Abstract
Post-disaster satellite imagery can document damaged buildings after an event; disaster preparation, however, requires decisions before those observations exist. This paper proposes a physically grounded framework to estimate the range of plausible impacts from Caribbean tropical cyclones using information available strictly before a declared forecast issue time. Pre-event optical and synthetic-aperture radar satellite observations would establish the location and characteristics of exposed buildings, transport corridors and land cover. Archived as-issued weather forecasts, terrain and hydrological information would then drive physically based wind, rainfall and inundation scenarios. An artificial intelligence model would estimate probabilistic asset damage conditional on those hazards, exposure and vulnerability assumptions. Economic loss would be calculated from transparent asset valuations, not inferred from satellite texture alone. The methodology separates post-event imagery used to label and verify observed damage from the information permitted during a prospective forecast. A proposed retrospective evaluation uses historical Caribbean cyclones, including Hurricane Beryl in July 2024, with complete event holdouts and temporal data audits. Skill would be assessed against conventional hazard–exposure–vulnerability baselines using classification scores, calibration, location error and monetary loss error. No casualty predictions, pre-landfall skill scores or operational system are claimed. Its contribution is a falsifiable end-to-end protocol for producing and testing decision-ready uncertainty estimates before a storm reaches exposed communities.
1. Introduction
Disaster impact becomes visible after an event. Satellite imagery then records damaged roofs, blocked roads and inundated neighbourhoods, often before conventional field surveys can cover every affected place. The urgent planning problem is different. Before a tropical cyclone arrives, disaster managers need estimates of which facilities may be affected and the range of resources that might be required. They must make those decisions without the benefit of a future damage image.
The proposed framework combines pre-event images, a physically interpretable hazard simulation and a probabilistic model of asset damage. It seeks to produce geographically explicit estimates rather than just a regional exposure total. A hospital with road access at risk has different operational implications from an exposed but unoccupied structure. A model should preserve those distinctions and expose the uncertainty arising from incomplete asset inventories and the forecast track.
This article presents an evaluation protocol, not results from an operating Caribbean impact-prediction platform. It does not claim to predict deaths or to replace an evacuation directive. The research question is whether imagery-derived asset features and physically plausible hazard scenarios add useful pre-event information beyond the established combination of hazard maps, exposure registers and vulnerability curves.
2. Scientific and operational background
UNDRR distinguishes the physical hazard from the people, infrastructure and assets exposed to it, and from the characteristics that make those exposed elements vulnerable [1]. This separation is fundamental. A storm's wind field is not a damage field; the same wind can produce different consequences for buildings with different construction, maintenance or protective features. Similarly, a model of total economic loss must include asset values and damage fractions, not merely the number of buildings within the storm footprint.
Satellite-based damage assessment research provides useful tools, but its evaluation task is often post-event. The xBD dataset, introduced by Gupta and colleagues, includes pre- and post-disaster satellite imagery with building damage annotations [2]. When post-event imagery is available, a model can learn to classify observed physical damage. Those results do not prove the model can predict the same damage before a storm: at forecast time the post-event image, damage label and realised hazard are not yet available.
Historical event selection can draw on NOAA's International Best Track Archive for Climate Stewardship (IBTrACS), which harmonises tropical-cyclone track records [3]. A best-track record is a retrospective estimate and must not be mistaken for what a forecaster knew in real time. Rigorous testing needs archives of forecasts as issued, with their corresponding publication times. Otherwise, the experiment inadvertently gives the model information discovered after the decision deadline.
Optical imagery can help delineate buildings and land cover under suitable atmospheric conditions. Synthetic-aperture radar provides complementary microwave measurements that can be acquired despite cloud cover, although geometric distortions and sensor-specific interpretation remain important. Neither modality reliably reveals every structure's strength or all occupants. Missing vulnerability information must remain a disclosed limitation rather than being filled with apparently precise but invented labels.
3. Data architecture and temporal boundary
3.1 Forecast issue time
Define a decision time t₀ for each historical case. Every input used to predict impacts must have been available to a decision-maker by t₀. This includes publication latency as well as the nominal observation time. A satellite acquisition from the preceding day might have arrived too late for an earlier decision point, and the final archived cyclone track is not a valid replacement for the track forecast available at t₀.
Permitted predictors include pre-event Sentinel-1 or Sentinel-2 imagery where available, building and road inventories, terrain, land-cover maps, historical exposure data and forecast hazard products issued before the deadline. A documented version manifest would store the provider, acquisition date, release date, coordinate reference system, geographic coverage and available uncertainty for each data layer. Changes in an asset register after the event require special treatment, because they may contain knowledge of the disaster itself.
3.2 Post-event truth held aside
Post-event imagery, damage surveys, official loss assessments and observed storm trajectories are retained only for model training in historical training events or evaluation in completely held-out test events, as appropriate. Within a test event, the post-event record cannot be passed to an agent asked to simulate what would have been known before landfall. A test set should be protected at the event level, not merely at the pixel level, because nearby tiles can encode the same storm and damaged neighbourhood.
3.3 Hazard generation
A physical hazard ensemble would combine wind-field estimates, rainfall forcing and flood routing over terrain. The precise level of physical complexity must suit the inputs and use case. A physically based model need not resolve every turbulent eddy to provide a useful inundation scenario, but it must state which processes are parameterised, what the boundary conditions are and how uncertainty enters. Coastal surge, if included, requires additional bathymetry, ocean boundary conditions and modelling validation; otherwise, the study should state that surge is excluded.
4. Probabilistic damage and loss formulation
4.1 Conditional asset damage
Let Dᵢ be the damage category or fractional damage for asset i. Let Hᵢ describe the event-specific hazard at that location; Eᵢ the observed asset representation; and Vᵢ the vulnerability information known before the storm. The proposed model estimates:
The estimated loss for an asset is the expected fraction of damage multiplied by its independently specified replacement or repair value Cᵢ. Integrating across the hazard ensemble gives:
A credible risk assessment would report a distribution of total losses rather than a single point estimate. The uncertainty decomposition should distinguish forecast-track spread, rainfall and wind intensity, inundation-model error, building-detection error, vulnerability-model error and missing economic values. Correlation is important: several structures in the same coastal settlement may be affected by the same uncertain hazard, so individual uncertainties cannot simply be treated as independent.
4.2 Verification measures
For a binary threshold such as moderate-or-worse damage, the Brier score evaluates probabilistic calibration and discrimination through mean squared probability error:
Damage-category F1, precision and recall capture classification performance but can conceal badly calibrated probabilities or physically displaced hotspots. Spatial evaluation must therefore include location error and scale-sensitive metrics. Economic loss should be verified in monetary units and appropriate relative errors, acknowledging that percentage errors become unstable when the observed loss is close to zero.
5. Interactive hazard illustration
6. Proposed retrospective study
6.1 Event and geographic holdouts
One planned historical case is Hurricane Beryl during July 2024, subject to the availability of archived as-issued forecasts, suitable satellite acquisitions, credible asset inventories and independently sourced observations of damage. At least one other Caribbean event should be withheld from model selection and training to test transfer between islands and storm types. An event cannot contribute imagery tiles to both training and evaluation merely because those tiles are geographically distinct.
Source layers must be aligned on spatial coverage and legal use. Data quality should be reported for buildings omitted by mapping, cloud-limited optical imagery, floodplain topography and economically informal assets. Exposed population can be reported using transparently stated census or gridded population assumptions; it is not a verified estimate of fatalities. Social vulnerability needs separate consideration rather than being inferred solely from roof appearance or neighbourhood location.
6.2 Comparison methods
The system should be evaluated against an established hazard–exposure–vulnerability baseline without satellite-image representation learning, a model with static hazard layers but no dynamic forecast, and an ablated version without physical hazard features. Training and model selection should use only training and validation events, with hyperparameters frozen before test events are opened. All inputs and results need reproducible data manifests, software versions and exact issue-time cutoffs.
| Output | Planned validation | Key limitation |
|---|---|---|
| Building damage class | Event-held-out F1, sensitivity and specificity | Incomplete post-event labels |
| Damage probability | Brier score, calibration plots and reliability by subgroup | Small rare-damage samples |
| Impact geography | Spatial overlap and displacement error | Geolocation and segmentation uncertainty |
| Economic loss | Absolute monetary error and uncertainty coverage | Valuation and insurance gaps |
| Preparedness usefulness | Blinded decision exercise with response officials | Operational decisions are context-dependent |
Table 1. Proposed validation plan; these are not reported experimental results.
6.3 Decision-oriented evaluation
Emergency managers may need to decide where to pre-position equipment, whether a transport route may become impassable and which shelters require further inspection. Such decisions can be simulated in a retrospective tabletop assessment with officials viewing forecast-time products only. The assessment must measure false alarms and missed critical facilities alongside time saved and whether uncertainty is understood. Faster map production alone is not proof of a better emergency decision.
7. Discussion, limitations and ethical boundaries
A visually persuasive impact map can invite misplaced confidence. A classifier may distinguish roof types in one country but fail where construction and image acquisition differ. The same hazard intensity can produce different losses because drainage, maintenance, code compliance and building value differ. Flood models are sensitive to terrain quality and unresolved local channels. All those limitations require explicit error reporting and sensitivity analysis.
Missing data are not neutral. Informal settlements, small agricultural assets and communities with fewer records may be underrepresented in both training labels and economic valuations. The project should measure coverage by locality and social vulnerability, and explain where estimates are too uncertain for resource allocation. Personal data and high-resolution imagery of homes should be governed by lawful purpose, minimisation and access controls.
Finally, estimating population exposed is not the same as predicting injury or death. Mortality models would require separate epidemiological information, verified outcomes and a dedicated independent validation study. The proposed framework makes no such claim. Its intended use is to help officials understand plausible damage and exposure under a forecast, not to deliver definitive instructions about individual survival.
8. Conclusion
Satellite imagery has a role in pre-impact assessment when it is used to characterise exposure before a storm, paired with physical hazard scenarios and honest uncertainty about vulnerability. A credible empirical contribution will depend on strict issue-time data separation and comparisons with conventional methods, not on visually realistic damage maps. The proposed research offers a test of that contribution; its forecasting accuracy and operational value remain to be established.
References
- United Nations Office for Disaster Risk Reduction (2017). Disaster risk terminology. https://www.undrr.org/terminology/disaster-risk.
- Gupta, R. et al. (2019). Creating xBD: A dataset for assessing building damage from satellite imagery. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. https://openaccess.thecvf.com/content_CVPRW_2019/html/cv4gc/Gupta_Creating_xBD_A_Dataset_for_Assessing_Building_Damage_from_Satellite_CVPRW_2019_paper.html.
- NOAA National Centers for Environmental Information. International Best Track Archive for Climate Stewardship (IBTrACS). https://www.ncei.noaa.gov/products/international-best-track-archive.
- Copernicus Data Space Ecosystem. Sentinel-1 and Sentinel-2 missions and data access. https://dataspace.copernicus.eu/.