Document Type : Original Research Paper
Authors
1 Department of Geospatial Information Systems , Faculty of Surveying Engineering, K. N.Toosi University of Technology, Tehran, Iran
2 Department of Physical Geography and Bolin Centre for Climate Research, Stockholm University, Stockholm, Sweden
Abstract
Background and Objectives: Iran's northern provinces (Golestan, Mazandaran, and Gilan) are simultaneously exposed to floods, landslides, and land subsidence, yet most prior studies have addressed these hazards in isolation. Existing susceptibility models often rely on single machine learning algorithms without systematic hyperparameter tuning and treat their predictions as opaque outputs, limiting their value for evidence-based decision-making. This research aims to develop an optimized and interpretable framework that produces joint susceptibility maps for all three hazards across this region, supporting integrated risk planning.
Methods: A hybrid model combining Support Vector Regression with the Two-phase Mutation Grey Wolf Optimizer (SVR-TMGWO) was built separately for each hazard. Twenty years of recorded flood and landslide events (2000–2020) were compiled from the Geological Survey of Iran and the national watershed-management authority, while subsidence locations were derived from Sentinel-1 InSAR analyses. Eighteen conditioning factors spanning topographic, hydrological, anthropogenic, and soil-physical variables were prepared for each pixel. The TMGWO algorithm simultaneously tuned the three SVR hyperparameters (C, ε, γ) for each hazard, producing distinct configurations that reflect the differing physics of each process. Model accuracy was evaluated using AUC-ROC, RMSE, R², and adjusted R² on independent test data. To overcome the black-box limitation of machine learning, Shapley Additive Explanations (SHAP) were computed at both global and local scales, exposing the contribution of each factor to the prediction. Finally, the three susceptibility maps were combined into an eight-class multi-hazard map.
Findings: TMGWO consistently improved test-phase AUC-ROC over the base SVR model, reaching 0.8404 for flood, 0.9329 for landslide, and 0.9642 for subsidence, while narrowing the training-test gap and indicating reduced overfitting. SHAP identified elevation as the leading driver across all three hazards but revealed contrasting secondary controls per process. The multi-hazard map showed that triple-overlap zones were spatially restricted and concentrated along the mountain-plain transition.
Conclusion: The proposed SVR-TMGWO framework, combined with SHAP interpretability, produced spatially explicit and physically meaningful susceptibility maps for three coexisting hazards in northern Iran. The resulting eight-class multi-hazard product highlights priority areas for integrated land-use planning and infrastructure protection. Future research should incorporate temporal dynamics, compare deep learning alternatives such as convolutional neural networks, and extend the framework to full risk assessment by including exposure and vulnerability components.
Main Subjects
)