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

This research investigates the differences between compound and cumulative hazards using machine learning and fuzzy logic for a mountainous area in northern Iran. The study focuses on four main hazards: flood, avalanche, rockfall, and landslide. Three machine learning models - Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Machine (SVM) - were used to create models for each hazard individually.

These models were trained using 21 variables such as topography, climate, geology, land cover, and proximity at a resolution of 30 meters.

The performance of the models was tested using ROC-AUC, accuracy, precision, recall, and F1-score, based on field-validated data. The results showed that RF performed best for flood and rockfall (90.96% and 91.63% respectively) while SVM had competitive performance for avalanche (85.70%) and landslide (86.20%), indicating different dominant processes and sensitivity to nonlinear and threshold-based relationships in each model.

To go beyond single-hazard assessment, a Fuzzy Logic-based integration framework was applied using AND, OR, and GAMMA operators to represent alternative hazard interaction mechanisms. The AND operator identified high-confidence compound hazard hotspots in specific areas, while the GAMMA operator captured broader zones of cumulative susceptibility due to shared environmental factors.

When comparing the two integration methods, the AND operator showed higher predictive reliability (ROC-AUC = 92.20% and F1-score = 93.83%) but GAMMA provided a more generalized but spatially comprehensive susceptibility pattern.

The study concludes that combining the selection of appropriate ML models with suitable integration logic can lead to robust multi-hazard assessment frameworks for complex mountainous environments.