Machine Learning Outperforms Classic Statistics in Mapping Deadly Himalayan Landslide Risk

On a 67-kilometre stretch of National Highway 7 in India’s Chamoli district, the road between Nandprayag and Vishnuprayag winds along the Alaknanda river through some of the most landslide-prone terrain on Earth. An average of 1.25 landslides occur per kilometre of this corridor, which forms part of the pilgrimage route to Badrinath. A new study published in Discover Geoscience has now compared three modelling approaches—Frequency Ratio, Random Forest, and XGBoost—to determine which best predicts where the next slope failure is likely to occur, and the results offer a clear verdict: machine learning wins.

The stakes could hardly be higher. Landslides rank among the world’s most destructive natural hazards, and a global analysis of fatal non-seismic landslide events recorded 4,862 such disasters between January 2004 and December 2016, killing 55,997 people, with Asia the hardest-hit continent. India carries a disproportionate share of the burden: roughly 0.42 million square kilometres, or 12.6 percent of the country’s land surface excluding snow-covered terrain, is susceptible to landslide hazard. In the Himalayan region alone, landslides claimed more than 5,000 lives between 2007 and 2015, and the June 2013 extreme rainfall disaster in Uttarakhand destroyed over 250 villages and killed approximately 6,000 people. The 2021 Chamoli disaster further devastated hydropower infrastructure, bridges, and settlements in the Rishiganga and Dhauliganga valleys.

The research team, led by Sidra of Jamia Millia Islamia in New Delhi, built their analysis on an unusually detailed foundation: a landslide inventory of 852 manually digitised landslide locations, compiled from Google Earth satellite imagery spanning 2012 to 2022. Each landslide was mapped as a polygon and then converted to presence points, which were split into a training set of 627 points and an independent testing set of 225. To give the machine learning classifiers the negative examples they need, the team generated 852 non-landslide points in stable areas, carefully avoiding a 150-metre buffer around known landslides to reduce spatial autocorrelation and enforcing a minimum spacing of 50 metres to prevent clustering.

Fourteen conditioning factors fed the models, each chosen for its documented influence on slope instability and its relevance to the geology, hydrology, and human geography of the corridor. Topographic variables—elevation, slope, aspect, and curvature—were derived from the ALOS PALSAR digital elevation model at 30-metre resolution. Hydrological factors included drainage density and distance to streams, while geological inputs comprised lithology, geology, distance to faults, and distance to lineaments drawn from Geological Survey of India data at 1:50,000 scale. Environmental and anthropogenic layers included land use and land cover, the Normalized Difference Vegetation Index derived from Sentinel-2 imagery, distance to roads from OpenStreetMap, and rainfall data from the India Meteorological Department.

The Frequency Ratio model, a classical bivariate statistical technique, quantified the association between each factor class and landslide occurrence by dividing the percentage of landslide pixels in a class by the percentage of the study area that class occupies. Values above one signal elevated susceptibility. The results were telling: the steepest slopes, between 45 and 78.18 degrees, showed the highest frequency ratio of 1.36, while barren land registered a striking 3.76 and built-up areas 2.64. Bare soil and rock with sparse vegetation scored 3.62, and areas within 486 metres of streams scored 1.56. Intriguingly, susceptibility was highest at the lowest elevations, between 815 and 1,130 metres, likely reflecting the concentration of road construction, settlements, and land-use modification at these altitudes.

Where the statistical model treats each factor in isolation, the machine learning approaches capture the nonlinear interplay that governs real slope failures. Random Forest, an ensemble method that grows hundreds of decision trees on bootstrap samples and aggregates their votes, was trained with 500 trees using the scikit-learn library. XGBoost, a gradient boosting algorithm that sequentially adds trees to correct the residuals of previous ones while penalising model complexity through a regularised objective function, was trained with 400 trees at a maximum depth of five and a learning rate of 0.05. Both models received the same balanced presence-absence dataset.

Performance was evaluated using the area under the receiver operating characteristic curve, a standard measure of a model’s ability to discriminate between landslide and non-landslide locations, alongside success rate curves and a rigorous spatial cross-validation scheme. The GroupKFold algorithm partitioned the data into spatial blocks so that all samples within a block went entirely to either training or testing, minimising spatial leakage. Random Forest achieved the highest ROC-AUC of 0.880, with a mean spatial cross-validation score of 0.888 plus or minus 0.021. XGBoost followed closely at 0.867, with a spatial mean of 0.876 plus or minus 0.026. The Frequency Ratio model trailed at 0.761, with a spatial mean of 0.805 plus or minus 0.025.

The spatial concentration of predicted risk was equally revealing. In the Random Forest map, the very high susceptibility class covered just 10.4 percent of the corridor yet contained 66.8 percent of all mapped landslides, while the very low class spanned 26.6 percent of the area with a negligible 0.3 percent of events. XGBoost performed similarly, concentrating 65.3 percent of landslides within 13.4 percent of the area. The success rate curves for both machine learning models rose steeply at the outset, indicating sharp discrimination between stable and unstable ground, whereas the Frequency Ratio curve climbed gradually, reflecting its weaker ability to delineate high-risk zones.

Feature importance analysis from both machine learning models converged on a consistent story. Land use and land cover emerged as the single most influential factor in Random Forest, contributing 0.1608 of the total importance, followed by elevation, aspect, NDVI, and slope—together accounting for roughly 52 percent of model importance. XGBoost likewise ranked land use and land cover first, with elevation and distance to streams next. The prominence of road proximity, which ranked sixth in Random Forest and carried weight comparable to slope itself, underscores how slope cutting, excavation, and drainage alteration during highway construction destabilise the very corridors that connect Himalayan communities.

The authors acknowledge limitations, including a temporal mismatch between the landslide inventory compiled from 2022 imagery and the 2024 land cover and vegetation datasets, and the absence of field verification of the susceptibility zones. Even so, the maps produced offer highway planners, engineers, and disaster managers a practical tool for prioritising slope stabilisation, guiding development, and targeting monitoring along one of India’s most critical and most fragile mountain routes. As climate extremes intensify and infrastructure expands across the Himalaya, the study makes a compelling case that data-driven machine learning, validated in space as well as in random splits, should become the default instrument for anticipating where the mountains will give way.

Subject of Research: Comparative landslide susceptibility mapping using Frequency Ratio, Random Forest, and XGBoost models along the NH-7 corridor in Chamoli district, Uttarakhand, India

Article Title: Landslide susceptibility mapping using Frequency Ratio, Random Forest, and XGBoost Models along NH-7 between Nandprayag and Vishnuprayag

Article References: Sidra, Dwivedi, D., Faruque, U., Raina, S. I., Sundaram, Siddiqui, L., Prakash, S., & Siddiqui, M. A. (2026). Landslide susceptibility mapping using Frequency Ratio, Random Forest, and XGBoost Models along NH-7 between Nandprayag and Vishnuprayag. Discover Geoscience, 4(1), Article 312. https://doi.org/10.1007/s44288-026-00683-3

Image Credits: AI Generated

DOI: 10.1007/s44288-026-00683-3

Keywords: landslide susceptibility, Random Forest, XGBoost, Frequency Ratio, machine learning, Himalaya, Uttarakhand, NH-7, Chamoli, GIS, remote sensing, disaster risk management