Understanding the global subnational migration patterns driven by hydrological intrusion exposure

Migration is a key mechanism by which people avoid risks linked to natural disasters, poverty, and conflict, seeking safety and better living conditions. As climate change intensifies the hydrological cycle (e.g., flooding, extreme precipitation, sea-level rise), migration pressures are expected to increase, aligning with SDG 11 and 13 priorities. This study conceptualizes hydrological intrusion as incremental surface hydrological events (floods, sea-level rise, storm surge) and examines how these affect subnational migration. While climate can catalyze movement, migration is a complex process governed by economic, social, cultural, and political factors. Some literature reports limited or context-dependent impacts of hydrological risks on migration; others show that resource scarcity can trap populations in high-risk areas. Following IPCC/UNDRR, climate risk is framed by hazard, exposure, and vulnerability. Prior work has focused more on hazards, with mixed or biased measures, and often at national scales. Moreover, migration responses can be nonlinear and threshold-dependent, but many studies use linear models. To fill these gaps, this study estimates population-weighted hydrological intrusion exposure (HIE) from satellite-observed surface water dynamics and derives subnational migration rates (Mig_R) from population stocks net of natural growth. It then applies interpretable nonlinear ensemble models with country-year fixed effects to analyze global subnational migration responses to hydrological intrusion.

Prior studies have linked sea-level rise and flooding to migration via direct and indirect damages (e.g., loss of housing and infrastructure, salinization, agricultural impacts), and have estimated substantial displacements associated with rainfall anomalies. However, findings vary due to heterogeneity across places and populations; some evidence suggests environmental risks do not uniformly drive migration, with adaptation and immobility common in resource-constrained settings. Risk assessment frameworks highlight hazard, exposure, and vulnerability, yet empirical work has emphasized hazards using proxies such as extreme precipitation deviations, disaster counts, and floodplain area. These approaches can misrepresent risk where high rainfall/floodplains coincide with productive livelihoods, and disaster databases exhibit reporting biases (e.g., higher losses in wealthier economies). Exposure estimates (e.g., population in LECZ or floodplains) also diverge due to heterogeneous population distributions. Migration analyses have increasingly used bilateral datasets but largely at country level, despite climate-induced moves often being short-distance and internal. Many models are linear, potentially missing threshold effects and nonlinear dynamics. This study complements prior work by constructing a globally consistent, population-weighted exposure index from remote sensing, analyzing subnational units worldwide, and explicitly modeling nonlinear effects and neighborhood contrasts in exposure (HIE_D).

Study scope: 249 countries/regions, 46,776 subnational units (mostly municipal/county level). Migration estimation: Subnational migration is computed from annual changes in population stock (WorldPop, ~100 m resolution) after removing natural growth (birth and death rates from World Bank). Stocks are adjusted to align with UN DESA country totals. Migration rate (Mig_R) is migration divided by total population; group-specific Mig_R computed for minor (0–20), adult (20–65), elder (65+), male, and female. To limit outlier influence in small-population units, 0.1% Winsorization is applied. Exposure (HIE): Using Landsat-derived global surface water dynamics (30 m), resampled to 100 m to match population data, the authors compute a population-weighted hydrologic exposure index per unit for 2015–2020 (2015 excluded due to data gaps in some steps). Null water pixels are mirrored as null in population rasters to avoid underestimating exposure. HIE is Pop-weighted mean of surface water occurrence; HIE_D quantifies difference between a unit’s HIE and its neighboring units. Hazard: Hydrological hazard is proxied by Rx5day (annual maximum 5-day precipitation total) from ERA5-Land, a globally consistent measure of extreme precipitation related to flood risk. Vulnerability: Environmental/geographic vulnerability includes average annual temperature (TEMP), precipitation (PRCP), elevation (DEM from SRTM), and mean distance to water (Water_Dis). Socioeconomic vulnerability includes night-time lights per area (NTL_PA), GDP per area (GDP_PA) from calibrated gridded GDP, urbanization rate (Urban_R), and average years of education (Edu). Modeling: An interpretable nonlinear ensemble learning framework is used: LightGBM for prediction and SHAP for interpretability. Country-year fixed effects are included to mitigate unobserved confounders at those levels (e.g., religion, social/humanities, national policies). Model performance and feature effects are analyzed globally and within income-group and demographic subgroup splits. Validation: Aggregated subnational migration estimates are compared to WPP2022 country-level migration, yielding R^2 = 0.89, supporting reliability for subnational inference, while noting inability to recover OD flows or migration durations due to stock-based estimation.

• Global subnational migration patterns: Aggregating subnational estimates to countries yields R^2 = 0.89 versus WPP22, validating the approach. Average global Mig_R is −0.59%, implying net out-migration is common though small rates produce large impacts at scale. Oceania shows extensive out-migration; Europe and America are major destinations; hotspot analysis identifies out-migration clusters in Middle Africa, Central Asia, Southern Europe.
• Exposure drives migration: SHAP-based importance ranks exposure among the top drivers. Increasing HIE tends to increase out-migration at low-to-moderate exposure levels but with diminishing marginal effects. When HIE < 6.74%, increasing HIE by 1 SD (1.51%) raises out-migration by about 0.09% (95% CI −0.11% to −0.06% per regression curve). Above HIE ≈ 6.74%, the outward migration effect attenuates or reverses, consistent with resource constraints and entrapment.
• Neighborhood contrast intensifies effects: HIE_D (a unit’s exposure relative to neighbors) substantially amplifies migration responses. When HIE_D > 0, a 1 SD increase in HIE_D (127.42%) is associated with a 0.34% rise in Mig_R (95% CI ≈ −0.36% to −0.33% by curve), indicating stronger sensitivity when safer nearby alternatives exist.
• Socioeconomic factors vs exposure: Although socioeconomic variables often appear relatively important, their marginal effects are smaller globally: 1 SD increases in GDP_PA and Edu correspond to average Mig_R changes of 0.05% (95% CI 0.04–0.06%) and 0.06% (95% CI 0.05–0.007%), respectively. For HIE_D, the average Mig_R change is 0.16% (95% CI −0.17 to −0.15%), exceeding socioeconomic effects. GDP_PA shows a leap-frogging mechanism: a strong attraction when exceeding domestic average, then quickly flattening.
• Income-group heterogeneity: High-income regions show resilience; at high exposure (HIE ≥ 2.40%), inward migration occurs, likely due to strong local economies (GDP_PA > mean by 0.30 SD) and connectivity. HIE_D effects are pronounced, consistent with flexible relocation across well-connected urban systems. Middle-income regions mirror global nonlinearity with a lower suppression threshold (HIE < 5.00%): a 1 SD rise in HIE lowers Mig_R by ~0.09% (95% CI −0.09% to −0.08%), and they are comparatively insensitive to HIE_D. Low-income regions display oscillatory responses consistent with constraints; HIE_D is especially influential (relative importance 10.52%). When local HIE exceeds neighbors, a 1 SD increase in HIE_D (168.14%) decreases Mig_R by 0.47% (95% CI −0.49% to −0.45%), indicating predominant short-distance or neighborhood migration under resource limitations.
• Demographic heterogeneity: Median Mig_R by age: minors −1.41%, adults −0.71%, elders +2.23% (outflows of younger groups toward select areas; older groups show net inflows/return). Seniors’ HIE effects diminish around zero at HIE > 0%, and HIE_D effects hover near zero, implying entrapment/low mobility among the elderly. By sex, HIE/HIE_D effects are marginally stronger in men (importances: HIE 14.82%, HIE_D 13.04%), consistent with higher earnings responsiveness.
• Nonlinear thresholds and S-shaped response: Both exposure contrast (HIE_D) and hazard (Rx5day) exhibit thresholds beyond which out-migration increases: HIE_D tipping around −37.15% and Rx5day around 135.66 mm. Migration responses follow an S-shaped trajectory: initial resistance, then adaptation with increased mobility once thresholds are crossed, and eventual plateau/entrapment as risks intensify (plateau when HIE ≲ 6.74%).
• Policy-relevant implications: While economic improvements reduce out-migration up to national-average levels (≈0.40% Mig_R improvement within ±1 SD around mean GDP density), sustained reductions in out-migration are better supported by mitigating hydrological risks and reducing exposure disparities with neighboring regions.

The study directly addresses how hydrological intrusion shapes subnational migration by combining a new, population-weighted exposure metric with nonlinear, interpretable models controlling for country-year effects. Findings show that both absolute exposure (HIE) and relative exposure to neighbors (HIE_D) significantly influence out-migration, but effects are nonlinear and heterogeneous across income levels and demographics. Exposure disparities (HIE_D) emerge as especially potent levers, indicating that perceived alternative risks in nearby areas drive movement more than absolute risk alone. Although climate-related variables explain modest shares of variance overall—echoing literature that climate is a weaker predictor than socioeconomic factors—their effects can be decisive near thresholds, with substantial local impacts. The S-shaped response clarifies why linear models understate or mischaracterize environmental migration: people resist moving at low risk, migrate once thresholds are exceeded, and can become trapped as risks intensify and resources dwindle. Importantly, high-income areas can attract migrants despite high exposure due to resilience and economic pull, while low-income areas face a double burden of high vulnerability and limited mobility, leading to short-distance moves or immobility. These insights underscore the relevance of risk mitigation, regional equity in exposure, and tailored support for vulnerable groups to manage population displacement effectively.

This work offers the first global observational analysis of subnational migration responses to hydrological intrusion using a consistent, population-weighted exposure index from remote sensing and an interpretable nonlinear modeling framework. It reveals that exposure and especially exposure disparities with neighboring areas are key drivers of out-migration, operating through nonlinear, threshold-dependent dynamics that vary by income level and demographic group. The findings reframe expectations that economic development alone will retain populations, highlighting the long-term value of mitigating hydrological risks and narrowing inter-regional exposure gaps. Policy directions include strengthening early warning and preparedness, enhancing infrastructure and social support in disadvantaged regions, and prioritizing risk reduction alongside economic development to curb out-migration. Future research should integrate finer-grained mobility data (e.g., mobile phone traces) to recover bilateral flows and durations, extend the time horizon to capture long-term dynamics, and refine hazard representation beyond proxies to better resolve local disaster processes.

• Migration estimation relies on annual population stocks net of natural growth; it cannot recover origin-destination flows, migration duration (temporary vs permanent), or specific bilateral patterns.
• Hydrological hazard is proxied by Rx5day rather than comprehensive disaster records due to global reporting biases and under-coverage of small events.
• The analysis period is relatively short (primarily 2015–2020 due to remote sensing data gaps around 2015), limiting long-term inference on evolving dynamics.
• Potential residual confounding remains despite country-year fixed effects and extensive covariates (e.g., unmeasured cultural or policy factors).
• Small-population units can exhibit volatile rates despite Winsorization; neighborhood definitions for HIE_D may influence sensitivity analyses.
• Remote sensing resampling and null-pixel handling, while designed to avoid underestimation, may still introduce measurement uncertainties in exposure estimates.