The Mount Marapi study integrated InSAR-derived deformation with geospatial variables and compared KNN, Random Forest, SVM and Gradient Boosting for landslide-susceptibility mapping. Subsidence and persistent slope fluctuations were observed, while slope, elevation and rainfall were the most influential variables.
Key findings
- Random Forest achieved 94.33% accuracy and 91.30% precision. Seismic activity and extreme rainfall were interpreted as additional triggers; curvature, soil type and local deformation had weaker individual influence but remained relevant in combination.
Why this matters globally
Satellite-machine-learning integration is transferable to hazard-prone regions with limited ground monitoring. A susceptibility map is not a time-specific forecast and must be connected to field observations and operational warning thresholds.
Thai researcher contribution
Khon Kaen University-affiliated authors contributed to an Indonesia-focused framework integrating InSAR and machine learning for landslide assessment.
Limitations to consider
The abstract does not specify spatial train-test separation, class imbalance or external validation. Accuracy may be optimistic if nearby locations occur in both training and testing. Susceptibility is not the same as exposure, vulnerability or societal risk.