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

Human displacement due to extreme weather events, particularly tropical cyclones (TCs), is a significant global challenge. TCs are the second leading cause of displacement after floods, averaging 9.3 million people displaced annually (2017-2020). Displacement impacts vary, from short-term evacuations to long-term displacement due to damage and economic disruption. International and national aid is often required for post-disaster recovery, encompassing shelter, water, food, healthcare, and long-term support. Anticipatory action, such as evacuation planning and humanitarian aid coordination, is crucial in mitigating the negative effects of TCs. Current planning largely relies on weather forecasts, lacking crucial information on population exposure and community resilience. Impact forecasts, which translate weather information into risk by incorporating social variables, offer a significant advancement. While tools exist for specific regions (e.g., the US), a globally consistent impact forecast for human displacement is lacking. This research presents a proof-of-concept implementation of such a global forecast, aiming to provide standardized and comparable information for more effective anticipatory action decision-making.

Existing literature highlights the growing need for improved forecasting of displacement due to tropical cyclones. Studies such as those by the Internal Displacement Monitoring Centre (IDMC) provide valuable data on displacement numbers, but lack the integration of this data with meteorological forecasts in a sophisticated manner. The World Meteorological Organization (WMO) and the International Federation of Red Cross and Red Crescent Societies (IFRC) have advocated for anticipatory action and forecast-based financing, but these efforts often rely on simpler weather forecast thresholds. Research on evacuation planning in the US has shown the benefit of integrating weather forecasts with other data sources, but these models are not globally applicable. This study leverages the CLIMADA platform, previously used for long-term climate impact assessments and operational impact forecasts of building damages, to create a globally consistent displacement forecast that addresses the limitations of current methods. The integration of hazard, exposure, and vulnerability datasets within a probabilistic framework allows for a more nuanced and comprehensive understanding of displacement risk than previous approaches.

This study employs the open-source probabilistic risk assessment platform CLIMADA to generate TC-related displacement impact forecasts. The methodology integrates three key datasets: 1) TC track forecasts and wind fields from the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System (ECMWF-IFS), providing hazard information; 2) global population distribution data from the Gridded Population of the World (GPW v4) dataset, representing exposure; and 3) a set of calibrated impact functions, defining the vulnerability of people to displacement based on wind speed. The maximum 1-minute sustained wind speed at 10 m above ground is used as a proxy for TC hazard intensity. The impact functions, in the form of sigmoid curves, are calibrated using reported displacement data from the IDMC database for events between 2008 and 2020. To account for regional differences in vulnerability, separate impact functions are calibrated for ten regions globally. A probabilistic approach is adopted, incorporating uncertainties in the meteorological forecasts (using ensemble members), population estimates (homogeneously scaling total population), and impact functions (using an ensemble of functions calibrated to each past event). A quasi-Monte Carlo sampling method with over 8000 runs is used to capture the full range of plausible outcomes, complemented by a sensitivity analysis using the Sobol' method to assess the relative contributions of meteorological, exposure, and vulnerability uncertainties to the overall uncertainty in the displacement forecast. The analysis was performed for TC Yasa in Fiji and extended to all worldwide TC events with displacement records from 2017 to 2020.

The case study of TC Yasa in Fiji demonstrates the model's capacity to provide spatially explicit displacement forecasts. The forecast reveals a right-skewed distribution of potential displacement numbers, ranging from 3,500 to 450,000, with a mean of 172,000 (considering only meteorological uncertainty) and 123,391 (considering global uncertainty). The reported displacement from the IDMC was 23,414. The global uncertainty analysis, incorporating uncertainties in hazard, exposure, and vulnerability, produced a more realistic mean forecast closer to the actual reported number. The sensitivity analysis for TC Yasa showed that at a two-day lead time, the impact forecast uncertainty was most sensitive to the impact function (sensitivity index 0.411), followed by the meteorological forecast (0.320), while the total population had minimal impact (0.005). The analysis across all TC events from 2017 to 2020 revealed a general trend: meteorological uncertainty dominates at longer lead times, while vulnerability uncertainty becomes more significant as landfall approaches. However, significant variation across individual events was observed. The model shows a bias towards overestimation of displacement compared to IDMC data, particularly for longer lead times, this is likely due to the differences between the historical TC tracks used for calibration and the real-time ECMWF forecast tracks.

This study successfully demonstrates the feasibility of producing spatially explicit, impact-based forecasts for TC-related displacement using publicly available, open-source data. The probabilistic approach is superior to deterministic models, offering a fuller representation of the uncertainty surrounding displacement estimates. The skewed nature of the impact distributions underscores the importance of considering the full range of plausible outcomes, particularly the tail risks of high-impact scenarios, which are often underrepresented by average measures. The sensitivity analysis highlights the dynamic interplay between meteorological uncertainty and vulnerability at different forecast lead times, providing crucial guidance for decision-makers planning anticipatory actions. At longer lead times, emphasis should be placed on meteorological forecasts; closer to landfall, community-level knowledge of vulnerability becomes paramount. However, variations exist across individual events, suggesting that the predictive accuracy and uncertainty sources vary with the meteorological conditions of specific storms.

This research presents a novel, open-source, globally applicable model for forecasting TC-related displacement, providing vital information for anticipatory action. The model's probabilistic nature effectively captures uncertainty, highlighting the importance of considering the full range of potential outcomes. The sensitivity analysis offers valuable insights into the dynamic shift in the relative importance of meteorological and vulnerability uncertainties across different forecast lead times. Future work should focus on refining the model by incorporating additional hazard factors (storm surge, rainfall), improving exposure data and their uncertainties, and further investigating the complex relationships between hazard, exposure, vulnerability, and displacement. Expanding the model to include other impact types and hazards is also a key direction for future research.

The model has some limitations. While it integrates key datasets, some uncertainties remain unaddressed, particularly in the population exposure data (homogeneous scaling is applied due to the lack of uncertainty data), the relationship between displacement and TC hazards is simplified. The model considers only the wind speed as a proxy for the overall hazard, neglecting other factors like storm surge, flood, and rainfall. Compound hazards and cascading effects are not explicitly modeled, limiting its ability to capture all potential contributing factors to displacement. The bias towards overestimation might result from differences in historical and forecast track intensity. Additionally, the model does not directly capture the impacts on individuals who choose to remain or become trapped in impacted areas.