the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Aug 2026
Validating Remote-Sensing Measures of Natural Hazards: Granular-Level Links to Insured Loss during a Cyclone
Abstract. Remote-sensing products are widely used after disasters as indicators of incurred damage, yet it remains uncertain which mapped surface-disturbance remote-sensing signals are most informative about residential damages. This study examines the damage from the 2023 Cyclone Gabrielle in New Zealand by linking four publicly available remote-sensing layers—SAR-detected standing water, post-event wetness, soil- or silt-related disturbance, and inferred slope-related disturbance—to residential insurance claims from the public insurer at a fine spatial scale. We construct claim-rate and insurance payout outcomes and estimate cross-sectional models, investigating their association with the data from remote-sensing products. The two insurance outcomes capture recorded claim intensity relative to local building stock and the monetary intensity of insured loss. Slope-related disturbance is most strongly associated with claim occurrence and loss variation in hazard-positive rural areas. Wetness-related disturbance becomes the strongest predictor of loss severity once claims are observed. Standing-water and soil-related indicators provide smaller but more stable signals, and their composite index is positively associated with both claim rates and payouts. Urban areas show higher baseline loss levels, whereas rural areas show stronger marginal responses to additional physical disturbance. Our findings show that remote-sensing indicators are not interchangeable hazard proxies. Their value as proxies for disaster damage depends on the physical signal captured, the damage outcome measured, and the settlement context in which damage occurs.
- Preprint
(2635 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Review for egusphere-2026-3539', Anonymous Referee #1, 03 Sep 2026
-
AC1: 'Reply on RC1', Ke Wang, 26 Sep 2026
General comment
Comment 1
The study has a slightly inconsistent conceptualisation of how hazards and losses are linked. It often ignores the fact that losses are not a direct function of the hazard intensity but are also driven by the exposure and vulnerability of dwellings. I would recommend using the standard IPCC disaster risk definition of Risk = Hazard x Exposure x Vulnerability for consistency across the manuscript.
Response: Thank you for this comment. We agree that some of the current wording may imply a more direct relationship between mapped physical disturbance and realised insured loss than intended. The remote-sensing indicators used in this study capture different forms of mapped surface disturbance, whereas the insurance outcomes record realised insured residential losses. We will revise the relevant wording throughout the manuscript, including the passages identified by the reviewer, to distinguish these two types of measures more clearly and to avoid implying a direct or one-to-one relationship between them. This conceptual clarification does not alter the empirical specifications reported in the submitted manuscript; rather, it narrows and clarifies the interpretation of the estimated associations.
Comment 2
The approach using mesh blocks, which seem to represent administrative units, to normalise both the remote sensing and claims data is counterintuitive as it seems to throw away the most valuable properties of the two types of data which is spatial granularity. Lumping together the data in large unevenly sized geographical units removes the potential to study the heterogeneous effects that drive the variation in damages in different areas, often at much smaller scales than represented in the mesh blocks used in this study. Based on the description of the data and the robustness checks, my understanding is that you would be able to pair hazard indicators directly with the specific geolocation of a claim avoiding issues of spatial mismatches between hazard indicators and claims described in the discussion. I assume there is a reason why the authors have not done that, but the study currently lacks a justification for why they decided to aggregate data on administrative units that seem to have nothing to do with either how claims are reported or how the hazard indicators are constructed.
Response: Thank you for this helpful comment. We use 2023 meshblocks as a common small-area spatial framework in which we can construct the area-level insurance outcomes, relate claims and payouts to the mapped LINZ building-outline count (Land Information New Zealand, 2023), measure the affected-area share of each remote-sensing indicator, and link these measures consistently to the urban–rural classification used in our heterogeneity analysis. A claim-location analysis would answer a different spatial question and would not, by itself, reproduce these area-level measures or provide the same basis for comparing mapped disturbance–loss relationships across settlement contexts.
At the same time, we agree that aggregation may introduce spatial mismatch, particularly where mapped disturbance and claim-generating locations occur in different parts of the same meshblock. We therefore propose a complementary finer-scale analysis using the available geocoded claim locations. Specifically, we would construct 50 m and 100 m buffers around each claim location and record whether each buffer intersects the footprint of each remote-sensing indicator. These distances would be treated as sensitivity choices rather than as validated optimal radii for our study. Gao et al. (2025) tested buffers of 50, 100, 150, and 200 m around enterprise insurance-claim locations in a flood-modelling application and found that 50–100 m buffers provided a practical balance in that specific setting, while also cautioning that larger buffers could inflate apparent correspondence.
The two analyses therefore serve different purposes. The meshblock framework is used to examine area-level variation in mapped disturbance and insured loss, including urban–rural heterogeneity, whereas the buffer analysis would provide a local spatial-correspondence check around observed claim locations. We would therefore treat the latter as a complementary analysis rather than as a replacement for the main meshblock framework.
References:
Gao, D., Yin, J., Yang, Y., Mai, J., Lu, Y., Liang, P., Wang, W., Yuan, X., & Yu, D. (2025). Validating typhoon-induced rainstorm flood inundation modelling with insurance claims. Journal of Hydrology, 657, 133178. https://doi.org/10.1016/j.jhydrol.2025.133178
Land Information New Zealand. (2023). NZ Building Outlines (2023 version) [Data set]. LINZ Data Service. https://data.linz.govt.nz/layer/101290-nz-building-outlines/
Comment 3
If I interpret footnote 1 on page 5 correctly, then the claims data used in the study exclude any flood damages to residential dwellings. I find this confusing as the study uses at least two flood-related hazard indicators and previous studies on cyclones have shown that flood damages to dwellings are one of the main drivers of total damages during cyclones or hurricanes. If the claims data actually do not contain information about flood losses to private dwellings, the authors would need to explain how to interpret the results if presumably a large part of the overall damage is not included in the data. If the claims data do contain flood damages to dwellings, it would be good to rephrase footnote 1 to make this point clear.
Response: Thank you for raising this point. The distinction between the types of flood-related losses covered by EQCover is important for interpreting our insurance outcomes and deserves a fuller explanation.
Under the EQCover arrangements applicable to the North Island severe-weather events, direct flood damage to the home itself (the building) is covered by the homeowner’s private insurer rather than by EQCover. EQCover does, however, cover flood- and storm-related forms of damage to insured residential land, that is the land under and around the house and any accessway. This includes debris and silt inundation from floodwater or high winds and land scour or evacuation caused by overland water flow. Any landslip damage resulting from the storm is covered by EQCover for both the home itself and the insured residential land. EQCover for the building is capped at NZ$ 300,000, and for the residential land the cap is the assessed value of the land. Any damage above the cap is covered by private insurance. Insured residential land primarily includes the land under and up to eight metres around the home or insured outbuildings, the land under the main accessway within 60 metres of the home, and certain associated land structures, such as supporting walls (Toka Tū Ake EQC, 2023).
This coverage structure is relevant to the interpretation of the insurance outcomes and the flood- and slope-related remote-sensing indicators, but it does not establish a one-to-one correspondence between a mapped disturbance and an individual claim. S-Flood represents a time-specific standing-water signal, while D-Wet represents a post-event wetness-related surface-change signal. These indicators may identify surface conditions that overlap spatially with forms of insured residential-land damage, such as silt or debris inundation and land scour. Similarly, D-Slope is an inferred slope-related disturbance indicator that may be spatially associated with locations of landslip-related loss; it is not a direct or independently validated landslide inventory.
It is also important to distinguish the scope of losses covered from the institutional reach of the insurance arrangement. EQCover is New Zealand government natural-hazard insurance and is automatically attached to every private home-insurance policy that includes fire cover (all do). Estimates suggest that more than 95% of all residential buildings are insured under this private-public partnership system (Nguyen et al., 2020), and because the public insurer is responsible for the first tranche of damages, it is likely that the EQC recorded claims provide a good representation of overall scope of losses (even if not the overall monetary value of these losses).
Accordingly, we interpret our outcomes as claim and payout measures recorded in the NHC/EQCover data, rather than as a complete measure of all residential damage caused by Cyclone Gabrielle. The estimated associations for the flood-related indicators should therefore be interpreted as their correspondence with the insured losses recorded in these data, rather than with the full extent of direct flood damage to dwellings. We will correct footnote 1 by replacing “six metres” with “eight metres” and expand the explanation of the coverage arrangements and their implications for interpreting the results.
References:
Toka Tū Ake EQC. (2022, September 30). EQCover building cap begins doubling for homeowners from 1 October 2022.https://www.naturalhazards.govt.nz/news/eqcover-building-cap-begins-doubling-for-homeowners-from-1-october-2022/
Toka Tū Ake EQC. (2023, March 3). EQCover community information: North Island severe weather. https://www.naturalhazards.govt.nz/our-publications/eqcover-community-information-cyclone-gabrielle/
Comment 4
I think LogPayout should be normalised by the replacement value of buildings in an area to avoid losses being primarily a function of how expensive it is to fix a building. Averaging payouts over a larger geographical area can potentially be seen as an alternative to normalising payouts per claim, but comes with the limitations mentioned under point 2).
Response: Thank you for this suggestion. We understand the motivation for normalising monetary losses by the replacement value of the exposed building stock. If both the numerator and denominator referred to the same building assets, such a measure could provide a loss ratio relative to asset value. However, the payout outcome available in this study does not allow us to construct that measure consistently.
First, the NHC data extract available to us does not include property-level building replacement values. Second, the EQCover arrangements applicable to the North Island severe-weather events distinguish between damage to the home and damage to insured residential land. EQCover may cover landslip damage to both the home and insured residential land, while debris and silt inundation and land scour are covered for insured residential land only (Toka Tū Ake EQC, 2023). To the extent that the payout variable includes payments associated with insured residential land, dividing total payouts by building replacement value alone would introduce a numerator–denominator mismatch: part of the numerator could relate to land, whereas the denominator would value only the building.
We therefore retain the current definition of LogPayout, but will describe it more precisely. LogPayout is defined as the natural logarithm of one plus total NHC payouts in a meshblock divided by the mapped LINZ building-outline count. It is a building-outline-normalised log payout measure; it is not a property-value-normalised loss ratio or the average payout per affected building. We will also state explicitly that the measure does not adjust for differences in building values and repair costs across locations. These differences may therefore remain in the outcome and constitute a limitation of its interpretation.
References:
Toka Tū Ake EQC. (2023, March 3). EQCover community information: North Island severe weather. https://www.naturalhazards.govt.nz/our-publications/eqcover-community-information-cyclone-gabrielle/
Comment 5
Overall, the regression results are well described and seem robust, but I would suggest to also discuss the fact that the overall explained variance in all presented model specifications is quite small (Rsq < 0.1). This is very common when studying losses and damages from natural hazards, as the processes are complex and cannot be easily captured by individual hazard metrics, but if the declared goal of the paper is to validate remote sensing measures of natural hazards, this seems an important point to discuss. The results suggest that while the chosen hazard indicators do explain some of the variability in claim occurrence and payout, the unexplained variability is very high.
Response: Thank you for this suggestion. We agree that model fit is important for interpreting the results and should be discussed more explicitly. For precision, however, the reported fit statistics are not uniformly below 0.1. Across Tables 3 and 4, 14 of the 16 baseline OLS specifications have R² values below 0.1. The exceptions are the S-Flood ClaimRate specification in Table 4, with an R² of 0.1881, and the D-Soil LogPayout specification in Table 4, with an R² of 0.1136. Several robustness specifications also report R² values above 0.1. We will therefore discuss model fit by specification rather than characterising all models as having R² values below 0.1. We will also distinguish ordinary R² from the pseudo-R² measures reported for nonlinear models.
The R² values refer to the fit of the full specifications, which include the remote-sensing indicator, the urban indicator, and their interaction. They should not be interpreted as the proportion of outcome variation explained uniquely by the remote-sensing indicator. More generally, statistical significance does not imply high predictive accuracy or a comprehensive model of residential loss. The analysis is designed to estimate conditional associations between mapped physical-disturbance indicators and recorded insurance outcomes, rather than to produce a complete predictive model of disaster losses.
We will revise the Discussion accordingly. We will state that model fit is modest in many of the baseline specifications and varies across indicators, outcomes, samples, and model forms. The remaining variation may reflect measurement error and factors not fully represented in the specifications, potentially including differences in exposure, vulnerability, building characteristics, insurance coverage, and other local conditions. We will therefore interpret the results as evidence of statistically detectable but incomplete associations. This qualification is consistent with the paper’s conclusion that the four remote-sensing indicators exhibit heterogeneous and non-interchangeable relationships with the recorded insurance outcomes.
Specific comment
Comment L37ff
‘This distinction matters for loss analysis because insurance claims better reflect realised damage rather than meteorological intensity alone. Remote-sensing indicators can therefore provide a more direct spatial link between physical disturbance and observed residential losses and damages.’ These two sentences confuse hazard, risk and vulnerability. Both the meteorological intensity and the used remote sensing indicators are just different ways to measure hazard intensity. Losses are a function of properties exposed to the measured hazard and how vulnerable these properties are to the hazard.
Response: Thank you for this clarification. We agree that the current wording does not sufficiently distinguish between measures of hazard-related physical conditions and observed insured-loss outcomes. Conventional meteorological variables and the remote-sensing indicators used in this study describe different aspects of the event, but neither directly measures property damage or insured loss. The relationship between these measures and recorded insurance outcomes may also depend on the exposure and vulnerability of affected assets.
Our intention was to distinguish conventional meteorological measures from spatially explicit remote-sensing measures of surface conditions and disturbance, rather than to imply that the latter provide a direct measure of loss or damage. We will therefore revise these sentences to remove that implication and to make the conceptual distinction identified by the reviewer explicit.
We propose to revise the passage as follows:
“Conventional meteorological variables, such as rainfall and wind speed, characterise atmospheric forcing associated with cyclone impacts. The remote-sensing indicators used in this study provide instead spatially explicit information on mapped surface conditions or disturbances: time-specific standing water, post-event wetness, soil- or silt-related disturbance, and inferred slope-related disturbance. These indicators do not directly measure property damage or insured loss. We compare their meshblock-level affected-area shares with recorded NHC claim and payout outcomes to examine their statistical associations, recognising that the observed insurance outcomes may also reflect differences in exposure, vulnerability, insurance coverage, and other local factors.”
Comment L164
This “standing water” image has not been validated against ground observations of flood extent.’ That seems like a useful thing to do as I would assume that standing water can be the result of various processes including in-situ rainfall which is different from standing flood water from rivers. Is there a reason why the authors have not done that?
Response: Thank you for raising this point. We agree that the standing-water classification does not identify the hydrological process responsible for the detected water. We therefore do not interpret S-Flood as a validated map of riverine flooding or as a process-specific flood footprint.
S-Flood is based on the LINZ Sentinel-1 standing-water layer derived from imagery acquired on 14 February 2023. In this study, it is used as a time-specific indicator of SAR-detected standing water. We did not conduct an independent ground validation because no contemporaneous ground-reference dataset suitable for comparison with the satellite acquisition was available in the materials used for this study. A direct validation would require reference observations sufficiently close to the acquisition time and with adequate spatial coverage for comparison with the mapped layer.
Accordingly, the purpose of S-Flood in this study is to represent a snapshot of detected standing water, rather than a complete or process-specific measure of flood extent. This also distinguishes it from the other three remote-sensing indicators, which were derived from changes between pre-event and post-event optical imagery. We compare the meshblock-level affected-area share of S-Flood with recorded NHC claim and payout outcomes to examine their statistical association; this comparison does not constitute a validation of the standing-water classification or identify the process responsible for the detected water.
We will revise the manuscript to make these limitations explicit and consistently describe S-Flood as a time-specific standing-water indicator rather than as a complete or validated flood-extent product.
Comment L188f
What is the rationale for normalising LogPayouts by the total number of buildings and not by the number of affected buildings in a mesh block? Normalising by the total number of buildings makes it difficult to distinguish between areas that have few buildings with high damages and areas with many buildings with low damages.
Response: Thank you for this helpful comment. We agree that normalising aggregate payout by the total mapped building-outline count and normalising it by the number of affected properties would produce measures with different interpretations. We will clarify the rationale and limitations of the denominator used in the submitted analysis.
The current variable is defined as the natural logarithm of one plus total NHC payouts in a meshblock divided by the number of LINZ building outlines in that meshblock. Its purpose is to scale aggregate recorded NHC payouts by a common measure of mapped built development and thereby reduce differences arising solely from meshblock size. However, the LINZ denominator is a count of mapped roof outlines, which may include houses, garages, and large sheds; it should therefore not be interpreted as an exact count of insured dwellings or affected properties.
Dividing total payout by the number of uniquely affected properties would answer a different question: it would provide a measure closer to average recorded payout conditional on a property having an associated claim. Such a measure could help distinguish a meshblock containing a small number of high-payout properties from one containing a larger number of lower-payout properties. The current LogPayout variable does not make that distinction.
The submitted analysis constructs BuildingCount as the number of LINZ building outlines and ClaimCount as the raw number of NHC residential claims in each meshblock. It does not construct a separate count of uniquely affected buildings or properties. Moreover, an NHC residential claim does not necessarily imply direct damage to a dwelling, because some covered losses may relate to insured residential land. We therefore do not treat the claim count as equivalent to the number of affected buildings.
We will revise the manuscript accordingly. In particular, we will replace the potentially misleading description of LogPayout as the “average paid NHC residential claim amount per building” with the more precise description “the log-transformed aggregate NHC payout normalised by the mapped LINZ building-outline count.” We will also state explicitly that this outcome does not distinguish whether a given aggregate payout arises from a small number of high-payout claims or a larger number of lower-payout claims.
Comment L288ff
It would be good to get at least a bit more detail about the remote sensing indicators used and how they were constructed. For example, for D-Wet, how was the threshold defined to distinguish between higher soil moisture levels following rainfall and higher soil moisture levels caused by flooding?
Response: Thank you for this helpful comment. We agree that the current manuscript provides only a brief description of how the Dragonfly remote-sensing indicators were constructed, and that additional methodological detail would make their interpretation clearer.
We would first like to clarify that we did not define a separate soil-moisture threshold to distinguish rainfall-induced wetness from flood-induced wetness in our empirical analysis. D-Wet is derived from the source Dragonfly product, where the corresponding source category is labelled “flood”; in this study, we relabel it as D-Wet or post-event wetness to distinguish it from the SAR-based S-Flood indicator. The submitted manuscript currently describes the Dragonfly products as being generated from Sentinel-2 imagery by comparing a pre-event NDVI composite (10 January–10 February 2023) with a post-event composite (19–21 February 2023), with large NDVI decreases subsequently classified into wetness, deposited-soil, and slope-related categories.
Accordingly, D-Wet should not be interpreted as a direct quantitative measure of soil moisture or as a flood-extent layer defined by a threshold chosen by us. Rather, it is a source-defined post-event wetness/flood-related surface-change class. This is also why the submitted manuscript interprets D-Wet more cautiously as a “wetness-related surface signal rather than the full flood extent.”
We agree, however, that the present description could more clearly distinguish the classifications defined in the source product from the subsequent processing undertaken in this study. Some implementation details are documented in an unpublished technical report supplied by the data provider. Because this report is not publicly available and was shared on a non-circulation basis, we are unable to reproduce or disclose its detailed implementation information. Nevertheless, we will revise the data section to provide the clearest high-level account possible of the product-generation workflow and the basis of the source classifications using the information available to us. We will also distinguish explicitly between the categories and processing steps originating from the source product and the subsequent relabelling, spatial aggregation, and analytical thresholds introduced in our study. Although the non-public status of the report limits the level of methodological detail that can be provided, these revisions will clarify the distinction raised by the reviewer and explain how the source products were interpreted and processed in our analysis.
This clarification will also sharpen the distinction between the two water-related indicators used in the study. S-Flood is a time-specific SAR-detected standing-water snapshot from 14 February 2023, whereas D-Wet is a post-event wetness-related surface-change signal derived from the Dragonfly product. Neither indicator, as used in this study, identifies the specific hydrological process responsible for the mapped signal. We will revise the manuscript so that these different measurement concepts and their respective limitations are more transparent.
Comment L423
‘This sample’ repetition
Response: Thank you for noting this. We will remove the duplicated phrase “This sample” at L423.
Citation: https://doi.org/10.5194/egusphere-2026-3539-AC1
-
AC1: 'Reply on RC1', Ke Wang, 26 Sep 2026
-
RC2: 'Comment on egusphere-2026-3539', Anonymous Referee #2, 16 Sep 2026
Validating remote-sensing measures of natural hazards: Granular-level links to insured loss during a cyclone
This paper examines the relationship between residential property insurance claims—specifically the proportion of claim counts and payout ratios—and remote sensing indicators in regions affected by the 2023 cyclone in New Zealand. As remote sensing indicators, the study employs flood extent estimated from SAR imagery (S-flood), alongside three indicators derived from optical sensors (D-wet, D-soil, and D-slope). Through statistical analysis, the paper reveals that both the claim count ratio and the payout ratio show relatively higher correlations with S-flood and D-soil, and that the performance of these indicators varies between urban and suburban/rural areas. Based on these findings, the paper concludes that estimating damage-related metrics from remote sensing indicators is still challenging.
The use of insurance loss data is invaluable for evaluating and estimating damage scale, and the study deserves credit for utilizing such a precious dataset. However, significant flaws remain regarding the remote sensing indicators selected. While S-flood measures inundated areas using SAR imagery—a metric clearly relevant to cyclone damage—the other three indicators are calculated from pre- and post-disaster changes in NDVI. Because these NDVI-based indicators inherently include seasonal vegetation fluctuations, they lack a direct connection to actual disaster damage. Performing statistical analyses with metrics that have such weak physical ties to actual losses makes it virtually impossible to infer meaningful relationships with monetary damage, leaving the validity of the selected remote sensing indicators questionable.
Furthermore, NDVI is primarily an index used to assess the condition of vegetation and green spaces. In contrast, residential properties (buildings and immediate lots) subject to insurance claims are predominantly non-vegetated, built environments. Asserting that a drop in NDVI in surrounding farmland or woodland directly correlates with physical destruction or inundation of residential structures represents a major leap in physical logic.
If the authors intend to conclude that estimating damage from remote sensing indicators is difficult, it is essential to explore indicators that directly represent cyclone-induced physical damage. A study that skips the critical step of evaluating indicator validity and simply fits basic statistical models remains insufficient. Linking data that fails to separate baseline background noise from true disaster damage directly to insurance claims yields extremely low academic validity and utility.
For these reasons, the reviewer concludes that the paper suffers from flawed methodological premises regarding the physical and disaster-related validity of the input data. As the reliability of its conclusions cannot be guaranteed, this manuscript is recommended for Reject.
Citation: https://doi.org/10.5194/egusphere-2026-3539-RC2 -
AC2: 'Reply on RC2', Ke Wang, 26 Sep 2026
Comment 1
This paper examines the relationship between residential property insurance claims—specifically the proportion of claim counts and payout ratios—and remote sensing indicators in regions affected by the 2023 cyclone in New Zealand. As remote sensing indicators, the study employs flood extent estimated from SAR imagery (S-flood), alongside three indicators derived from optical sensors (D-wet, D-soil, and D-slope). Through statistical analysis, the paper reveals that both the claim count ratio and the payout ratio show relatively higher correlations with S-flood and D-soil, and that the performance of these indicators varies between urban and suburban/rural areas. Based on these findings, the paper concludes that estimating damage-related metrics from remote sensing indicators is still challenging.
The use of insurance loss data is invaluable for evaluating and estimating damage scale, and the study deserves credit for utilizing such a precious dataset. However, significant flaws remain regarding the remote sensing indicators selected. While S-flood measures inundated areas using SAR imagery—a metric clearly relevant to cyclone damage—the other three indicators are calculated from pre- and post-disaster changes in NDVI. Because these NDVI-based indicators inherently include seasonal vegetation fluctuations, they lack a direct connection to actual disaster damage. Performing statistical analyses with metrics that have such weak physical ties to actual losses makes it virtually impossible to infer meaningful relationships with monetary damage, leaving the validity of the selected remote sensing indicators questionable.
Response: Thank you for raising this concern. We agree that NDVI-based change detection can, in principle, be affected by seasonal vegetation dynamics and other background changes. We also recognise that our original description of the three Dragonfly products was brief. In the submitted manuscript, D-Wet, D-Soil, and D-Slope were described based on the source-product methodological information available to us at the time. That information describes the use of Sentinel-2 imagery before and after Cyclone Gabrielle, with NDVI-based change detection used to identify post-event wetness/flood-related surface change, deposited silt or soil, and slope-related change.
We would, however, like to clarify that these products are not generic NDVI measures constructed for our study. They are event-specific mapped classes produced by Dragonfly Data Science for Cyclone Gabrielle. Their intended use was to identify and map different categories of event-related surface change, rather than simply to report raw vegetation-index change. In our study, we use the source categories labelled “flood”, “silt”, and “slip”, relabel them as D-Wet, D-Soil, and D-Slope for interpretation, and aggregate their mapped areas to the meshblock level. We did not construct the source classifications using the insurance outcomes.
Following the reviewer’s comment, we contacted Dragonfly directly to obtain further clarification on the underlying methodology. The additional methodological information indicates that, for the non-flood impact classes represented in our study by D-Soil and D-Slope, the final products were not generated by treating a simple, unadjusted pre-event/post-event NDVI difference as the impact measure. Candidate pixel-level NDVI changes were identified relative to the mean and standard deviation of NDVI behaviour over the preceding year. This step was intended to screen candidate changes against expected temporal background variability rather than treating every observed NDVI change as an event impact.
The candidate changes were subsequently subjected to additional supervised classification. The classification model used 500 manually labelled impact and no-impact locations identified from aerial imagery, together with terrain information, to further filter and classify the detected changes (Dragonfly Data Science, personal communication, 2026).
Dragonfly also provided us with an unpublished draft technical report that documents the product-generation methodology in greater detail. We now cite this document as an unpublished technical report and use it to support the general methodological account provided above. Because the report was shared on a non-circulation basis, we do not circulate it or reproduce its detailed non-public implementation information, figures, or other contents in the response or revised manuscript (Dragonfly Data Science, unpublished technical report, n.d.).
Taken together, the additional methodological clarification and the unpublished technical report directly address the concern that background variability was not considered when identifying candidate event-related surface changes. They show that, for the non-flood classes, NDVI change was one input into a broader detection and classification workflow rather than the final indicator by itself. The workflow explicitly incorporated a historical pixel-level reference distribution, manually labelled aerial-image information, and terrain information.
We agree that these procedures cannot eliminate all measurement error. Seasonal vegetation change, land-use change, cloud contamination, and other non-disaster processes may still affect the classifications, and the workflow does not by itself prove that every mapped change represents verified disaster damage. We will make these limitations clearer. However, the existence of residual uncertainty is different from assuming that the products were constructed from a simple, unadjusted NDVI difference with no consideration of background variability. For the non-flood classes, the available methodological information shows that the workflow was designed to distinguish unusual candidate event-related surface changes from expected background variation, although necessarily imperfectly.
D-Wet needs to be considered separately because the methodological clarification described above was specifically provided for the non-flood impact classes. D-Wet corresponds to the Dragonfly source category labelled “flood” and is interpreted in this study as a post-event wetness/flood-related surface-change class. We do not interpret it as a direct quantitative measure of soil moisture, a complete or independently validated flood-extent layer, or an indicator capable of distinguishing rainfall-induced wetness from flood-induced wetness. We will describe its construction only to the extent supported by the source information available to us.
This distinction is also important for the interpretation of the indicators. We agree that NDVI itself is not a direct measure of structural or monetary damage. Rather, NDVI change is one input into a broader event-specific workflow intended to identify cyclone-related surface change or disturbance. D-Wet, D-Soil, and D-Slope should therefore be interpreted as mapped surface-change or disturbance indicators, rather than as raw vegetation indices or direct measurements of property damage.
A separate question is whether these mapped surface-change indicators correspond to recorded residential insurance outcomes. We do not assume that relationship a priori. Testing whether, and to what extent, the source-provided indicators correspond to recorded NHC claims and payouts is precisely the empirical question examined in this paper.
Comparing remotely mapped conditions with external damage or insurance records has been used in published empirical studies. Thomas et al. (2023) assess remote-sensing flood indicators using spatial accuracy, temporal consistency, and correlation with damage data. Zischg et al. (2018) and Gao et al. (2025) use geolocated insurance claims to evaluate flood-inundation models, while Miura et al. (2022) examine relationships between remotely identified building damage and area-level insured losses. These studies do not validate the Dragonfly products used here, but they support the use of damage or insurance data to evaluate whether mapped indicators contain information relevant to observed losses.
Our results should therefore be interpreted as evidence of statistical associations between the mapped indicators and recorded insurance outcomes, rather than as proof that the indicators directly measure monetary damage, that every mapped change was caused by Cyclone Gabrielle, or that the source classifications have been completely validated. The remaining classification uncertainty is an important limitation, but it does not make the empirical question or the observed associations inherently uninformative. Nor does the use of NDVI as one component of a broader product-generation workflow make meaningful empirical analysis virtually impossible.
We will revise the manuscript accordingly. We will clarify that the Dragonfly products are event-specific mapped surface-change classes rather than raw NDVI measures; distinguish the source-product processing from the subsequent relabelling, meshblock aggregation, and analytical thresholds introduced by us; and provide a clearer high-level description of the product-generation methodology while respecting the non-public status of the detailed technical documentation. We will also expand the discussion of residual seasonality, background change, classification uncertainty, and the absence of independent ground validation. The unpublished technical report will be cited as such, but its detailed non-public contents will not be reproduced or circulated.
References:
Thomas, M., Tellman, E., Osgood, D. E., DeVries, B., Islam, A. S., Steckler, M. S., Goodman, M., & Billah, M. (2023). A Framework to Assess Remote Sensing Algorithms for Satellite-Based Flood Index Insurance. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16, 2589–2604. https://doi.org/10.1109/JSTARS.2023.3244098
Zischg, A. P., Mosimann, M., Bernet, D. B., & Röthlisberger, V. (2018). Validation of 2D flood models with insurance claims. Journal of Hydrology, 557, 350–361. https://doi.org/10.1016/j.jhydrol.2017.12.042
Gao, D., Yin, J., Yang, Y., Mai, J., Lu, Y., Liang, P., Wang, W., Yuan, X., & Yu, D. (2025). Validating typhoon-induced rainstorm flood inundation modelling with insurance claims. Journal of Hydrology, 657, 133178. https://doi.org/10.1016/j.jhydrol.2025.133178
Miura, H., Murata, Y., Wakasa, H., & Takara, T. (2022). Empirical estimation based on remote sensing images of insured typhoon-induced economic losses from building damage. International Journal of Disaster Risk Reduction, 82, 103334. https://doi.org/10.1016/j.ijdrr.2022.103334
Comment 2
Furthermore, NDVI is primarily an index used to assess the condition of vegetation and green spaces. In contrast, residential properties (buildings and immediate lots) subject to insurance claims are predominantly non-vegetated, built environments. Asserting that a drop in NDVI in surrounding farmland or woodland directly correlates with physical destruction or inundation of residential structures represents a major leap in physical logic.
Response: Thank you for raising this point. We agree that NDVI itself is not a direct measure of damage to residential structures, and we do not interpret it in that way. However, our analysis does not assume that a decline in NDVI in surrounding farmland or woodland directly implies physical destruction or inundation of a dwelling. The comment therefore characterises the empirical relationship more directly than it is specified in our study.
First, the Dragonfly layers are not intended simply to represent vegetation loss in surrounding farmland or woodland. The Cyclone Gabrielle Impact Map was developed by Dragonfly Data Science in collaboration with the Social Wellbeing Agency to map and visualise different categories of cyclone impact. The official project materials describe the tool as displaying physical and infrastructure impacts and identify mapped layers for inundated areas, deposited silt, and slips or bare soil on steep slopes (Dragonfly Data Science, n.d.; Social Wellbeing Agency & Dragonfly Data Science, n.d.). The interactive map also reports contextual measures, including the number of buildings located within mapped impacted areas (Social Wellbeing Agency & Dragonfly Data Science, n.d.). Importantly, the number of buildings within a mapped impacted area is a measure of spatial overlap or proximity; it does not establish that every building in that area was physically damaged. Consistent with this limitation, we use the Dragonfly layers as mapped surface-change or disturbance indicators rather than as building-damage inventories.
Second, the insurance outcomes in our study should not be interpreted as measures of damage to residential structures alone. Under the EQCover arrangements relevant to the February 2023 North Island severe-weather events, covered damage could relate to insured residential land and, for some hazard types, to the home itself (see detailed response to reviewer #1). The official Toka Tū Ake EQC guidance states that landslip damage may be covered for both the home and insured residential land, whereas debris and silt inundation from floodwater or high winds and land scour/evacuation may be covered for insured residential land. The same guidance explains that insured residential land primarily includes the land under and up to eight metres around the home or insured outbuildings, the land under the main accessway within 60 metres of the home, and certain associated land structures (Toka Tū Ake EQC, 2023).
These coverage rules establish a plausible category-level connection between some mapped surface-disturbance classes and some forms of covered residential-property damage. For example, mapped silt- or soil-related disturbance may be spatially relevant to claims involving debris or silt on insured residential land, while inferred slope-related disturbance may be relevant to landslip damage affecting insured residential land and, in some cases, the home. However, the coverage rules do not establish that a particular mapped polygon caused a particular claim, or that every claim and payout in our analytical data relates to these specific damage mechanisms.
The relationship examined in our study is therefore not that “NDVI decline in surrounding vegetation implies destruction of a dwelling.” Rather, the remote-sensing products provide mapped surface-change or disturbance classes, while the insurance data provide separately recorded NHC claim and payout outcomes. We examine whether the meshblock-level extent of the mapped indicators is statistically associated with those insurance outcomes. We do not assume a one-to-one physical equivalence between a disturbed pixel and damage to a residential structure.
At the same time, we agree that the reviewer raises a valid spatial-scale concern. Because the main analysis aggregates the mapped indicators to the meshblock level, some mapped disturbance may occur away from the residential properties associated with claims, particularly in larger rural meshblocks. This limits the extent to which a meshblock-level association can be interpreted as local physical correspondence around an affected property.
To address this concern, we will add a complementary claim-location analysis using geocoded claim locations. We will construct sensitivity buffers around the claim locations—for example, 50 m and 100 m—and test whether those buffers intersect the mapped remote-sensing indicator footprints. Similar claim-location buffer methods have been used to evaluate the spatial correspondence between insurance claims and modelled flood extent (Gao et al., 2025). We will interpret this analysis as a local spatial-correspondence check rather than as proof that a mapped disturbance caused an individual claim or as evidence that the associated residential property was directly identified as damaged by remote sensing.
We will also revise the manuscript to clarify the composition and interpretation of the NHC outcomes and to remove language that could suggest that the remote-sensing layers are direct measures of structural damage. The intended interpretation is that the Dragonfly layers provide event-related mapped surface-change or disturbance classes, whereas the insurance variables represent recorded NHC claim and payout outcomes within the applicable coverage arrangements. The empirical analysis evaluates the extent to which these two separately constructed dimensions correspond statistically, while recognising that exposure, vulnerability, insurance coverage, spatial aggregation, and classification uncertainty may all affect that correspondence.
References:
Dragonfly Data Science. (n.d.). Cyclone Gabrielle impact map. https://www.dragonfly.co.nz/case-studies/cyclone-impact-map
Gao, D., Yin, J., Yang, Y., Mai, J., Lu, Y., Liang, P., Wang, W., Yuan, X., & Yu, D. (2025). Validating typhoon-induced rainstorm flood inundation modelling with insurance claims. Journal of Hydrology, 657, 133178. https://doi.org/10.1016/j.jhydrol.2025.133178
Social Wellbeing Agency, & Dragonfly Data Science. (n.d.). Cyclone Gabrielle impact map [Interactive map]. https://d1econosvb0ksc.cloudfront.net/
Toka Tū Ake EQC. (2023, March 3). EQCover community information: North Island severe weather. https://www.naturalhazards.govt.nz/our-publications/eqcover-community-information-cyclone-gabrielle/
Comment 3
If the authors intend to conclude that estimating damage from remote sensing indicators is difficult, it is essential to explore indicators that directly represent cyclone-induced physical damage. A study that skips the critical step of evaluating indicator validity and simply fits basic statistical models remains insufficient. Linking data that fails to separate baseline background noise from true disaster damage directly to insurance claims yields extremely low academic validity and utility.
Response: Thank you for this comment. We would first like to clarify an important point regarding the objective and conclusion of the paper. The paper does not seek to establish the general conclusion that “estimating damage from remote-sensing indicators is difficult.” That statement reflects a broader challenge discussed in the manuscript concerning the translation of mapped hazard-related conditions into economic losses; it is not the principal conclusion of this study. The central question of the paper is whether different remotely mapped surface-change or disturbance indicators provide comparable information about recorded residential insurance outcomes. Our results indicate that they do not: their statistical associations with recorded NHC claims and payouts vary according to the mapped signal, the insurance outcome examined, and the settlement context. This objective is stated in the submitted manuscript as an evaluation of the relative information content of the four indicators, rather than an attempt to determine whether remote sensing can estimate disaster damage in general.
We agree that the meaning of “validation” needs to be defined carefully. If validation refers to a pixel-level assessment of classification accuracy against an independent ground-reference dataset, this study does not perform, and does not claim to perform, that form of validation. We will therefore avoid describing our regression analysis as validation of the underlying remote-sensing classifications. More precisely, our analysis provides an application-relevant assessment of whether the source-provided mapped indicators contain information that is statistically associated with separately recorded insurance outcomes.
There is methodological precedent for using external damage or insurance data as one component of such an assessment, although these data do not replace independent classification-accuracy validation. Thomas et al. (2023), in a framework developed for assessing remote-sensing algorithms for satellite-based flood-index insurance, evaluate performance using three complementary criteria: spatial accuracy against high-resolution imagery, temporal consistency with river-water levels, and correlation with government-reported damage. They describe this as a “convergence of evidence” (p. 2591) across imperfect but informative reference data. Thus, correspondence with damage data is treated as one relevant criterion, while not being presented as a substitute for spatial or temporal validation.
A similar application-dependent principle appears in the flood-model literature. Zischg et al. (2018) argue that validation data and metrics should be adapted to the intended purpose of a model and examine geolocated insurance claims as an alternative validation dataset for flood-exposure applications. Gao et al. (2025) also use insurance claims to assess simulated flood inundation. These studies concern flood models rather than the Dragonfly surface-change products used in our study, and they therefore do not validate our classifications. They nevertheless demonstrate that recorded insurance outcomes can provide an application-relevant external benchmark when the question concerns the correspondence between mapped hazard conditions and insured impacts.
Miura et al. (2022) provide a related remote-sensing precedent. They compare building damage identified from post-disaster imagery with insured economic losses. Their results indicate that estimating loss building by building is challenging, whereas estimating approximate loss rates for spatially aggregated groups of buildings is more feasible because area-level building-damage ratios are correlated with area-level loss rates. Again, this study does not validate the Dragonfly classifications. It shows that examining statistical correspondence between remotely identified impacts and insured losses at an aggregated spatial scale has precedent in the published literature.
Our own empirical design should be interpreted within these limits. The Dragonfly layers are source-provided, event-specific mapped surface-change or disturbance classes, while the NHC variables are separately recorded claim and payout outcomes aggregated to the meshblock level. We do not assume that a mapped polygon represents verified damage to a particular residential property, or that a statistical association proves that a mapped disturbance caused an individual claim. The analysis asks the narrower question of whether variation in the mapped indicators corresponds systematically to variation in the recorded insurance outcomes.
As clarified in our response to the preceding comments, the available methodological information also indicates that the non-flood Dragonfly classes were not generated simply by treating an unadjusted pre-event/post-event NDVI difference as the final indicator. This clarification addresses the specific suggestion that background variability was not considered at all. It does not eliminate the possibility of residual seasonal effects, land-use change, cloud contamination, classification error, or other non-disaster influences. We will make these limitations explicit and will avoid describing the source classifications as completely validated.
We also agree that stronger evidence on local spatial correspondence would improve the assessment. We will therefore add a complementary claim-location analysis using buffers around geocoded claim locations and test whether those buffers intersect the mapped indicator footprints. Gao et al. (2025) use circular buffers of 50–100 m around claim locations as approximations of building footprints and report higher correspondence than when using claim points alone. They also caution that larger buffers can inflate apparent agreement. Guided by this precedent, we will use 50 m and 100 m sensitivity buffers and interpret the results as a local spatial-correspondence check, rather than as proof that a mapped disturbance caused an individual claim or as a substitute for independent pixel-level classification validation.
Taken together, we agree that the study does not provide independent pixel-level validation of the remote-sensing classifications, and we will state this limitation more clearly. We do not, however, agree that this makes the empirical comparison inherently invalid or uninformative. The study evaluates whether source-provided mapped indicators contain information that corresponds to separately recorded insurance outcomes, while explicitly recognising the limits of that correspondence. The additional methodological clarification establishes that, for the non-flood classes, the source workflow involved more than an unadjusted NDVI difference, and the proposed claim-location analysis will provide a more direct assessment of local spatial correspondence.
Finally, we will revise the Discussion to avoid the impression that the paper concludes that remote-sensing damage estimation is generally “difficult.” Our intended conclusion is narrower: the four mapped indicators are not interchangeable, and their associations with recorded insurance outcomes differ according to the mapped signal, the insurance outcome considered, and the spatial context. We will also distinguish this application-level assessment from pixel-level validation of the underlying remote-sensing classifications.
References:
Thomas, M., Tellman, E., Osgood, D. E., DeVries, B., Islam, A. S., Steckler, M. S., Goodman, M., & Billah, M. (2023). A Framework to Assess Remote Sensing Algorithms for Satellite-Based Flood Index Insurance. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16, 2589–2604. https://doi.org/10.1109/JSTARS.2023.3244098
Zischg, A. P., Mosimann, M., Bernet, D. B., & Röthlisberger, V. (2018). Validation of 2D flood models with insurance claims. Journal of Hydrology, 557, 350–361. https://doi.org/10.1016/j.jhydrol.2017.12.042
Gao, D., Yin, J., Yang, Y., Mai, J., Lu, Y., Liang, P., Wang, W., Yuan, X., & Yu, D. (2025). Validating typhoon-induced rainstorm flood inundation modelling with insurance claims. Journal of Hydrology, 657, 133178. https://doi.org/10.1016/j.jhydrol.2025.133178
Miura, H., Murata, Y., Wakasa, H., & Takara, T. (2022). Empirical estimation based on remote sensing images of insured typhoon-induced economic losses from building damage. International Journal of Disaster Risk Reduction, 82, 103334. https://doi.org/10.1016/j.ijdrr.2022.103334
Comment 4
For these reasons, the reviewer concludes that the paper suffers from flawed methodological premises regarding the physical and disaster-related validity of the input data. As the reliability of its conclusions cannot be guaranteed, this manuscript is recommended for Reject.
Response: Thank you for this overall assessment. In light of our responses above, however, we respectfully disagree with the conclusion that the study is based on fundamentally flawed methodological premises or that its findings therefore lack validity.
The concerns raised by the reviewer largely relate to the description and interpretation of the remote-sensing products, the insurance outcomes, and the scope of the conclusions that can be drawn from their statistical correspondence. In the revised manuscript, we will make these elements more explicit, provide a clearer high-level description of the product-generation workflow using information that can appropriately be reported and cited, and add a complementary claim-location buffer analysis as an additional spatial robustness check.
We will also revise the relevant wording in the Discussion to make clear that the paper does not conclude that remote-sensing damage estimation is generally difficult; this was a broader discussion point rather than the principal conclusion of the study.
These revisions will clarify and strengthen the existing research design and its interpretation while retaining the principal empirical framework.
Citation: https://doi.org/10.5194/egusphere-2026-3539-AC2
-
AC2: 'Reply on RC2', Ke Wang, 26 Sep 2026
-
RC3: 'Comment on egusphere-2026-3539', Anonymous Referee #3, 16 Sep 2026
Although the title presents the study as a validation exercise, the methodology describes an exploratory cross-sectional analysis. The study assesses correspondence with insurance outcomes but does not validate the accuracy of the remote-sensing measures against independent ground-truth hazard or damage observations.
The selection and construction of the variables requires clear justification. ClaimRate is calculated by dividing residential claims by the number of building footprints in each meshblock, but the building footprints may include garages, sheds and other roofed structures that are not equivalent to insured dwellings. The authors should explain why this is an appropriate denominator and assess the sensitivity of the results to this measurement mismatch. Similarly, hazard intensity is defined as the affected pixel area divided by the total land area of the meshblock. It is unclear why this represents residential hazard exposure, since affected pixels may cover areas without connection to a residential asset.
Re: the PCA discussion - Line 534, 541-543 highlights identical loading as something that supports the study but this is just a natural result of reducing the dimensions to two, you get sqrt(2)/2. A 2-variable PCA reveals 36.3% unexplained variance.
Citation: https://doi.org/10.5194/egusphere-2026-3539-RC3 -
AC3: 'Reply on RC3', Ke Wang, 26 Sep 2026
Comment 1
Although the title presents the study as a validation exercise, the methodology describes an exploratory cross-sectional analysis. The study assesses correspondence with insurance outcomes but does not validate the accuracy of the remote-sensing measures against independent ground-truth hazard or damage observations.
Response: Thank you for this comment. We agree that the scope of “validation” needs to be defined more precisely. If validation is understood as assessing the classification accuracy of the remote-sensing products against an independent ground-truth dataset of hazard extent or physical damage, our study does not perform that form of validation. The empirical analysis instead examines the correspondence between source-provided mapped surface-change or disturbance indicators and separately recorded NHC claim and payout outcomes.
Importantly, we do not treat insurance claims or payouts as ground truth for the underlying remote-sensing classifications. The two sets of measures capture different dimensions of cyclone impact. The remote-sensing indicators represent mapped surface-change or disturbance signals, whereas the NHC data provide recorded residential claim and payout outcomes within the applicable insurance-coverage arrangements. The empirical question is therefore whether, and to what extent, these separately constructed sets of measures correspond spatially and statistically, rather than whether the remote-sensing products reproduce a ground-truth hazard or damage map.
There is methodological precedent for using observed loss or insurance data as an external benchmark when evaluating hazard or remote-sensing information for loss-related applications. Thomas et al. (2023), for example, include correspondence with reported damage as one component of a broader framework for assessing remote-sensing algorithms. Zischg et al. (2018) use geolocated insurance claims as an alternative validation dataset for flood-model outputs, and Gao et al. (2025) similarly use insurance claims to evaluate typhoon-induced flood simulations. These studies do not imply that insurance outcomes establish the classification accuracy of a remote-sensing product, and we do not make that claim. Rather, they show that recorded loss or insurance data can provide an application-relevant external benchmark for assessing whether mapped hazard or impact information corresponds to recorded loss outcomes.
We nevertheless agree that the terminology surrounding “validation” needs to be used carefully, because it may otherwise be interpreted as referring to the stronger form of ground-truth classification validation described by the reviewer. We will therefore revise the relevant descriptions in the manuscript to distinguish more clearly between ground-truth validation of the underlying remote-sensing classifications, which is not undertaken in this study, and evaluation of their correspondence with separately recorded insurance outcomes, which is the focus of our empirical analysis.
This clarification does not change the empirical objective or design of the study. The analysis remains cross-sectional, and its purpose is to compare how four mapped surface-change or disturbance indicators relate to recorded residential insurance outcomes and whether their information content differs across loss outcomes and settlement contexts. We will make this scope more explicit throughout the manuscript and avoid wording that could imply that the study establishes independent ground-truth validation of the remote-sensing products.
References:
Thomas, M., Tellman, E., Osgood, D. E., DeVries, B., Islam, A. S., Steckler, M. S., Goodman, M., & Billah, M. (2023). A Framework to Assess Remote Sensing Algorithms for Satellite-Based Flood Index Insurance. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16, 2589–2604. https://doi.org/10.1109/JSTARS.2023.3244098
Zischg, A. P., Mosimann, M., Bernet, D. B., & Röthlisberger, V. (2018). Validation of 2D flood models with insurance claims. Journal of Hydrology, 557, 350–361. https://doi.org/10.1016/j.jhydrol.2017.12.042
Gao, D., Yin, J., Yang, Y., Mai, J., Lu, Y., Liang, P., Wang, W., Yuan, X., & Yu, D. (2025). Validating typhoon-induced rainstorm flood inundation modelling with insurance claims. Journal of Hydrology, 657, 133178. https://doi.org/10.1016/j.jhydrol.2025.133178
Comment 2
The selection and construction of the variables requires clear justification. ClaimRate is calculated by dividing residential claims by the number of building footprints in each meshblock, but the building footprints may include garages, sheds and other roofed structures that are not equivalent to insured dwellings. The authors should explain why this is an appropriate denominator and assess the sensitivity of the results to this measurement mismatch.
Response: Thank you for raising this important issue. We agree that the LINZ BuildingCount is not equivalent to the number of insured residential dwellings and that this creates a measurement limitation in the denominator used to construct ClaimRate. The LINZ Building Outlines dataset identifies mapped roof outlines and may include garages and large sheds (Land Information New Zealand, 2023). In the analytical building-outline data available to us, we do not have a sufficiently complete and validated structure-use classification that would allow us to distinguish residential dwellings reliably from these other structures across the study area. We therefore cannot construct a dwelling-only denominator without imposing an additional classification that is not supported consistently by the available analytical fields.
Our use of BuildingCount should consequently not be interpreted as an estimate of the number of insured dwellings. Its purpose is to normalise the number of NHC residential claim records by the mapped building stock within each meshblock. ClaimRate should therefore be interpreted as the number of NHC residential claim records relative to mapped building stock, rather than as the proportion or probability of insured dwellings that made a claim. We will revise the manuscript to make this distinction explicit.
At the same time, the NHC insurance outcomes should not be interpreted as relating exclusively to damage to the main dwelling structure. The EQCover guidance for the North Island Severe Weather events states that, depending on the hazard, coverage may apply to the home and/or insured residential land. It further explains that insured residential land can include land under and around insured outbuildings, explicitly giving sheds and garages as examples (Toka Tū Ake EQC, 2023). This means that the insurance outcomes considered in our study can reflect a broader set of insured residential losses than damage to the main dwelling structure alone. Importantly, however, this does not establish that every garage or shed represented in the LINZ Building Outlines dataset constitutes an insured exposure, and we do not assume such an equivalence. The mismatch between mapped building outlines and insured residential exposures therefore remains a limitation of the denominator.
Because the available LINZ data do not provide a sufficiently comprehensive classification for constructing a dwelling-only denominator, we cannot directly eliminate this measurement mismatch by recalculating ClaimRate using only insured residential dwellings. We will state this limitation explicitly in the revised manuscript and avoid interpreting ClaimRate as a dwelling-level claim incidence measure.
We will also add a complementary spatial robustness analysis using the available geocoded claim locations. Specifically, we will construct buffers around observed claim locations and examine their spatial correspondence with the remotely sensed hazard footprints (Gao et al., 2025). This addresses a related spatial concern in the construction of our exposure measures: because the main hazard variables are calculated as the affected area relative to total meshblock land area, some mapped physical disturbances may occur away from locations associated with residential claims. The buffer analysis allows us to examine whether remotely mapped physical disturbances occur in proximity to observed claim locations, providing a more spatially targeted assessment of the correspondence between the remote-sensing indicators and insured losses.
We emphasise that this additional analysis does not eliminate the BuildingCount denominator mismatch and will not be presented as doing so. Rather, it provides complementary evidence addressing the related concern that meshblock-level hazard measures may capture physical disturbances that are spatially separated from claim-generating locations. Together with a clearer interpretation of ClaimRate and an explicit acknowledgement of the denominator limitation, this analysis will allow us to define more precisely what the insurance and remote-sensing measures can—and cannot—represent.
References:
Gao, D., Yin, J., Yang, Y., Mai, J., Lu, Y., Liang, P., Wang, W., Yuan, X., & Yu, D. (2025). Validating typhoon-induced rainstorm flood inundation modelling with insurance claims. Journal of Hydrology, 657, 133178. https://doi.org/10.1016/j.jhydrol.2025.133178
Land Information New Zealand. (2023). NZ Building Outlines (2023 version) [Data set]. LINZ Data Service. https://data.linz.govt.nz/layer/101290-nz-building-outlines/
Toka Tū Ake EQC. (2023, March 3). EQCover community information: North Island severe weather. https://www.naturalhazards.govt.nz/our-publications/eqcover-community-information-cyclone-gabrielle/
Comment 3
Similarly, hazard intensity is defined as the affected pixel area divided by the total land area of the meshblock. It is unclear why this represents residential hazard exposure, since affected pixels may cover areas without connection to a residential asset.
Response: Thank you for raising this point. We agree that the share of affected land within a meshblock should not be interpreted directly as a measure of residential asset exposure. In our analysis, the hazard-intensity variables are constructed as the area identified as affected by each remote-sensing indicator divided by the total land area of the meshblock. They therefore measure the spatial intensity of mapped surface-change or disturbance signals within the meshblock, rather than the proportion of residential assets exposed to those signals. We will revise the manuscript to make this interpretation explicit and avoid language that could imply that these area-based measures directly represent residential hazard exposure.
The purpose of these measures is to characterise how extensively each type of remotely mapped surface-change or disturbance signal occurs within a common local geographic unit. The empirical analysis then examines whether meshblocks with a greater spatial intensity of mapped disturbance also exhibit different recorded residential insurance outcomes. This relationship is tested empirically rather than assumed. We therefore agree with the reviewer that affected pixels need not necessarily coincide with residential assets, particularly where mapped disturbance occurs in parts of a meshblock that are spatially separated from residential properties.
To address this concern more directly, we will add a complementary spatial analysis using the available geocoded claim locations. Specifically, we will construct buffers around recorded claim locations and examine whether the remotely sensed indicator footprints intersect with, or occur in proximity to, these locations. Claim-location buffer methods have previously been used to assess spatial correspondence between insurance claims and modelled flood extent (Gao et al., 2025). This provides a more spatially targeted assessment of whether the mapped surface-change or disturbance signals underlying our meshblock-level measures are also present near locations associated with recorded residential insurance claims.
This additional analysis does not convert the remote-sensing measures into direct measures of residential asset exposure. Claim locations represent locations associated with recorded insurance claims and do not provide a complete spatial inventory of all residential assets. We will therefore interpret the buffer analysis specifically as a test of the spatial correspondence between mapped surface-change or disturbance signals and recorded claim locations, rather than as a measure of the proportion of residential assets exposed.
Accordingly, we will revise the manuscript to (i) define the existing hazard-intensity variables more precisely as area-based measures of mapped surface-change or disturbance within each meshblock, rather than direct measures of residential exposure; (ii) explicitly acknowledge that some affected pixels may occur away from residential assets; and (iii) add the claim-location buffer analysis as a complementary, finer-scale assessment of the spatial correspondence between the remote-sensing indicator footprints and recorded claim locations.
References:
Gao, D., Yin, J., Yang, Y., Mai, J., Lu, Y., Liang, P., Wang, W., Yuan, X., & Yu, D. (2025). Validating typhoon-induced rainstorm flood inundation modelling with insurance claims. Journal of Hydrology, 657, 133178. https://doi.org/10.1016/j.jhydrol.2025.133178
Comment 4
Re: the PCA discussion - Line 534, 541-543 highlights identical loading as something that supports the study but this is just a natural result of reducing the dimensions to two, you get sqrt(2)/2. A 2-variable PCA reveals 36.3% unexplained variance.
Response: Thank you for this comment. We agree that the identical loadings in the two-variable PCA should not be interpreted as independent evidence supporting the composite index. In the current manuscript, we placed too much interpretive weight on the equal loadings of S-Flood and D-Soil (0.7071 for each variable). We will remove the language that treats these identical loadings as substantive support for the PCA-based index and report them descriptively without assigning them additional interpretive significance.
We also agree that the amount of variance not captured by the first principal component should be made explicit. In the two-variable PCA, PC1 explains 63.7% of the total variance, leaving 36.3% outside the first component. We will therefore avoid implying that PC1 provides a comprehensive representation of the information contained in S-Flood and D-Soil. Rather, PC1 captures the largest common dimension of variation between the two indicators while retaining only part of their total variation.
More generally, we will clarify the role of the PCA in the study. The PCA is intended as a supplementary dimension-reduction exercise that provides a data-driven composite summary of the common variation in S-Flood and D-Soil. It is not intended to demonstrate that the two indicators are equivalent or interchangeable, nor to provide a comprehensive measure of cyclone hazard exposure. The individual remote-sensing indicators remain the primary measures in our analysis, while the PCA-based index provides a complementary assessment of whether their shared variation is also associated with the observed insurance outcomes.
We will revise the relevant discussion accordingly by removing the interpretation attached to the identical loadings, explicitly reporting the variance explained and unexplained by PC1, and narrowing our interpretation of the PCA-based index to reflect its supplementary role in the analysis.
Citation: https://doi.org/10.5194/egusphere-2026-3539-AC3
-
AC3: 'Reply on RC3', Ke Wang, 26 Sep 2026
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 197 | 80 | 47 | 324 | 28 | 36 |
- HTML: 197
- PDF: 80
- XML: 47
- Total: 324
- BibTeX: 28
- EndNote: 36
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The study ‘Validating Remote-Sensing Measures of Natural Hazards: Granular-Level Links to Insured Loss during a Cyclone’ looks at the correlation between four different remote-sensing indicators and the occurrence and magnitude of residential losses represented through natural-hazard insurance claims following Cyclone Gabrielle, which hit the North Island of New Zealand in February 2023.
The study is an interesting and timely contribution to improvements in the validation of risk and hazard indices. While the results are interesting and worth publishing in NHESS, several key points made in the manuscript require further clarification. I would therefore recommend the manuscript for publication after major revisions.
General comments
I think the manuscript will do well in future reviews if the authors are able to address the following issues:
Specific comments:
L37ff: ‘This distinction matters for loss analysis because insurance claims better reflect realised damage rather than meteorological intensity alone. Remote-sensing indicators can therefore provide a more direct spatial link between physical disturbance and observed residential losses and damages.’ These two sentences confuse hazard, risk and vulnerability. Both the meteorological intensity and the used remote sensing indicators are just different ways to measure hazard intensity. Losses are a function of properties exposed to the measured hazard and how vulnerable these properties are to the hazard.
L164:’This “standing water” image has not been validated against ground observations of flood extent.’ That seems like a useful thing to do as I would assume that standing water can be the result of various processes including in-situ rainfall which is different from standing flood water from rivers. Is there a reason why the authors have not done that?
L188f: What is the rationale for normalising LogPayouts by the total number of buildings and not by the number of affected buildings in a mesh block? Normalising by the total number of buildings makes it difficult to distinguish between areas that have few buildings with high damages and areas with many buildings with low damages.
L288ff: It would be good to get at least a bit more detail about the remote sensing indicators used and how they were constructed. For example, for D-Wet, how was the threshold defined to distinguish between higher soil moisture levels following rainfall and higher soil moisture levels caused by flooding?
L423: ‘This sample’ repetition