A domain-informed optimization and machine learning framework for land subsidence susceptibility mapping and building exposure assessment: A case study of Tehran, Iran

Abstract

Land subsidence (LS) is a major geohazard causing irreversible damage to urban infrastructure and substantial socioeconomic and environmental impacts. This study presents a domain-informed optimization and machine-learning framework that combines Genetic Algorithm (GA), Bayesian-guided Genetic Algorithm (BayGA), and Marine Predators Algorithm (MPA) optimization with Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGB) models to map LS susceptibility (LSS) in the Tehran metropolitan area, Iran. A dual-resolution grid sampling was employed to automatically generate subsidence and non-subsidence samples from an open InSAR-derived inventory data. Twenty-six environmental and anthropogenic conditioning factors representing the principal controls on LS were incorporated. Rather than optimizing all predictors, the proposed framework retains essential process-related variables as baseline features and optimizes only the remaining variables using metaheuristic feature selection. This strategy preserves physically meaningful information while improving model interpretability and robustness over conventional data-driven feature-selection approaches. The results showed that all models achieved high predictive performance (ROC–AUC > 0.96) under the adopted random point-level validation strategy. BayGA-XGB achieved the highest accuracy (96%) and F1-score (95%), although its advantage over the other leading XGB- and RF-based models was marginal (0.2 to 0.8 percentage points in F1-score). SHapley Additive exPlanations (SHAP) identified elevation, sediment properties, geology, fault proximity, and groundwater-level decline as the dominant predictors. Overlay with Microsoft Building Footprints showed concentrated exposure in the southern and central districts. Groundwater trends for 2012–2019 provided a hydrogeological consistency check for the physical plausibility of the mapped susceptibility patterns. The framework enables interpretable regional LSS and exposure analysis for urban planning in Tehran and comparable metropolitan settings.