The increasing complexity and interdependence of urban physical, social, environmental, and infrastructural systems have made conventional approaches to urban resilience assessment increasingly insufficient for capturing the multidimensional and interconnected nature of urban vulnerability. Urban resilience is not determined by a single physical or socioeconomic factor; rather, it emerges from the interaction of infrastructure accessibility, population characteristics, spatial structure, environmental exposure, institutional capacity, and the continuity of critical urban functions. Consequently, effective resilience assessment requires analytical frameworks capable of integrating heterogeneous data sources, accounting for multiple dimensions of resilience, and identifying complex relationships among urban indicators.
Recent advances in geographic information systems, urban data infrastructure, and machine learning have created new opportunities for data-driven resilience assessment. Machine learning methods can capture nonlinear relationships and interactions that may not be adequately represented through conventional weighted indices. Nevertheless, an important methodological challenge remains: when machine learning models use a composite resilience index as their target variable, they may merely reproduce the mathematical structure created through the weighting procedure rather than provide independent insights into urban risk. This limitation reduces the added analytical value of machine learning and may obscure the distinction between the importance of an indicator in constructing a composite index and its contribution to functional risk.
From a civil defense perspective, resilience assessment should also extend beyond general resilience scoring toward identifying conditions that may lead to disruption of critical urban functions. Access to critical infrastructure, emergency services, transportation networks, communication systems, backup capacities, and the spatial distribution of vulnerable populations are particularly important in reducing the consequences of disruptive events. Therefore, there is a need for an integrated framework that simultaneously provides a composite measure of urban resilience and independently analyzes the determinants of functional risk.
Accordingly, this study aims to develop an intelligent framework for urban resilience assessment by integrating multi-source urban data, a hybrid Analytic Hierarchy Process–Principal Component Analysis (AHP–PCA) weighting approach, machine learning, and civil defense principles. The main research question is: How can multi-source urban data and machine learning be integrated into an independent and interpretable framework for assessing resilience and identifying factors associated with functional risk from a civil defense perspective?

Methodology
The study adopts an applied-developmental and descriptive-analytical approach based on a data-driven analytical framework. Multi-source data representing the physical, social, infrastructural, environmental, spatial, and managerial dimensions of urban resilience were organized at the level of urban analytical units. Following data quality control, preprocessing, standardization, and integration, twelve base indicators (I1–I12) were used to represent key dimensions of resilience and vulnerability.
A composite resilience index was first constructed using a hybrid AHP–PCA approach. In the AHP stage, strategic weights were determined through pairwise comparisons based on the relevance of indicators to vulnerability reduction, protection of critical infrastructure, continuity of urban functions, and resilience enhancement. In the PCA stage, data-driven weights were derived from the statistical structure of the standardized indicators using eigenvalues and factor loadings. The two weighting components were then combined with equal contributions (α = 0.5), followed by normalization. Sensitivity analysis was conducted to examine the stability of the composite index under alternative values of α.
To prevent the machine learning stage from directly reproducing the mathematical structure of the AHP–PCA index, functional risk was independently determined rather than being derived from the composite resilience score. Functional risk was assessed using complementary information, including Monte Carlo simulation of potential disruption of critical services, expert judgment and sensitivity analysis, and threat-related scenarios consistent with civil defense requirements. Urban units were subsequently classified into three functional risk levels: low, medium, and high.
The twelve base indicators were then used as input features, while functional risk level was used as the independent target variable. The composite resilience score and its derived classes were excluded from the machine learning feature set. Four algorithms—Logistic Regression, Decision Tree, Random Forest, and XGBoost—were compared. Model performance was evaluated using Accuracy, Precision, Recall, F1-Score, and AUC-ROC. Following model selection, SHapley Additive exPlanations (SHAP) were employed to examine the relative importance and direction of the effects of the input indicators. In addition, the ranking of indicators based on SHAP importance was compared with their ranking according to the final AHP–PCA weights, and Spearman rank correlation was used as a complementary measure of alignment between the two rankings.