Distinguishing compound and cumulative hazards using machine learning and fuzzy logic in multi-hazard susceptibility mapping

This study presents an integrated multi-hazard susceptibility assessment for a mountainous region in northern Iran, focusing on four major hazards: flood, avalanche, rockfall, and landslide. Three machine learning models Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Machine (SVM) were applied to model single-hazard susceptibility using 21 topographic, climatic, geological, land-cover, and proximity-related variables at 30 m spatial resolution. Model performance was evaluated using ROC-AUC, accuracy, precision, recall, and F1-score based on field-validated hazard inventories. The results demonstrate that RF achieved the highest predictive performance for flood (90.96%) and rockfall (91.63%), while SVM showed competitive performance for avalanche (85.70%) and landslide (86.20%) susceptibility, reflecting differences in dominant hazard processes and model sensitivity to nonlinear and threshold-based relationships. To extend beyond single-hazard assessment, a Fuzzy Logic-based integration framework was implemented using AND, OR, and GAMMA operators to explicitly represent alternative hazard interaction mechanisms. The AND operator identified spatially constrained but high-confidence compound hazard hotspots, whereas the GAMMA operator captured broader zones of cumulative susceptibility driven by shared environmental controls. Comparative analysis indicates that fuzzy integration improves overall multi-hazard prediction, with the AND operator achieving the highest predictive reliability (ROC-AUC = 92.20%, F1-score = 93.83%), while GAMMA provides a more generalized but spatially comprehensive susceptibility pattern. The proposed framework highlights the importance of combining model selection with integration logic and provides a robust basis for regional-scale hazard assessment and spatial planning in complex mountainous environments.