Integrating social vulnerability into high-resolution global flood risk mapping

The study addresses the gap in global flood risk mapping where social vulnerability is often ignored, resulting in maps that primarily reflect exposure of people or assets rather than risks to human well-being. Flooding affects billions globally, with low- and middle-income countries facing the greatest exposure and lowest coping capacity. Existing global models and maps, though increasingly detailed, typically do not incorporate spatial variation in vulnerability. The authors propose integrating social vulnerability—proxied by relative deprivation (poverty)—into high-resolution fluvial flood risk estimation to better identify hotspots where large populations with low coping capacity face flood hazards. The purpose is to support decision makers in prioritising adaptation and response under scarce resources by providing a vulnerability-adjusted risk index (VARI Flood) at 90 m resolution that can be applied globally and within countries.

Prior global flood risk studies commonly combine hydrological model outputs with gridded population, income, or asset data to estimate exposure, effectively producing exposure maps rather than risk maps. Social vulnerability, defined as susceptibility to harm due to social factors (e.g., poverty, status), is often omitted or only incorporated at coarse spatial scales (national or large subnational units). For instance, some studies assume uniform income within large administrative units or provide national-level vulnerability-resilience assessments, limiting local decision utility. Recent advances in global flood hazard and population modeling enable high-resolution mapping, but a lack of fine-scale, globally consistent vulnerability data remains a major limitation. The authors build on this literature by incorporating relative deprivation proxies (gridded GDP per capita and a multidimensional deprivation index) at high spatial resolution to adjust exposure-based estimates and produce a more human well-being–oriented assessment of flood risk.

• Study scope: 175 countries with available gridded GDP and poverty data, covering 98.6% of global population.
• Hazard modeling and AEP: Used Version 2 of the University of Bristol/Fathom Global Flood Model. Simulated fluvial flood depths at ~90 m (3 arc-seconds) for 10 return periods (5, 10, 20, 50, 75, 100, 200, 250, 500, 1000 years). Converted return periods to annual exceedance probability (AEP). For each pixel, classified flooded cells as those with depth >10 cm; assigned the maximum AEP (most frequent probability) across the stack to create a single AEP map.
• Population data: WorldPop (building-constrained) 2020 population counts at ~90 m, adjusted to UN totals. Constrained data were chosen to avoid overestimation of exposure common in unconstrained products.
• Social vulnerability proxies: (1) Gridded GDP (Chen et al.) at 1 km² (30 arc-seconds) for 1992–2019, converted to per capita by dividing areal GDP by WorldPop population; (2) Global Gridded Relative Deprivation Index (GRDI; multidimensional) at 1 km². Due to limited GRDI coverage (11% of GDP cells), GDP per capita is the primary proxy.
• Harmonizing resolutions: Up-sampled 1 km² GDP-based layers to ~90 m via nearest neighbor to match AEP and population grids.
• Expected Population Exposure (EPE): Computed as AEP × population for each cell (representing integrated exposure over a spectrum of probabilities, rather than a single return period threshold).
• Relative deprivation construction: Ranked cell-level GDP per capita into quintiles. For global comparison, used the global income distribution; for within-country analysis, used country-specific distributions. Inverted GDP quintiles so higher quintile implies higher vulnerability (poorer areas are more vulnerable).
• VARI Flood score: For each cell, divided EPE and vulnerability into quintiles, then computed VARI Flood Score = sqrt(EPE_quintile × Vulnerability_quintile), rounded to the nearest integer. Scores range 1 (lowest risk) to 5 (highest risk). Global applications use globally standardized quintiles; national applications use country-standardized quintiles.
• Aggregation and mapping: Results reported at multiple scales: cell (~90 m), admin-2 units (counties/districts), countries, and world regions. Country examples include Nigeria, Pakistan, and Vietnam.
• Tools: R used to prepare AEP rasters; Python 3.0 for analysis; QGIS for mapping. Replication code and datasets are provided via Figshare and data repositories.
• Sensitivity analyses: Explored alternative flood depth thresholds (50 cm, 100 cm) in case study countries; results confirm that incorporating vulnerability consistently shifts risk distributions even as absolute exposures change.

• Exposure baseline: ~2 billion people are exposed to a fluvial flood event using a 10 cm depth threshold.
• Global top-quintile comparisons (globally standardized):
  • EPE-only: 1.23 billion people live in areas in the top two risk quintiles (quintiles 4 and 5; totals 388 + 840 = 1,228 million).
  • VARI Flood (income-based vulnerability): 866 million in top two risk quintiles (568 + 298 million).
  • Using the multidimensional deprivation proxy (limited coverage): 867 million (EPE) vs 373 million (VARI Flood) in top two quintiles.
• Geography of risk shifts: Many densely populated or high-income regions (e.g., Australia, Europe, North America, Russia) drop in relative risk when vulnerability is incorporated.
• Administrative unit shifts: The highest-risk admin-2 unit changes in 75 of 175 countries (43%) under global standardization; with country-standardized quintiles, the highest-risk admin-2 changes in 73 countries (42%).
• Distributional change: Globally, EPE-only risk is monotonic with over half the population in high/very high risk (quintiles 4–5), while VARI Flood yields a more normal distribution with most people in lower three quintiles, reflecting the emphasis on socially vulnerable areas.
• Regional patterns: Incorporating vulnerability redistributes assessed risk toward regions with higher social vulnerability, altering cross-regional comparisons.
• Country examples:
  • Nigeria: Exposure-only maps highlight extensive risk along the Niger basin/delta and northern regions; VARI Flood reduces relative risk in the south and concentrates risk in the far northeast.
  • Pakistan: Similar overall patterns between methods, with a slight increase in high-risk populations in southern regions under VARI Flood.
  • Vietnam: Both approaches show high risk nationwide, but VARI Flood emphasizes hotspots in the Mekong River Delta and Central Coast.
• Resolution constraints observed at settlement scale: Coarser 1 km² vulnerability inputs produce visible tiling at 90 m mapping, partially limiting fine-grained differentiation.

Incorporating social vulnerability at high spatial resolution changes the perceived geography of flood risk by shifting emphasis from purely exposed or asset-dense areas to places where large populations face low coping capacity, better aligning risk estimates with potential human well-being losses. VARI Flood facilitates ethical and practical decision-making by making explicit trade-offs between protecting dense populations, valuable assets, and minimizing well-being losses among vulnerable groups. While intended as an illustrative global application rather than definitive national tallies, the approach demonstrates robust reweighting of risk distributions under multiple assumptions and thresholds. Uncertainties in flood hazard models can be substantial locally but tend to diminish when aggregating to coarser units (≥1 km² or admin-2). Incorporating additional hazards (pluvial, small rivers <50 km², coastal, and compound events) would change risk estimates but also add uncertainty, especially for pluvial models. The approach underscores the need for improved, high-resolution, and context-consistent social vulnerability datasets to enhance accuracy, particularly in rapidly urbanizing LMIC regions where intra-urban vulnerability heterogeneity is high. VARI Flood complements traditional exposure/asset-centric methods by foregrounding potential well-being losses and enabling multi-scale prioritization.

The study introduces VARI Flood, a 90 m resolution vulnerability-adjusted index that integrates fluvial flood exposure (AEP × population) with relative deprivation proxies to produce relative risk scores at global and national scales. The index reveals substantial shifts in the distribution and geography of risk when social vulnerability is considered, improving the identification of hotspots where well-being losses may be greatest. This framework supports resource prioritization for adaptation, mitigation, and response, especially in data-scarce contexts. Future research should: (1) improve the resolution and consistency of social vulnerability indicators; (2) assess and reduce uncertainties through multi-hazard integration (pluvial, coastal, small rivers, and compounding effects); (3) refine sensitivity to flood depth thresholds and local adaptation; and (4) further evaluate interactions between hazard, exposure, and socio-economic data uncertainties across scales.

• Social vulnerability data resolution: GDP and multidimensional deprivation data are at 1 km², coarser than ~90 m hazard and population grids, leading to tiling artifacts and potential underestimation of intra-urban variation.
• Proxy limitations: Poverty (relative deprivation) is an imperfect proxy for social vulnerability; factors like age, gender, race/ethnicity, governance, and emergency access are not captured globally.
• Hazard scope: Only fluvial hazards modeled; exclusion of pluvial, small rivers (<50 km²), coastal flooding, and compound events may omit important risks.
• Threshold sensitivity: The 10 cm flood depth threshold may not reflect local adaptation; results are sensitive to chosen thresholds, though the vulnerability-adjusted redistribution effect is consistent.
• Model uncertainties: Known uncertainties in large-scale flood hazard models can affect local predictions; interactions with socio-economic data uncertainties warrant further study.
• Data coverage: Multidimensional deprivation dataset has limited geographic coverage (11% of cells vs GDP), constraining alternative vulnerability analyses.
• Resampling effects: Nearest-neighbor upsampling of 1 km² socio-economic data to ~90 m may introduce inaccuracies at fine scales.
• Assumptions of relative measures: Global vs country-standardized quintiles affect comparability and prioritization depending on the decision context.