Feature-Enhanced Erroneous Outlier Detection in Hydrological Time Series Using Ensemble Methods

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The main focus of this study is data quality in hydrologic time series, specifically the identification of outliers to generate clean datasets for modeling and operational flood prediction. The authors identify existing issues in hydrologic data quality control and address these gaps by proposing and systematically evaluating a framework that employs multiple outlier detection algorithms applied to daily hydrologic datasets from five gaging stations. Specifically, they compare raw data in vanilla implementations with engineered variants that incorporate specifics of hydrologic data, and they validate the approach through the injection of synthetic outliers.

The authors undertook substantial effort to prepare the datasets, select statistical, machine-learning, and ensemble-based detection methods, and design the evaluation strategy. All selected algorithms, the experimental setup, and the evaluation procedures are clearly described. The results are well presented and discussed, although minor improvements to some figures are suggested (see technical details below).

Overall, the results are clearly interpreted, and the study successfully achieves its stated objectives. The proposed framework, which emphasizes hydrologic features rather than relying solely on raw data, demonstrates improved outlier detection performance with clear potential to enhance operational modelling and flood forecasting. The paper is well written and well organized, and I recommend it for acceptance.

 

Technical notes:

1. Line 410: w_i should be w_i 
2. Line 448: 𝑂_𝑖 and 𝑂_𝑗 should be lower case o_𝑖 and o_𝑗, respectively
3. Line 469: q_α should be q_α 
4. Line 470: 𝑞_{0.05} should be without {}, 𝑞_0.05 
5. Line 434: You should mention meaning of AUC (Area under the Curve)
6. Figs 2, 4, 5, 6, 7, 8, and 10: Please increase the font of axes labels and some legends if possible.

Author Response

Thank you for your positive feedback and for recognising the value of our study. We greatly appreciate your constructive review and recommendation for publication after minor revisions. We have carefully addressed all the comments and suggestions provided, making the necessary revisions to enhance the quality and clarity of the manuscript. Where direct implementation was not feasible, we have provided detailed justifications. The following responses address each of the specific comments, with all changes marked in the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Methodology

1. (pp.20-21) There is a need to provide further clarification on the process of the semi synthetic validation process. Specifically, the rationale for selecting a 5% contamination rate, distributing as 50% point, 30% contextual, and 20% collective outliers should be more clearly justified, instead of just referencing an established practice.
2. (pp.2-3, pp.11) The study presents an extensive feature engineering framework with 19 features later reduced to six key indicators. However, the current paper mainly describes the conceptual role of these features. Visual comparisons showing time series before and after feature enhanced detection is needed to demonstrate how engineered features change separability between normal and erroneous data.
3. (pp.5-6) The paper evaluates 13 individual detection algorithms alongside three ensemble models, but it should clearly describe how the parameter settings used for the different methods were selected. Whether through previous studies or empirical adjustments. In addition, a summary of parameter settings and some discussion of parameter sensitivity is needed to further strengthen the comparability across methods.

Experimental Setup, Results and Discussion

1. (pp.9-21) The paper should explain why certain algorithms consistently perform better under specific hydrological conditions by interpreting the implications of low agreement among some models and by relating observed performance trends to underlying hydrological processes rather than attributing them solely to algorithmic behavior.
2. (pp.9-21) Although the ensemble strategies appear effective, the paper should more explicitly demonstrate the quantitative benefits of adaptive ensembles compared with fixed combination ensembles, for example by providing direct performance comparisons with conventional static ensembles.

Conclusion

1. (pp.21) The paper places significant portion of discussion and interpretive content to Sections 3.4 Overall Discussion and 3.5 Limitations and Future Directions, while the Conclusion section itself is comparatively brief and largely reiterates the Results section. As a result, important interpretative findings and key takeaways are scattered instead of being aggregated. The conclusion can be improved by incorporating the pivotal takeaway already discussed in sections 3.4 and 3.5 into a more integrated concluding statement.

Author Response

Thank you for your positive feedback and for recognising the value of our study. We greatly appreciate your constructive review and recommendation for publication after minor revisions. We have carefully addressed all the comments and suggestions provided, making the necessary revisions to enhance the quality and clarity of the manuscript. Where direct implementation was not feasible, we have provided detailed justifications. The following responses address each of the specific comments, with all changes marked in the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

water-4089788

The present manuscript “Feature-Enhanced Erroneous Outlier Detection in Hydrological Time Series Using Ensemble Methods”,  aims to develop a framework for detecting erroneous outliers in daily hydrological time series through univariate machine learning.The article is well-structured, with advanced methods, and the conclusions are innovative and have practical guiding value, especially in terms drought monitoring and flood forecasting systems.However, we still need to address some minor points and details before the article can be accepted，The detailed comments are as follows:

* Abstract

1. The section in the Introduction regarding erroneous outlier detection in the context of hydrology could be moved to the Abstract to highlight the innovation of the study.

* Introduction

2. Table 1 could be moved to the supplementary materials, and the discussion of erroneous outlier detection in the context of hydrology in the Introduction is overly redundant.

* Methodology

3. Several elements, such as Table 3 and the pseudocode algorithm, should be moved to the supplementary materials, as the Methodology section is excessively redundant.

* Experimental Setup, Results and Discussion

4. Based on an understanding of hydrological processes and temporal patterns, 19 features were derived. These features were then reduced to six key features through correlation analysis. Could this process be elaborated on in detail? It is recommended to include the detailed content in the supplementary materials to enhance the transparency of the study.
5. The self-explanatory nature of Figure 10 is somewhat inadequate. It is recommended to optimize the color scheme or add annotations to improve readability.

* Conclusion

6. The explanation for why methods dependent on temporal context underperform at stations with incomplete data histories is suggested to be supported by references or verified through simple analysis to enhance its persuasiveness.

Author Response

Thank you for your positive feedback and for recognising the value of our study. We greatly appreciate your constructive review and recommendation for publication after minor revisions. We have carefully addressed all the comments and suggestions provided, making the necessary revisions to enhance the quality and clarity of the manuscript. Where direct implementation was not feasible, we have provided detailed justifications. The following responses address each of the specific comments, with all changes marked in the revised manuscript.

Author Response File: Author Response.pdf