Machine learning maps landslide danger across Ethiopia’s rift valley

Deep in the highlands fringing Ethiopia’s Great Rift Valley, the ground is quietly on the move. Steep slopes, seasonal downpours, fractured volcanic rock and expanding road networks combine to make the Main Ethiopian Rift one of the most landslide-prone landscapes in East Africa, and every rainy season brings new scars to hillsides, highways and farmland. Now a research team based at Arba Minch University has shown that a carefully built artificial intelligence pipeline can pinpoint where the next failures are most likely to occur, with a level of spatial precision that could reshape how the region plans its roads, settlements and disaster response.

The study, published in the journal Discover Sustainability, took on a deceptively simple question: given what we know about a landscape, can an algorithm learn to recognize the conditions that produce landslides? To answer it, the researchers compiled an inventory of 770 landslide locations drawn from field campaigns and satellite imagery interpretation, then trained four different machine learning models to distinguish unstable terrain from stable ground. The models tested were Random Forest, Support Vector Machine, XGBoost and a mean-probability ensemble that averaged the predictions of the individual algorithms.

Each model received the same ten conditioning factors as inputs, spanning the physical and human dimensions of slope stability: elevation, slope angle, lithology, land use and land cover, distance to roads, and other terrain attributes that geologists have long associated with slope failure. The dataset was split so that 70 percent of the mapped landslides trained the algorithms while the remaining 30 percent served as an independent test, a standard practice designed to reveal whether a model has genuinely learned the physics of instability or merely memorized its training examples.

The results were striking. Random Forest, an algorithm that builds hundreds of decision trees and pools their votes, achieved a ROC-AUC of 0.9516, the highest of the four, followed by the ensemble at 0.9301, XGBoost at 0.9192 and the Support Vector Machine at 0.8344. Success-rate analysis, which measures how well a model ranks the landscape from most to least dangerous, told a similar story: Random Forest scored 0.981 and XGBoost 0.973, comfortably ahead of the ensemble at 0.931 and the Support Vector Machine at 0.837.

But raw accuracy scores only tell part of the story, and the team went further by asking a more practical question: how much of the actual landslide activity does each model concentrate into its highest-danger zones? The answer revealed meaningful differences between the algorithms. Random Forest placed 675 landslides, or 87.66 percent of the inventory, inside its very-high-susceptibility class, which covered just 10.79 percent of the study area, producing a landslide density of 3.3501 landslides per square kilometer. XGBoost captured an even larger share, 90.78 percent, but needed 21.48 percent of the area to do so, effectively flagging more terrain as dangerous.

The ensemble model offered a different kind of value. By averaging the probability outputs of the individual algorithms, it produced a more balanced susceptibility map, containing 86.36 percent of landslides within 14.11 percent of the study area. Where Random Forest delivered the sharpest spatial discrimination and XGBoost the broadest capture, the ensemble landed in between, smoothing out the idiosyncrasies of any single learner. For planners who must weigh the cost of over-warning against the risk of under-warning, that balance may matter as much as raw precision.

Geographically, the high and very-high susceptibility zones clustered in the central, south-central, western and northern sectors of the study area, patterns that align with the field evidence documented by the researchers. Their photographic appendix records rotational slides and earth flows tearing through residual soils, roads that have settled vertically by 3.5 meters where water seeped beneath the pavement, retaining walls displaced 26 meters downslope, and debris masses that blocked entire road sections. Springs and seepage appear repeatedly at the failure sites, underscoring how rainfall infiltration triggers movement in weathered basalt and rhyolite terrain.

Beyond mapping danger, the study asked which factors the algorithms actually relied on, and the feature-importance analysis delivered a clear hierarchy. Elevation emerged as the dominant control, followed by lithology, land use and land cover, slope and distance to roads. That ranking carries a pointed message: while geologists cannot change the elevation or rock type of a hillside, land cover and road construction are human decisions, meaning that a substantial share of landslide risk in the rift is within the region’s own power to manage. Deforestation, agricultural expansion on steep slopes and poorly engineered road cuts all feed directly into the susceptibility the models detected.

The implications reach well beyond Ethiopia. Landslide susceptibility mapping has become a global priority as climate change intensifies rainfall extremes and as infrastructure expands into mountainous terrain, and the Main Ethiopian Rift offers a demanding test case where steep topography, variable lithology and rapid land-use change interact. The authors frame their work as a contribution to the United Nations Sustainable Development Goals, particularly SDG 11 on sustainable cities and communities, SDG 13 on climate action and SDG 15 on life on land, arguing that identifying landslide-prone areas before disaster strikes is far cheaper than rebuilding roads and communities afterward.

For the researchers, the takeaway is twofold. Random Forest provided the strongest overall spatial discrimination, making it the algorithm of choice when the goal is to concentrate attention on the smallest possible high-danger footprint. The ensemble, meanwhile, produced a comparatively balanced representation of susceptibility that may serve better for regional planning, where a more conservative map helps avoid both false alarms and missed hazards. Either way, the study demonstrates that machine learning, fed with field data and satellite observations, can convert the scattered evidence of past slope failures into actionable foresight, giving communities in the rift and beyond a fighting chance to plan around the terrain rather than react to it.

Subject of Research: Earth Sciences

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

Article References: Getahun, E., Jothimani, M., Shano, L., Dawit, Z., Regassa, H., Oyda, Y., Ekaso, D., & Gissila, B. (2026). A multi-algorithm machine learning approach for landslide susceptibility assessment using ensemble prediction and feature importance analysis in the Main Ethiopian Rift. Discover Sustainability. https://doi.org/10.1007/s43621-026-04851-w

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04851-w

Keywords: landslides, machine learning, Random Forest, Ethiopian Rift, geohazards, XGBoost, susceptibility mapping, disaster preparedness, multi-algorithm, machine, learning, approach