A decade of satellite imagery, census records, and explainable machine learning has exposed an uncomfortable paradox on China’s southeastern coast: the region’s wealthiest cities are also its most fragile when typhoons strike. In a study published on 28 August 2026 in the International Journal of Disaster Risk Science, a team led by Xiaoliu Yang and corresponding author Lu Gao of Fujian Normal University, together with colleagues at Western Michigan University, Beijing Normal University, and the Karlsruhe Institute of Technology, built a multi-scale diagnostic framework for urban resilience to typhoon disaster chains and applied it across Fujian Province for 2010, 2015, and 2020. Their verdict is blunt. Provincial resilience inched upward, yet the core districts of Xiamen, Quanzhou, and Fuzhou—cities synonymous with China’s coastal economic boom—moved in the opposite direction, while low-resilience zones clustered and expanded around them.

The problem the researchers set out to quantify is not the typhoon itself but the chain reaction it unleashes. A landfalling storm typically arrives as a compound package of strong winds, torrential rainfall, and storm surge, which then cascades into secondary hazards—riverine flooding, urban waterlogging, and infrastructure failure—that propagate across scales and amplify losses. Climate projections indicate that although future typhoon frequency remains uncertain, maximum wind speeds and rainfall intensity are expected to increase, raising the likelihood that extreme rain and surge will strike concurrently. Fujian offers a stark test case. Its 3,324-kilometer coastline sits squarely in the East Asian monsoon belt, and between 2000 and 2020 typhoons caused an average of 12.74 billion yuan in direct economic losses annually—58.3 percent of the province’s total natural-hazard losses. More than 76 percent of those typhoon-related losses stemmed from compound disaster chains rather than any single hazard, precisely the conditions under which conventional single-hazard risk assessments break down.

To measure resilience against such cascades, the team defined typhoon disaster chain urban resilience, or TDCUR, organized into three capacities: absorption, the ability to withstand the initial shock; coping, the ability to manage impacts as they unfold; and recovery and adaptation, the ability to bounce back and adjust over time. These capacities were operationalized through 21 indicators spanning terrain, vegetation, impervious surface coverage, drainage and flood-control infrastructure, economic conditions, and population—among them three chain-specific sensitivity indices for typhoon–rainstorm–waterlogging, typhoon–rainstorm–flooding, and typhoon–strong wind–storm surge sequences. All indicators were computed on 30-arc-second grids of roughly one kilometer, restricted to a fixed urban mask covering areas with at least 600 residents per square kilometer and totaling about 9,680.6 square kilometers, so every year was compared across identical territory. Indicator weights were specified in advance using an objective procedure combining GeoDetector-based spatial stratified heterogeneity statistics with recursive entropy weighting, and Monte Carlo perturbation of those weights quantified the uncertainty surrounding each score.

The framework then tackles a problem that plagues most resilience studies: the mismatch between fine-grained data and the administrative units where decisions are made. Grid-level scores were aggregated to counties, cities, and eight major river-basin units using a population-aware scheme. Each cell’s exposed population—area multiplied by population density—was converted into an effective exposure weight through an exponential penalty, with a coefficient of 0.7, that systematically down-weights cells whose baseline resilience is already high. Unit-level resilience is the exposure-weighted mean of its cells, scaled by the share of the unit’s census population living in exposed cells. The result is an index reflecting not just where resilience is weak but where weakness and human exposure overlap, giving provincial, municipal, and county authorities a common metric for cross-level planning.

The headline numbers show real but uneven progress. The provincial mean TDCUR rose from 0.4275 in 2010 to 0.4571 in 2020, a 6.9 percent gain, and areas classified as very high resilience expanded by 1,638.6 square kilometers. Kernel density estimation showed the distribution’s peak shifting from 0.47 to 0.51 before settling at 0.50, while its width narrowed by about 22 percent—evidence that regional disparities converged even as averages improved. Yet the spatial pattern is a tale of two provinces. Western inland cities, with lower urbanization intensity, richer vegetation cover, and more dispersed populations and assets, consistently posted higher absorption and coping scores. In the eastern coastal belt the picture inverted: the lowest values appeared not in poor or peripheral areas but in the dense urban cores of Fuzhou, Quanzhou, and Xiamen, where impervious surfaces cover more than 85 percent of the land and drainage systems were built to outdated standards. Core districts including Gulou and Taijiang in Fuzhou and Siming and Huli in Xiamen actually lost resilience over the decade.

The multi-scale aggregation sharpens the contrast. The five weakest counties in 2020 all lay in the east, averaging just 57.8 percent of the resilience of the strongest western counties such as Yongding District, which scored 0.5016. Among cities, Xiamen ranked last at 0.3252, while Putian posted the fastest improvement at +0.0166. On the watershed scale, the coastal directly discharging composite basin scored lowest at 0.3710, whereas the Jinjiang Basin gained +0.0269—roughly 1.6 times the provincial mean increase—hinting at the payoff of integrated river-basin management. Component analysis explains the mechanism: Jinjiang’s median absorption capacity of 0.103 sits 58 percent below the provincial maximum of 0.245, and coping capacity in Siming, Huli, and Jinjiang ran 40 to 50 percent below comparable inland areas by 2020. Recovery capacity remained highest along the coast, but most of its decade’s gains accrued inland—a diffusion of improvement that stopped short of the urban cores that need it most.

Hotspot analysis converted those gradients into statistically rigorous geography. Getis-Ord Gi* statistics revealed a landscape of clustered low resilience and dispersed high resilience: significant cold spots covered 447 square kilometers, or 4.6 percent of the urban domain, with more than 80 percent concentrated in the Xiamen–Quanzhou core, while hot spots shrank to a mere 18 square kilometers scattered across inland cities. Between 2010 and 2020, cold-spot area expanded by 48 percent as hot-spot area contracted by 51 percent—a deepening spatial polarization. Local Moran’s I analysis confirmed a persistent low–low clustering belt along the coast, though it weakened by 243.7 square kilometers over the decade. Spatiotemporal classification sorted the cold spots into three trajectories: persistent cold spots occupying 43.8 percent of classified area along the Xiamen–Zhangzhou–Quanzhou coastline, diminishing cold spots at 4.0 percent in recovering inland counties, and emerging cold spots at 2.4 percent on urban fringes—early warning zones where vulnerability is only beginning to condense.

The study’s methodological punchline is its use of explainable machine learning as a diagnostic layer. The team trained an XGBoost regression model—gradient-boosted decision trees with a maximum depth of 6, 600 estimators, a learning rate of 0.05, and L1 and L2 regularization of 0.5 and 1.0—on grid-level observations pooled from all three years, with the 21 indicators as features and Monte Carlo-averaged TDCUR as the response. Tuned by grid search with five-fold cross-validation, the model achieved a coefficient of determination of 0.9932 and a root mean squared error of 0.0062 on held-out data. The authors then applied SHapley Additive exPlanations, using the TreeSHAP algorithm to decompose every prediction into additive contributions from individual indicators. Three variables dominated: typhoon–strong wind–storm surge sensitivity, contributing 23.52 percent of mean absolute SHAP values; typhoon–rainstorm–flooding sensitivity at 15.72 percent; and impervious surface proportion at 10.06 percent. Segmented regression identified sharp slope changes—empirical turning points—at normalized values of −0.60 for surge sensitivity, −0.66 for flood sensitivity, and −0.79 for impervious surface, the last corresponding to roughly 85 percent imperviousness, beyond which resilience losses accelerate. The researchers stress these are features of the fitted response surface, not physical or regulatory thresholds, and the machine learning was strictly post hoc: it explains the index without redefining it.

Interaction analysis revealed that the top drivers do not act independently. Across all 231 indicator pairs, the strongest joint association belonged to the flood-sensitivity and surge-sensitivity pairing, with a mean absolute SHAP interaction of 1.88 × 10⁻³, accounting for 22 percent of the total among the top ten pairs; impervious surface interacting with flood sensitivity contributed another 13.8 percent, and with surge sensitivity 12.0 percent. Most interaction values were negative, meaning that where high sensitivities and high imperviousness coincide, predicted resilience drops further than either factor alone would suggest. Mapping summed absolute SHAP values per grid cell and aggregating them across scales produced an unambiguous target list: Huli, Taijiang, and Gulou districts all exceeded 0.18 at the county level; Xiamen, Quanzhou, and Fuzhou topped the city rankings; and the Jinjiang, Minjiang, coastal directly discharging, and Jiulongjiang basins averaged about 0.145. A robustness check using only five core indicators shifted the results by just 3 to 5 percent and preserved the spatial rankings almost perfectly, with a Spearman correlation of roughly 0.96.

For planners, the framework converts diagnosis into a spatial work list. In the coastal cores, where imperviousness exceeds 85 percent, the authors point to nature-based measures such as rain gardens and permeable pavements, combined with drainage-system upgrading, storm-surge defense enhancement, and impervious-surface management—precisely because the machine learning shows those pressures interacting. Inland, the priority is the opposite: protecting the ecological buffers and vegetation cover that underpin high absorption capacity, and siting regional emergency supply hubs, such as in Yongding District, where geography favors support for coastal disaster response. A dynamic monitoring system tracking persistent, emerging, and diminishing cold spots, with periodic SHAP reanalysis after major interventions, would let strategies be iterated over time—an approach tied to the UN’s Sustainable Development Goal 11 for resilient cities. The framework has limitations the authors acknowledge: proxy indicators cannot deliver engineering-grade precision, landslide cascades are not yet included, and spatial autocorrelation in the training data may inflate the model’s apparent fit, motivating spatial cross-validation in future work. But the method is transferable to any typhoon-exposed coastline, and its central message resonates far beyond Fujian: a decade of double-digit economic growth did not automatically buy safety, and in the very districts where coastal wealth concentrates, the machinery of resilience—drainage, buffers, and surge defenses—has quietly fallen behind the rising risks.