Where do slopes fail in Rudraprayag? A small-inventory landslide susceptibility study, and how evaluation choices change the answer
Abstract
Rudraprayag district in the Garhwal Himalaya (Uttarakhand, India) is highly landslide-prone and lies on the pilgrimage route to Kedarnath. No landslide inventory was accessible for this study, so I built a preliminary one semi-automatically: bare, steep patches detected in Sentinel-2 imagery were verified one by one on high-resolution imagery, giving 38 landslides. I trained random forests on four terrain factors (elevation, slope, aspect and distance to river) and evaluated them with spatial cross-validation. Evaluation choices changed the results substantially. With background (non-landslide) points drawn across the whole district, a model that knew only each point's coordinates scored almost as well as the terrain model (AUC 0.83 vs 0.85). Adding vegetation and land cover raised the AUC to 0.93–0.95, but only because the landslides had been detected as bare ground. Results from any single random draw of background points were unreliable: across 20 draws, the same model's AUC ranged from 0.42 to 0.86. With background points matched within 3 km of each landslide and results reported across 20 random draws, the terrain model reaches a within-area AUC of 0.74 on average (95% of draws: 0.52–0.85) and beats a location-only model in 90% of draws. A map averaged over 20 models places 87% of held-out landslides (95% range 73–94%) in its two highest classes, which cover 40% of the area, compared with 49% of nearby stable slopes. Satellite rainfall (NASA GPM IMERG, 1998–2025) shows that the monsoon brings about 72% of annual rainfall, and that June 2013, the month of the Kedarnath disaster, was the wettest June on record, although 2013 was only the third wettest monsoon overall. The results are preliminary. The main contributions are an open inventory and pipeline, and a worked demonstration of how common sampling and evaluation choices can overstate the skill of small-inventory susceptibility models.
Related
- Landslide Susceptibility Modelling and Infrastructure Exposure in the Zillertal, Austrian Alps
- Machine-learning modelling of regional landslide susceptibility in eThekwini, South Africa
- Landslide Susceptibility Modeling and Mapping: A Comparison of Frequency Ratio and Decision Tree Models
- Role of Topo-Climatic Factors in Landslide Susceptibility Mapping: A Machine Learning Perspective
- A single framework for assessing flash flood and landslide susceptibility: an application to the Mediterranean Liguria region, Italy
BibTeX
@misc{makker2026where,
title = {{Where do slopes fail in Rudraprayag? A small-inventory landslide susceptibility study, and how evaluation choices change the answer}},
author = {Gauri Makker},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x5kb9j},
url = {https://eartharxiv.org/repository/view/15237/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Where do slopes fail in Rudraprayag? A small-inventory landslide susceptibility study, and how evaluation choices change the answer
AU - Gauri Makker
PY - 2026
DA - 2026/09/27
PB - EarthArXiv
DO - 10.31223/x5kb9j
UR - https://eartharxiv.org/repository/view/15237/
AB - Rudraprayag district in the Garhwal Himalaya (Uttarakhand, India) is highly landslide-prone and lies on the pilgrimage route to Kedarnath. No landslide inventory was accessible for this study, so I built a preliminary one semi-automatically: bare, steep patches detected in Sentinel-2 imagery were verified one by one on high-resolution imagery, giving 38 landslides. I trained random forests on four terrain factors (elevation, slope, aspect and distance to river) and evaluated them with spatial cross-validation. Evaluation choices changed the results substantially. With background (non-landslide) points drawn across the whole district, a model that knew only each point's coordinates scored almost as well as the terrain model (AUC 0.83 vs 0.85). Adding vegetation and land cover raised the AUC to 0.93–0.95, but only because the landslides had been detected as bare ground. Results from any single random draw of background points were unreliable: across 20 draws, the same model's AUC ranged from 0.42 to 0.86. With background points matched within 3 km of each landslide and results reported across 20 random draws, the terrain model reaches a within-area AUC of 0.74 on average (95% of draws: 0.52–0.85) and beats a location-only model in 90% of draws. A map averaged over 20 models places 87% of held-out landslides (95% range 73–94%) in its two highest classes, which cover 40% of the area, compared with 49% of nearby stable slopes. Satellite rainfall (NASA GPM IMERG, 1998–2025) shows that the monsoon brings about 72% of annual rainfall, and that June 2013, the month of the Kedarnath disaster, was the wettest June on record, although 2013 was only the third wettest monsoon overall. The results are preliminary. The main contributions are an open inventory and pipeline, and a worked demonstration of how common sampling and evaluation choices can overstate the skill of small-inventory susceptibility models.
ER -