Dear Referee #1,

Thank you for taking the time to consider our work, and for the positive feedback and critiques provided in your comments. We have written a short reply below to continue the discussion while we await for the Discussion period to end.

Sincerely,

The authors

Comment 1. Operational validity and target definition

The manuscript is framed as an impact-based early-warning study, but some modelling choices make the prediction task closer to detecting or updating already emerging impacts. In particular, the model uses flood-related emergency calls from the previous 1–3 hours as predictors, which may strongly benefit cases where impacts have already started. The authors should distinguish more clearly between predicting the first occurrence of impacts and predicting the continuation of ongoing impacts. Additional experiments without previous-call features, or separate results for “before first call” and “after first call” situations, would make the operational contribution clearer. The daily rain filter also requires clarification: if it uses full-day information retrospectively, the evaluation may not represent a real-time forecasting setting. The definition of the positive class, positive rate, class weighing, and evaluation metrics should also be stated more explicitly.

Response:

The operational validity is a major focus of our work, and we thank the reviewer for raising this issue. Regarding the use of emergency call data from the previous 1-3 hours as predictors, we acknowledge how these features contribute significantly to the models' performance. To prove this, we actually conducted some ablation experiments in Section 6.2.3 where the models were trained ignoring previous call data, and observed a reduction in CSI for the high-density model of -11.6 ± 4.95, allowing us to identify these features as the most influential. We chose to maintain them in the final configuration as they would be available during a operational deployment, potentially benefiting situations with unfolding impacts.

To address the distinction between predicting the first occurrence of impacts versus their continuation, we conducted the study in Section 6.2.4 (Model performance depending on rain stage). By presenting the performance of our high-density model (which obtained the best results and most consistent explanations) across different rainfall stages, we believe the operational concerns of the reviewer are answered. More specifically, we identify a performance drop at the start of a rainfall event, which can serve as a proxy for "before the first call", compared to the subsequent stages. We argue that using a physical magnitude (i.e., rainfall) instead of the appearance of the first impacts results in a more consistent definition of what an "event" is, and we hope this argumentation correctly addresses the reviewer's concerns.

Regarding the real-time use of our methods, the retrospective daily rain filter was primarily used to accelerate training and mitigate class imbalance in our experiments. In an operational, real-time setting, the model can still be run every hour without issue, as even after filtering out entire days without significant precipitation, most of the dataset still consists of samples that did not register neither rain nor impacts. In the revised version, however, we will make this clearer as the referee suggests. Furthermore, we will also revise the definitions of the rest of the concepts proposed.

Comment 2. Evaluation design and comparability across temporal configurations

The walk-forward evaluation is reasonable in principle, but the current implementation makes comparisons across temporal configurations difficult to interpret. Although the authors control the number of positive samples, the validation and test periods differ across configurations. Thus, differences in performance may arise not only from model robustness or training history, but also from differences in the test data themselves, such as negative-sample counts, positive rates, rainfall regimes, event severity, seasonality, or spatial distribution of impacts. Because validation sets also differ, hyperparameter tuning and threshold selection are performed under different validation distributions. The results shown in Figure 7 also do not indicate a clear trend.

Response:

We acknowledge the reviewer's concerns regarding the temporal configurations employed. The irregular and sparse nature of impacts poses many challenges when evaluating these systems, and when choosing a particular temporal cross-validation scheme, one must also consider their related trade-offs.

We experimented with multiple alternatives before deciding on our approach. For instance, fixing temporal periods (e.g., specific months) resulted in highly inconsistent numbers of both positive and negative samples per set, leading to high variance in performance metrics. On the other hand, fixing the total number of samples (positive and negative) led to sets with drastically different positive-to-negative ratios, and varying temporal periods. Moreover, for both previous cases, we could not ensure that the validation and test sets contained any positive samples, which was especially common in the medium-to-low density municipalities.

In the end, we chose to fix the number of positive instances (impacts) per set. When comparing the different schemes, this configuration yielded the most stable results with the lowest standard deviation in model metrics, providing a more reliable setting for comparing performance and interpretability across configurations. While we acknowledge that this leads to different test periods and background distributions, we have found that stabilizing the positive class seems to be the most effective way to derive meaningful insights from such an imbalanced dataset. We will update the manuscript to provide a more detailed justification of why we chose this methodology, and will additionally address the lack of clear trends in Figure 7.

Comment 3. Interpretation, uncertainty, and presentation

The SHAP, ablation, and rainfall-stage analyses are useful, but their interpretation should be more cautious. The low-density model has very weak and unstable performance, so feature-

importance or sensitivity analyses for this group should not be used to support strong conclusions; similar caution is needed for the medium-density model. The authors should distinguish robust findings from the high-density model from exploratory findings in lower-density settings. In addition, SHAP and ablation results should not be interpreted causally, especially given the likely correlation among rainfall, radar, warning, and call-history predictors. The rainfall-stage analysis is interesting, but the claimed “warm-up” phase of the first hours needs clearer reasons. More generally, the manuscript would benefit from restructuring: methodological details in the Introduction should be reduced, the short Related Work section could be integrated or better aligned with the research gap, and several redundant or unclear sentences should be revised.

Response:

We appreciate these insightful notes regarding the interpretation of our explainability analyses and the overall structure of the manuscript. We fully agree that the performance of the medium- and, especially, low-density models is less stable. Consequently, we will revise the text to distinguish clearly between the more robust findings from the high-density model and the exploratory insights from the lower-density approaches, with the goal to temper the conclusions and avoiding over-interpretation. Moreover, we will clarify that the interpretation of the SHAP and ablation results represents feature contributions within the model, and not necessarily causal relationships.

Regarding the rainfall-stage analysis, we will expand our discussion on the mentioned "warm-up" phase to be more grounded, and additionally make changes to target the proposed restructuring (which we agree could be beneficial) and the rest of the technical corrections.