Authors - Virgel William Afaga, Patrick Andrew Balang, Dana Wynnette Binwag, Emmanuel Paolo Bromeo, Mark John Bumacod, Carl Allan Calsiman, Juliana April Cendana, Roderick Makil,Dulthe Carlo Munar Jr. Abstract - Benguet Province, Cordillera Administrative Region, Philippines, is highly susceptible to landslides due to its rugged topography, complex geology, frequent typhoon tracks, and extensive mining and road construction. Existing hazard maps rely on static statistical methods and coarse rainfall averages that cannot capture the dynamic triggering conditions of individual storm events. This paper presents a dynamic landslide susceptibility model built on Random Forest (RF) and Extreme Gradient Boosting (XGBoost) trained on thirteen environmental conditioning factors across five domains (topographic: elevation, slope, aspect, distance to streams; geological: rock type, soil type; land cover: LULC, NDVI, NDWI; climatic/hydrological: mean annual rainfall, event rainfall, antecedent rainfall; anthropogenic: distance to roads) derived from high-resolution satellite imagery and event-specific rainfall records. Training used a balanced 16,158-sample dataset (50:50 landslide/non-landslide) from the MGB-CAR geohazard inventory, split 60:20:20 for training, validation, and testing. XGBoost outperformed RF on all metrics: AUC-ROC 0.8903, accuracy 81.78%, precision 81.87%, recall 81.62%, and F1 81.75%; the performance difference was statistically significant (McNemar's test: χ² = 6.22, p < .013). Spatial validation via the Seed Cell Area Index (SCAI) confirmed that High and Very High susceptibility classes captured 69.87% of inventoried landslides within only 36.3% of the provincial area. Expert review by four MGB-CAR geoscientists yielded Likert mean scores above 4.0 for conditioning factor appropriateness, inventory quality, and feature importance plausibility. A fully automated monthly update pipeline was deployed—completing the full cycle from remote-sensing data retrieval to web-map publication in approximately 31 minutes—demonstrating operational feasibility for dynamic hazard monitoring using open-source tools and free satellite data.
Open Zoom