: 29 June 2025
(This article belongs to the Section Statistical Hydrology)

Round 1

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

The authors have addressed my comments, and I am happy with the answers

Reviewer 2 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

I congratulate the authors for their efforts and thoroughness and recommend the publishing of this article, which would make a great addition to the scientific community.

Best regards,

The reviewer.

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

Comments and Suggestions for Authors

In this work, the author proposes a convolutional neural network (CNN)-based approach for stochastic generation in the field of hydraulics. This approach is intended to address the limitations typically associated with generative adversarial networks (GANs), such as the requirement of a high segmentation level for the discriminator component. The manuscript is written in a clear and coherent manner, and the scientific methodology is sound and justified.Upon addressing the following comments, I strongly recommend the publication of this paper.  

Major issues:

  • In the interest of aligning the discussion with the journal's scope, it is recommended that the author orient the introduction with a focus on the utilisation of generative ML approaches within the domain of hydrology. The authors may choose to expand upon the utilisation of Generative Adversarial Networks (GANs) in the extant literature pertaining to hydrology, with a particular emphasis on accentuating their significance and the constraints that have been identified in this context. The following is a non-exhaustive list of papers with which the author may enrich the introduction :
    • Ferchichi, A., Chihaoui, M., & Ferchichi, A. (2024). Spatio-temporal modeling of climate change impacts on drought forecast using Generative Adversarial Network: A case study in Africa. Expert Systems with Applications, 238, 122211.
    • Karimanzira, D. (2024). Mass Conservative Time-Series GAN for Synthetic Extreme Flood-Event Generation: Impact on Probabilistic Forecasting Models. Stats, 7(3), 808-826.
    • Belhajjam, R., Chaqdid, A., Yebari, N., Seaid, M., & El Moçayd, N. (2024). Climate-informed flood risk mapping using a GAN-based approach (ExGAN). Journal of Hydrology, 638, 131487.
    • Ma, Z., Mei, G., & Xu, N. (2024). Generative deep learning for data generation in natural hazard analysis: motivations, advances, challenges, and opportunities. Artificial Intelligence Review, 57(6), 160.
  • In Figure 9, the author provides a justification for the inadequacy of the modelling framework by introducing measurement error. Nevertheless, there is an absence of empirical evidence to support this assertion. It is recommended that the author performs a data cleansing procedure on the measurement prior to commencing a comparison, as this may otherwise compromise the integrity of the results and the validation of the work.
  • As illustrated in Figures 10, 11 and 12, the author conducts a comparative analysis of the climacogram, thereby demonstrating the efficacy of the CNN-based approach. Nevertheless, all comparisons are conducted visually. It is recommended that the author should endeavour to establish a method of quantification of the errors with reference to the aforementioned statement, in order that the veracity of the statement may be substantiated.
  • The author demonstrates the capacity of the CNN-based approach to conserve statistical properties during the development of the generation scheme. Nevertheless, the author has developed the methodology and comparison solely for a single study case. It is therefore evident that the results may be compromised if the generalisation is not ensured.
  • The author has provided a justification for the utilisation of CNN as an alternative approach to GAN-based methods, albeit in a theoretical context. Nevertheless, the absence of any utilisation of a GAN-based approach within the scope of the work prompts the question of the work's motivation. It is incumbent upon the author to develop a GAN-based approach and to assess the limitations identified in the introduction. Furthermore, it is essential to ensure that the CNN-based approach has effectively addressed the problem.

Minor isssues:

  • Figure 5 is not quite visible.
  • Figure 6 should be recentred.
  • It is evident that Figures 7, 8 and 9 require further refinement. The frequency data is only barely legible. It is imperative that the caption and title of the y-axis explicitly state that the log scale is employed, as the presence of negative frequency values is a source of concern.

Comments for author File: Comments.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Review report - "Just CNN for Stochastic Modeling: A Discriminator-Free Approach"

By Evangelos Rozos.

 

A. Summary.

This research presents a well-written exploration of a novel stochastic simulation scheme. It proposes utilizing a single convolutional neural network (CNN), functioning as a generator, to model the statistical properties of complex time series. This approach ingeniously replaces the traditional discriminator component of a Generative Adversarial Network (GAN) with a specifically designed cost function. The model's effectiveness is subsequently assessed by comparing it with established classical methods, specifically the first-order autoregressive scheme (AR1) and the Hurst-Kolmogorov (HK) model. The originality lies in the use of a single Convolutional Neural Network (CNN) as a generator for time series data, replacing the discriminator component of a traditional GAN with a specifically designed cost function.

 

 However, it requires major revisions due to inherent limitations in the traditional models selected for comparison, which may hinder a comprehensive evaluation of the proposed method's true capabilities.

 

 B. Major observations.

  1.  A key proposal to enhance this research is the utilization of higher-order ARMA(p, q) models for both p and q. This approach would enable the capture of more complex patterns and dynamics within the data, which might be overlooked by a simpler AR(1) model. The current comparison against an AR(1) model limits the scope of captured dynamics to "Markovian processes," potentially missing important longer-range dependencies mentioned later in the discussion.
  2.  While the convolutional neural network (CNN) used as a generator, despite its simpler stochastic scheme, demonstrates superiority over the classical alternative models presented (namely AR(1) and HK models), the choice of these classical alternatives is wrong. For a more robust comparison and to achieve a better fit to the data, the classical alternative models should have included higher-order ARMA(p, q) models, as previously suggested, or even Autoregressive Fractionally Integrated Moving Average (ARFIMA) models. The text itself highlights the limitations of AR(1) in capturing long-range dependence and the potential for HK models to lose important dynamics. Employing higher-order ARMA or ARFIMA models (https://blog.quantinsti.com/arfima-model/) would significantly enhance the predictive power and reliability of the alternative models, providing a more rigorous benchmark for the CNN generator.
  3.  In the context of the AR(1) and HK models, higher orders for p and q in ARMA models can more effectively explain the serial dependence structure of time series data. The manuscript acknowledges issues like long-range dependence and persistence up to 1350 time units when describing the CNN model. By incorporating models capable of capturing such long-range dependence, such as higher-order ARMA or ARFIMA, the research could be more rigorous.

 

C. Conclusion.

The article should be rewritten with these suggestions in mind, as I consider the chosen comparisons to be wrong and not sufficiently challenging for a thorough evaluation of the proposed CNN-based stochastic scheme.

Best regards,

The reviewer

Back to TopTop