Ensemble Learning for Spatial Modeling of Icing Fields from Multi-Source Remote Sensing Data

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a real-time icing grid field construction framework based on multi-source remote sensing data fusion and applies ensemble learning models to predict the spatial distribution of icing thickness. The research topic is of practical significance, the dataset is diverse, and the methodology is comprehensive. However, there are substantial flaws in the methodological framework, model evaluation strategy, result interpretation, and the analysis of model generalizability. Major revisions are required for the manuscript to meet the standards for publication.

Major comments:

1. In Section 3 (Results), only two references are cited. The manuscript does not sufficiently compare or discuss its findings in the context of existing studies on icing grid modeling, such as classical physical models, approaches combining numerical weather prediction (NWP), or alternative machine learning techniques. This omission weakens the academic depth of the paper and makes the work appear more like a technical report than an original research contribution.
2. The manuscript shows that both machine learning models and the Empirical Bayesian Kriging Interpolation (EBKI) method perform poorly on the validation set but significantly better on the testing set. This unusual trend raises questions about the data partitioning strategy. For example, splitting the dataset chronologically (first 70% of time steps for training, last 30% for testing) may lead to uneven distribution of cold wave events. If the validation set contains more extreme weather conditions while the testing set reflects relatively stable climates, the model performance comparison could be misleading.
3. While the research objective is to develop a "real-time icing grid field," the model is trained and tested exclusively on winter 2023 data, with intensive observations concentrated in a narrow window from late January to late February. There is no validation across different years or seasons, making it difficult to assess the model's robustness under long-term operational conditions or during climatic anomalies. Moreover, the current methodology essentially performs "real-time interpolation" rather than "real-time forecasting".
4. While the use of metrics such as the correlation coefficient (R), root mean square error (RMSE), and critical success index (CSI) helps quantify the statistical performance of the models, the paper does not adequately relate these differences to the operational needs of power grid maintenance. The threshold of 2 mm used for CSI calculation is not supported by engineering rationale. Moreover, the performance of the models under alternative thresholds is not evaluated or quantified, which may affect the generalizability and robustness of the conclusions.
5. The paper highlights the challenge of interpretability in machine learning models and applies the SHAP algorithm for model explanation. However, the physical interpretation of key meteorological features—such as relative humidity, wind speed, temperature, and precipitation—is not discussed in sufficient depth with respect to their role in the icing process. The assertion that there is “no interaction” among features based solely on SHAP values lacks statistical testing and should be supported by additional analysis.
6. The study shows a significant rise in RMSE during four cold wave events (Fig. 7), but the potential causes are not systematically analyzed. For example, do rapid temperature drops and sudden increases in wind speed during these events result in coupled meteorological extremes that fall outside the distribution seen in the training set?
7. The study relies on proprietary icing monitoring data from the China Southern Power Grid, which external researchers cannot access. This raises concerns about the reproducibility of the experiments. Furthermore, the architecture of the stacking model lacks transparency. The manuscript does not provide sufficient details on the construction and configuration of RF, XGBoost, LightGBM, or the stacking ensemble itself, which impedes reproducibility and independent validation of the findings.
8. The study focuses on icing prediction for power transmission lines but does not consider how material characteristics (e.g., conductor type, surface coatings) may influence the icing process.

Author Response

please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This study is very interesting and meaningful. The authors proposed a machine learning-based approach for constructing real-time icing grid fields using 1339 online terminal-monitoring datasets provided by the China Southern Power Grid Research Institute Co., Ltd. during the winter of 2023. Five machine learning algorithms were applied and evaluated, results indicated that the stacking model performed best. These findings offer a new approach for generating real-time icing grid fields.

Specific comments:

1. Line 134: Please change “Changes over time in the number of pylons (icing value > 0). The gray shaded areas labeled 'first,' 'second,' 'third,' and 'forth' represent the four cold wave events.” to “(c) Changes over time in the number of pylons (icing value > 0). The gray shaded areas labeled 'first,' 'second,' 'third,' and 'forth' represent the four cold wave events.”
2. In this study, only 1339 online terminal-monitoring datasets, covering a total of 413 time points during the winter of 2023, were used for training and validation. The sample size is relatively small for a data-driven approach. It would be better if the authors expanded the sample size.
3. Table 1 shows that the improvements in error evaluation metrics on the testing set are better than those on the validation set. Does this imply that all five machine learning models suffer from overfitting?
4. In section 4, figure 9 shows that the feature importance ranking of “lon” is higher than that of “lat”, what might account for this difference?

Comments on the Quality of English Language

The English language quality of this manuscript is generally clear and technically sound, but there are minor grammatical and stylistic issues that could be refined for improved readability and academic rigor.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed all my concerns, and the manuscript can be accepted now.