Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript presents a new k-nearest neighbors (kNN)–based method for sub-seasonal precipitation forecasting at weather stations. Tests were conducted at five Oklahoma stations with long-term observation records.

While the method is useful in the operational practice, the manuscript in its current form has weaknesses in methodological clarity and physical interpretation.

 

#1

The methodology section provides a conceptual flow (Figures 1–3), but several crucial aspects are unclear:

• The search domain of GFVs (1–180 spans × 1–20 days) is vast; the paper should specify whether all combinations were tested exhaustively and how computational efficiency was managed.
• It is mentioned that 45 years were used for GFV selection and 5 years for testing, but it is unclear whether cross-validation or bootstrapping was employed to prevent overfitting.

 

#2

• Reliability diagrams (Figure 9, p. 14) show large deviations from the 1:1 line at high precipitation bins, suggesting systematic bias and low calibration — this limitation should be explicitly discussed.

 

#3

The paper mainly focuses on algorithmic description, with little discussion of physical mechanisms behind the results:

• Why does GEM perform differently at humid (Idabel) versus arid (Hooker) stations?
• How do seasonal patterns (Figures 6–8) relate to known drivers of Oklahoma rainfall variability (e.g., ENSO, Gulf moisture flux)?
• Could the apparent overforecasting in wet months reflect the influence of normalization on variance?

A physically grounded interpretation would greatly strengthen the paper’s scientific contribution beyond pure statistical forecasting.

 

#4

The station-based investigation may largely miss extreme precipitation events, i.e., localized convective rainfall events, which fundamentally affects the predictability.

Such characteristics of localized heavy rainfall are summarized in the following studies: 

https://doi.org/10.1016/j.atmosres.2024.107802

https://doi.org/10.1007/s13143-019-00128-7

To address this issue, the proposed framework may be applied to the gridded-rainfall datasets.

I recommend that the authors discuss such future directions in the discussion section.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The article introduces a novel k-nearest neighbors (kNN) method for sub-seasonal precipitation forecasting, which predicts by identifying numerous monthly temporal patterns. Results show that this new method consistently outperforms climatological forecasts and existing kNN iterations across multiple metrics (RMSE, mean relative error, Nash-Sutcliffe coefficients) at five stations in Oklahoma. The topic is highly relevant for agricultural and water resource management, and the empirical results are encouraging. However, the current presentation lacks sufficient depth in the methodological novelty and the physical explanation of the chosen features. The reliance on a limited geographic scope and the absence of comparisons with state-of-the-art non-kNN models limit the potential impact and generalizability of the conclusions. Therefore, the article requires significant revisions to strengthen methodological details and expand the scope of its validation before it can be recommended for publication.

Main Comments:

1. The article does not clearly define how many days are grouped, nor does it provide a solid statistical or physical rationale for why this specific temporal averaging/grouping scheme is superior for sub-seasonal precipitation forecasting compared to standard daily or weekly aggregates. This is the core research contribution and requires rigorous detail.

2. All experiments and validations in the article use data from only five weather stations in Oklahoma. Although the results are favorable within the study area, the applicability of this "new kNN" feature engineering method to different climate regions (such as tropical, maritime, and mountainous areas) has not been tested. Global precipitation patterns are highly heterogeneous. To enhance the manuscript’s impact, authors should expand the validation scope to include at least two stations within different climate classifications (e.g., coastal areas and arid western regions).

3. Significant progress in sub-seasonal forecasting has been made with machine learning methods beyond kNN, such as Random Forest (RF), Support Vector Machine (SVM), and various deep learning models (e.g., LSTM). Without comparisons to these modern high-performance benchmarks, it is difficult to assess its competitive advantage in the field. It is recommended that authors include comparisons with at least one representative state-of-the-art machine learning model (e.g., Random Forest or a simple neural network architecture) commonly used for time series or precipitation forecasting, using the same feature set and validation sites. This is crucial for establishing the true contribution of the method.

4. The article mentions using "numerous monthly temporal patterns" for prediction but does not clearly list external predictors or feature variables (e.g., ENSO index, sea surface temperature, atmospheric circulation fields), other than the internal historical precipitation data. The article needs to clarify what constitutes the complete feature vector used by the kNN algorithm. It is suggested to explicitly list and describe all predictor variables (features) used in the kNN search space. If only historical precipitation data is used, it needs to be clearly stated, and its limitations in capturing teleconnections should be addressed in the discussion.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors This work certainly deserves publication. The only changes needed are to make the work more relevant to the topic of the journal in which it is published and to improve its readability. In particular, it would be helpful to analyse your model's performance in more detail regarding training and forecasting. I'm sure you've determined the "best" possible models, but it would be very important to further explain the methodologies that ensure you've reached the optimal model. In particular, it would be important to understand how much past data is needed to develop a good model, and how long into the future the forecast remains reliable before the model needs to be retrained based on the arrival of new real-world data. This discussion is very important for practical purposes and would certainly enhance your already excellent work. Regarding the presentation of the results currently in the paper, I would recommend improving some figures. Meanwhile, it would be interesting to include a geographical and physical description of the area under investigation. The only figure is Figure 4, which is essentially a map of the isohyets. How is it possible that they are so jagged? Why is there a white region? It would be helpful to include North and a length scale (especially for latitude, the scale is hard to understand...). Figures 6, 7, and 8 are difficult to understand. I understand that you want to maintain the same scale on both axes, but I'm having trouble making it out. Perhaps it would be better to use smaller, but filled, points. Perhaps the space (now left blank) between one inset and the next could be better utilised. For Figure 8, I suggest you use slightly longer y-axes than the current ones. There's room on the page. Here, too, a better choice of symbols and colours could be of help. Regarding Figure 9, it is referred to as a "reliability plot," but it provides the same information as Figures 6 and 7, which are, instead, scatter plots. Why is there this difference in the name, which could confuse the reader? Perhaps it's because you connected the points with line segments to show the trend of the forecast over time? It would be crucial to explain (even in the caption) what the line segments in the figure represent, if they make sense. Otherwise, they should be removed. For this figure, too, I would keep the same scales on the axes, so that the exact forecast line is at 45 degrees.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes a novel GEM-kNN method to improve monthly precipitation prediction. The method is innovative, the experimental design is rigorous, and the result analysis is detailed. The manuscript is well structured, clearly written, and has publication potential. However, there is still room for improvement in the methodological explanation, experimental comparisons, and discussion of results.

1. The paper should more clearly explain how GEM selects the “optimal” configuration from multiple GFVs. Is the selection based solely on RMSE? Has the risk of overfitting been considered?

2. Details of GFV construction: It is recommended to further explain why “group averaging” was chosen as the feature, rather than other aggregation methods (e.g., maximum, variance).

3. The test data used in the paper cover the period 2004–2008. It is recommended to clarify whether more recent data were used for validation, in order to enhance the timeliness and generalization capability of the method.

4. The SVM feature construction is consistent with SotA approaches, but the kernel function, parameter settings, and other details are not described. These should be added.

5. The manuscript mentions that GEM still performs inadequately in predicting extreme precipitation. It is recommended to analyze the underlying reasons and propose possible directions for improvement.

6. For some stations, the p-values are not significant (e.g., Chandler and Idabel). It is recommended to discuss what these results imply for the broader applicability of the method.

Author Response

Please see the attachment.

Author Response File: Author Response.docx