We thank the reviewer for the feedback, and provide some clarifications and responses below.

Line 146: 

A: The definition of damage comes from the Department of Social Welfare and Development that we cite in Section 3.1. The exact definition of the extent of damage is subject to interpretation of the assessors and the municipality. We assume that this definition does not vary between municipalities, but it is true that it may present a noise in the data.

Line 158:

A: Apologies for not catching this broken reference. It should be referring to Figure 1. We will fix it in the next version of the paper.

Lines 235-240: 

A: We chose the cutoff of feature correlation at absolute values larger than 0.99 to make sure that whole feature groups were not removed. We will clarify in the next version that we used the absolute value in the next version of the paper. The largest negative correlation between the remaining variables  is -0.85 between HAZ_dis_track_min and HAZ_v_max.  Further, we also experimented with the VIF method for removing additional features, but this decreased the performance of the model. Ultimately, the model we chose – XGBoost – is robust to multicollinearity, thus we chose to include more variables in order to let the model choose the most important ones, while still being able to have some performance gains through including the highly correlated variables. 

Lines 325-327:

A: In lines 325-327 we are speaking about the results shown in Table 4, on the average error, which measures the difference between estimated and actual damage values. This measure, overall, is very close to 0 for the first two municipality-based models (first two columns of the table), and becomes more negative for the grid-based models (third through fifth columns). Both, “Average Error” and “Root Square Mean Error” measure error, but whereas the first preserves the direction of the error, the second does not. Thus, our AE results are distinct from those measured using RMSE in Table 3.

Table 3:

A: It is true that the M-local model has the best performance on Bin 5 – that with the most damage. However, 2SG-Global+ is as we correctly state in line 319 the best performing model in bin 4. This bin is equally if not more important than bin 5. It is equally important as it also covers areas with damages larger than 10% which would contribute to the triggering of Anticipated Action. It is more important as the number of damaged areas in this bin is significantly larger, overall if bins 4 and 5 were merged the weighted error would be 15.18 for 2SG-Global+ vs 15.53 for M.local, showing the better performance of 2SG-Global+ in the high damage bins. In any case, in this study we are not necessarily looking for the best model per se, but it is seeking to achieve a good performance in models that use different sets of features – those that are more available globally, and which make the prediction at a more granular (grid) level. Further, we explore the performance of the models on the high-damage areas in section 4.3 Action Trigger Application, wherein we demonstrate the performance of the models in detecting the municipalities that were in most need of assistance. This experiment showed that the proposed grid-based model resulted in a better true positive, false positive, and false negative rates, which would have improved the use of limited anticipatory action resources. Finally, when we state in line 348 that “we now evaluate our best model (2SG-Global+)” we refer to the best global model we have introduced, not necessarily the best overall model. We will clarify that in subsequent versions of the paper.

We are glad to clarify the additional comments by the reviewer, and we will add these clarifications to the next version of the manuscript, if given the opportunity.

1. Line 141: 

A: We use 3726 0.1-degree cells that overlap with land. We will add this information to the next version of the paper.

2. Line 152-154: 

A: We defined the “low” threshold for both rainfall and windspeed by utilizing long-term average observation. This decision is rooted in the understanding that housing damage is unlikely to occur under conditions of below-average rainfall and wind speed. We will add this clarification in the next version of the paper.

3. Line 158:

A: Transformation weights are the proportion of the number of buildings in each grid and municipality intersection with respect to the total number of buildings in each municipality. The number of buildings within each grid in a given municipality and the total number of buildings in a municipality are determined by counting the number of building centroids from Google Open Buildings data and using shape-files for the borders of the municipalities obtained from the Humanitarian Data Exchange (HDX). We are trying to ascertain what proportion of the houses in each municipality should be assigned to the grid. For example, for a municipality wholly contained within a grid, all the houses in the municipality would be assigned to the grid and the weight would be 1. To convert the percentage of damaged houses from the municipality level to the grid level, the number of houses damaged and the total number of houses per municipality are multiplied by the weights to get the total number of houses and houses damaged per grid. The percentage of houses damaged are computed from the two values.

4. Line 159: 

A: Yes, OSM stands for OpenStreetMap. We will update this in the next version of the paper.

5. Line 163-164

A: The back-transformation weights follow the same principle as those of the transformation weights but in the opposite direction, ascertaining what proportion of the houses in each grid should be assigned to the municipality. They are the proportion of the number of buildings in each grid and municipality intersection with respect to those in each grid. Similarly, for a grid wholly contained within a municipality, all the houses damaged in the grid would be assigned to the municipality. 

6.  Line 385:

A: We apologize for the vagueness of the language here. Indeed it is the same measure of being damaged. This refers to 10% of all the houses falling into the NDRRMC category of "completely damaged" (so not inhabitable based on visible inspection). We will remove the “severely” from the next version of the paper.

7. 

A: Thank you for pointing out this awkward language. We will modify it in the next version of the paper. We will furthermore again revise the whole text of the manuscript for grammar errors and incomplete sentences. 

8. Discussion & Conclusions:

A: Thanks for the suggestion. We already hinted as this in the discussion when we state that 

“It should be noted that the grid-cell model can distribute some of its predictions to municipalities outside the municipalities of the target training data, with two consequences. It allows the uncovering of areas that might have been overlooked in the data collection (e.g., areas with mild damage that have not been reported but are helpful to train the model). However, it also might allocate damage to areas of minor importance (e.g., neighboring villages with very few houses).”

We will extend these sentences in the manuscript stating that ”Furthermore, this redistribution of data to the grid-level leads to a more fine-grained and uniform spatial data distribution. This augmentation of datapoints (as can be observed in Figure 2) quite possibly has a positive influence. Additionally, also the inclusion of socio-demographic features like the RWI index and terrain type has certainly helped in improving model performance.”

We thank the reviewer for this thoughtful comment.

The current impact-based forecasting model uses a storm surge susceptibility map, which is -we fully understand- a static map and hence a simplification as it does not cover the dynamics. We acknowledge the importance of and are aware of the developments in the area of storm surge modeling. We are in a project around Destination Earth to assess the use of GTSM, see https://stories.ecmwf.int/destine-digital-twins-to-anticipate-the-devastating-effects-of-flooding-in-coastal-areas/index.html. Future developments of the IBF Model might include hybrid modeling where the machine learning model exchanges the static storm surge map with a hydro-meteorological dynamic model. We speak about this briefly in the paper: “For example, for storm surge, dynamic models, such as the Global Tide and Surge Model, are being developed (Bloemendaal et al., 2019). Incorporating these dynamic models into an ML model might improve the secondary hazard features used in grid- and municipality-based models. The new technologies being developed as part of Digital Earth (Annoni et al., 2023) can play a role here.”  We would be happy to add further citations to the final version of the paper.

We would like to thank the reviewer for their time and very helpful comments. We provide a few clarifications below, and hope to have an opportunity to add these to the manuscript in the next version.

1. Thanks for pointing out the inconsistencies in our use of abbreviations. In the subsequent version of the paper we will proof-read it and standardize their use.

2. We will examine the use of “disaster” language and change it when it is not appropriate.

3. We use the statistics provided by the US agencies (NOAA & JTWC), which report 1-min sustained wind values, so no conversion is done. https://www.ncei.noaa.gov/sites/default/files/2021-07/IBTrACS_version4_Technical_Details.pdf The rest of the agencies report their data in 10-minute sustained values, but we do not use these.

4. Indeed, the typhoon is classified as a very strong typhoon by Japan Meteorological Agency and category 4-equivalent typhoon by JTWC. We are using the Saffir-Simpson scale here, despite the geography of the typhoons studied, and we will clarify this in the text. We will also clarify the use of Saffir-Simpson scale in the caption of Figure 3.