Mapeo de susceptibilidad a deslizamiento mediante comparación de evaluación multicriterio y machine learning en los Andes Ecuatorianos

Abstract

CONDITION FOR PUBLICATION OF PROJECT. Landslides represent a significant hazard in the mountainous regions of the Ecuadorian Andes, where their occurrence is conditioned by the geomorphological, geological, and anthropogenic characteristics of the terrain. This study evaluates landslide susceptibility in a region of the Andes in Ecuador by applying three methodologies: a hybrid spatial multicriteria evaluation (SMCE) approach that combines the Analytic Hierarchy Process (AHP) and Frequency Ratio (FR); the Random Forest (RF) machine learning algorithm; and its optimized version using Principal Component Analysis (PCA). A landslide inventory was built, and 14 conditioning factors derived from open-access data were integrated, maintaining a spatial resolution of 3 meters. Model validation was performed using ROC curves and statistical metrics. The results show that the best-performing model is RF with the original conditioning factors (AUC = 0.947, accuracy = 94.86%), mainly influenced by road and drainage distance. On the other hand, although the hybrid SMCE model presented lower performance (AUC = 0.843), it offers greater interpretability through the relative weight of each factor. The combination of PCA with the RF algorithm reduced dimensionality while maintaining good accuracy (AUC = 0.906). All models consistently identify terrain modifications in urban and rural areas as zones of high susceptibility. This work demonstrates that RF with original conditioning factors is the most effective technique for landslide mapping in complex geodynamic contexts. Keywords: Landslide susceptibility, Geological Hazard, Random Forest, Principal Components, Geospatial Multicriteria Evaluation.