Spatial Assessment of Fire-Casualty Vulnerability in Industrial Complexes: Comparing AHP-Derived Weights with SHAP-Based Feature Importance

Fires in industrial complexes cause property damage and casualties, but the spatial conditions under which fires result in casualties are rarely assessed. This study proposes a two-stage framework that combines expert-based Analytic Hierarchy Process (AHP)–geographic information system (GIS) assessment with machine-learning feature-importance analysis of observed fires. Seven spatial indicators—four vulnerability indicators (building density, aged-building share, fire-engine access, and hazardous-facility proximity) and three reverse-coded mitigation indicators (fire-station response time, firefighting water-supply coverage, and evacuation-facility accessibility)—were defined with largely common scoring rules. In Stage 1, weights from an AHP survey of 62 experts were applied in GIS to map fire-casualty vulnerability on a 250 m grid in the Myeongji Noksan National Industrial Complex, Busan, Republic of Korea. In Stage 2, scores for the same indicators were assigned to 5417 factory fires recorded in Busan’s 41 industrial complexes during 2019–2023 (32.9% with casualties); four classifiers were trained to predict casualty occurrence, and Shapley additive explanations (SHAP) were computed for XGBoost, which achieved the highest recall (0.621). The map identified a localized high-vulnerability zone in the south-central part of the complex, where hazardous-facility proximity was concentrated, although recorded casualty fires were not clearly concentrated in these cells. SHAP ranked aged-building share and building density highest, whereas experts emphasized fire-engine access and response time. Comparing the two perspectives shows where expert emphasis and data-driven importance diverge and offers an exploratory basis for spatial prioritization and for re-examining indicator weights.