Impact-based forecasting of tropical cyclone-related human displacement to support anticipatory action

The study addresses the need for impact-based forecasting to support anticipatory humanitarian action for tropical cyclones (TCs), which displace an average of 9.3 million people annually (2017–2020). Current anticipatory planning largely relies on weather forecasts, missing key social and exposure nuances that determine displacement outcomes. The authors propose a globally consistent, regionally calibrated impact forecast that integrates meteorological forecasts (hazard), population exposure, and vulnerability (impact functions) within the open-source CLIMADA platform. They emphasize quantifying uncertainty across hazard, exposure, and vulnerability to inform decision-making. A case study of TC Yasa (Fiji, December 2020) and a sensitivity analysis across all recorded TC displacement events (2017–2020) explore how uncertainties affect forecast outcomes and how their relative importance shifts with forecast lead time.

The paper situates its contribution within prior work on impact-based forecasting and decision support for evacuations, largely focused on the U.S., combining weather forecasts with traffic information and risk communication platforms (e.g., CHIME). While these tools support evacuation planning, the authors identify a gap: no globally consistent tool provides impact information for human displacement. The work builds on the CLIMADA platform, previously used for long-term impact assessments (e.g., building damages, heat-related mortality) and operational forecasting for winter storms, extending it to TC-related displacement. It leverages the IDMC displacement database as the most comprehensive global source for calibration of displacement vulnerability functions and notes ongoing issues with reporting variability and data gaps in displacement metrics.

- Platform: CLIMADA v3.3 (Python) to combine hazard, exposure, and vulnerability into spatially explicit displacement risk at ~4 km (150 arc-second) resolution on land, globally. - Hazard: ECMWF-IFS tropical cyclone track ensemble forecasts (51 members), updated 00/12 UTC, with positions, pressures, and wind speed at 6-hour intervals interpolated to 1-hour. Maximum 1-minute sustained wind at 10 m used as proxy for TC intensity. Wind fields computed using a revised hurricane pressure–wind model. Real-time tracks via ECMWF FTP; archived forecasts (2017–2020) via TIGGE; historical best tracks via IBTRACS for calibration. - Exposure: Gridded Population of the World v4 (GPWv4) from SEDAC, allocating census-based population uniformly within administrative units. - Vulnerability (impact functions): Sigmoid-type functions relating wind speed to probability of displacement. Calibrated using IDMC displacement records (2008–2020). Countries grouped into 10 regions (building on Eberenz et al., with Oceania split into Australia/New Zealand and Pacific Islands) to ensure sufficient samples; separate regional functions derived. - Deterministic calibration (for meteorological variability-only analysis, Fig. 2): One impact function per region minimizing root mean square fraction (RMSF) error across events to reduce relative error spread. - Uncertainty representation (for global uncertainty analyses, Figs. 3–5): Calibrate one function per historical event minimizing RMSE; per-region bundles of functions represent vulnerability uncertainty. Functions within the 80% confidence interval around the regional best estimate used for sampling. - Uncertainty analysis: Global (simultaneous) variation of inputs using quasi-Monte Carlo (Sobol) sampling (>8000 samples). Inputs varied: (i) forecast ensemble member (meteorology), (ii) total population scaling uniformly in [80%, 120%], (iii) impact function sampled from regional bundle within 80% CI. For each sample, compute spatial impact and aggregate to total displaced. - Sensitivity analysis: First-order Sobol indices (Saltelli approach) quantify the contribution of each uncertain input to output variance. Conducted for TC Yasa at multiple lead times and across all recorded TC displacement events (2017–2020) at lead times from 3.0 to 0.5 days before landfall. - Case study and lead-time experiments: TC Yasa (Fiji, Dec 2020), with forecasts at +48 h and at 0.5-day steps from 3.5 days to near landfall. Computation time for full uncertainty distribution (~8000 runs) ~10 minutes on a 2.3 GHz quad-core Intel i7 with 16 GB RAM. - Validation/bias: Comparison with IDMC reports and IBTRACS-based modelling indicates an overestimation bias in ECMWF-track-driven forecasts (0.5-day lead), potentially due to longer simulated track lifetimes; no bias correction applied in this version (details in Supplementary S4).

- TC Yasa case (Fiji, +48 h lead): Considering only meteorological ensemble variability (51 members), the forecasted total displaced averaged ~172,000 (172.205K), with outcomes ranging from ~3,500 to ~450,000. Distribution is right-skewed with a long high-impact tail, driven by whether tracks intersect densely populated areas (e.g., Viti Levu cities: Suva, Lautoka, Nadi). - Global uncertainty analysis (varying hazard, exposure, vulnerability): For TC Yasa at +48 h, the forecasted distribution remained strongly right-skewed, with mean ~123,391 displaced; mean and peak probability closer to the IDMC-reported 23,414. First-order Sobol sensitivity indices at +48 h: impact function (vulnerability) 0.411; meteorological forecast 0.320; total population 0.005, indicating dominant contributions from vulnerability and meteorology, negligible from homogeneous population scaling. - Lead-time dependence (TC Yasa): Global-uncertainty distributions are wider at longer lead times and narrow approaching landfall. Sensitivities shift: meteorological uncertainty contribution decreases towards landfall; vulnerability contribution increases; exposure scaling remains near zero across lead times. - Multi-event analysis (2017–2020 TC displacement events): Forecasts show an overall overestimation bias relative to IDMC reports (Supplementary S4). Sensitivity patterns mirror the Yasa case in aggregate: meteorology dominates at longer lead times; vulnerability grows in importance at shorter lead times. However, event-level variability exists (e.g., TC Harold shows high vulnerability sensitivity even at long lead times) despite comparable forecast uncertainty magnitudes. - Practical guidance: For longer lead times, prioritize meteorological expertise on track/intensity evolution; near landfall, emphasize local knowledge of community vulnerability for anticipatory action planning. - Operational feasibility: The open-source implementation runs quickly on modest hardware and is transferable to other hazards and impact types.

The study demonstrates that spatially explicit, probabilistic impact forecasts for TC-related displacement can be generated in near-real time using open data and open-source tooling (CLIMADA). By translating weather forecasts into risk via exposure and vulnerability, the approach provides actionable information (numbers and locations of people at risk) beyond traditional weather products, supporting anticipatory action (e.g., evacuations, aid pre-positioning). Critically, the authors show that forecasted impacts are highly skewed and that ensemble means can obscure tail risks that drive major humanitarian responses; thus decisions should consider full uncertainty distributions. Uncertainty arises from meteorological variability, exposure estimates, and vulnerability functions; a global uncertainty framework reveals how their interactions shape outcomes. Sensitivity analyses clarify which uncertainties matter most at different lead times, offering time-specific guidance: focus on meteorology further out; focus on vulnerability near landfall. While validated against IDMC data, forecasts show a tendency to overestimate (partly due to ECMWF track characteristics), underscoring the need for continued calibration and potential bias correction. The framework’s flexibility and openness can foster collaboration and broader adoption across stakeholders and hazards.

The paper contributes a globally consistent, regionally calibrated, open-source system to forecast TC-related displacement impacts using meteorological ensembles, population exposure, and calibrated vulnerability functions in CLIMADA. It provides probabilistic, spatially explicit forecasts and quantifies uncertainties and sensitivities, enabling more informed anticipatory actions. Key insights include the strong skew of impact distributions, the shifting dominance of uncertainty sources with lead time, and operational feasibility at low computational cost. Future work should: (i) incorporate multiple and compound TC sub-hazards (rainfall, storm surge) and cascading impacts, (ii) improve exposure uncertainty representation beyond homogeneous scaling, (iii) refine vulnerability calibration with improved displacement data and local context, (iv) address biases between forecast and historical track datasets (e.g., bias correction), and (v) extend impact forecasting to additional impact types (e.g., housing damage, mortality) and hazards (floods, heatwaves, storm surge, drought).

- Hazard representation: Displacement linked solely to wind speed; storm surge, pluvial/fluvial flooding, and compound/multi-hazard interactions are not explicitly modelled. Maximum sustained wind is used as a proxy but may miss important drivers and compounding effects. - Exposure uncertainty: Population exposure uncertainty is modelled only as a uniform total scaling (80–120%) due to lack of detailed uncertainty from GPWv4; locational/cluster uncertainties are not represented, likely underestimating exposure sensitivity. - Vulnerability data and calibration: IDMC displacement data, while the most comprehensive, contain reporting variability and gaps; vulnerability calibrated at regional scale, not local, possibly missing local adaptation and heterogeneity. Calibration strategies differ between deterministic and uncertainty analyses. - Forecast data biases: Differences between ECMWF forecast tracks and IBTRACS historical tracks (e.g., longer forecast lifetimes) contribute to overestimation; no bias correction applied in this version. - Impact scope: Displacement is modelled as a direct function of hazard; indirect causes (e.g., infrastructure failures, loss of services) and trapped/immobile populations are not explicitly represented. - Generalisation: Overestimation bias observed in 2017–2020 events suggests further validation and refinement are needed before operational adoption at scale.