Work

A multi-algorithm machine learning approach for landslide susceptibility assessment using ensemble prediction and feature importance analysis in the Main Ethiopian Rift

Year: 2026

Type: article

Abstract: Landslides are a major geomorphological hazard in the Ethiopian highlands and margins of the Rift Valley, where steep terrain, seasonal rainfall, variable lithology, land-cover modification, and human activities interact to promote slope instability. This study assessed landslide susceptibility in the Main Ethiopian Rift using Random Forest (RF), Support Vector Machine (SVM), XGBoost, and a mean-probability ensemble model. A landslide inventory comprising 770 locations was prepared from field investigation and Google Earth interpretation, and ten conditioning factors were used as model inputs. The dataset was divided into 70% for training and 30% for independent validation, while model performance was evaluated using ROC-AUC, success-rate analysis, landslide-density validation, and factor-importance analysis. RF achieved the highest ROC-AUC of 0.9516, followed by the ensemble (0.9301), XGBoost (0.9192), and SVM (0.8344). Success-rate analysis similarly showed strong spatial-ranking performance for RF (0.981) and XGBoost (0.973), compared with the ensemble (0.931) and SVM (0.837). Landslide-density validation further demonstrated clear differences in spatial concentration among the models. RF concentrated 675 landslides (87.66%) within the very-high-susceptibility class, which covered only 10.79% of the study area, producing the highest density of 3.3501 landslides/km 2 . XGBoost captured the largest proportion of landslides in the very-high class (90.78%) but over a larger area (21.48%). In contrast, the ensemble provided a more balanced distribution, containing 86.36% of landslides within 14.11% of the study area. High- and very-high-susceptibility zones were concentrated mainly in the central, south-central, western, and northern sectors of the study area. Factor-importance analysis identified elevation, lithology, land use/land cover, slope, and road distance as the dominant controls on susceptibility. These findings demonstrate that RF provided the strongest overall spatial discrimination. At the same time, the ensemble produced a comparatively balanced susceptibility representation, providing useful information for landslide-risk reduction, infrastructure planning, and slope management in the Main Ethiopian Rift. By identifying landslide-prone areas and supporting risk-informed land-use planning, road-slope management, and disaster preparedness, the study contributes to the Sustainable Development Goals, particularly SDG 11 on sustainable cities and communities, SDG 13 on climate action, and SDG 15 on life on land.

Source: Discover Sustainability

DOI: 10.1007/s43621-026-04851-w