Every monsoon, landslides in Nepal block highways, bury homes, and cut off villages. Many are triggered by rainfall, which means, in principle, they come with a warning sign: the rain itself. The challenge is turning rain data into warnings precise enough to act on and rare enough to be trusted.
Two different questions
Landslide early warning answers two questions, and it helps to keep them separate:
- Where? Which slopes are susceptible. This changes slowly: geology, slope, land cover, roads.
- When? When rainfall pushes susceptible slopes past the tipping point. This changes by the hour.
ML helps with both, in different ways.
"Where": susceptibility mapping
A susceptibility map scores each grid cell by how likely it is to fail. Inputs typically include:
- Slope angle and aspect from a digital elevation model
- Geology and soil type
- Land cover, from satellite imagery
- Distance to roads and rivers. Road cuts destabilise slopes, and this matters a lot in Nepal.
- A landslide inventory: where landslides have happened before
Classic approaches use logistic regression or frequency ratios. ML models like random forests and gradient boosting often perform better on these tabular features and give feature importance, which helps geologists sanity-check the model.
"When": rainfall thresholds
The classic tool is an intensity–duration threshold: a curve showing how intense rainfall must be, over a given duration, to trigger landslides. Above the curve, warn. Below it, don't.
This idea dates back decades in the research literature and is still the backbone of many operational systems. Variants use cumulative rainfall over several days, because soaked soil fails with less rain. That's why late-monsoon landslides can occur after moderate storms.
ML improves on single thresholds by:
- Combining antecedent rainfall, soil moisture, and short-term intensity in one model
- Learning regional differences rather than one national curve
- Using satellite rainfall (such as GPM) where gauges are sparse
Data sources a developer can use
| Data | Source type |
|---|---|
| Elevation | Global DEMs (e.g. Copernicus DEM, SRTM) |
| Land cover | Sentinel-2 derived products |
| Rainfall | Ground stations plus satellite rainfall products |
| Soil moisture | Satellite and reanalysis products |
| Landslide inventories | National agencies, research datasets, NASA's global landslide catalogue |
More in Open Climate Datasets for ML Projects in 2027.
Why Nepal is a hard case
- Steep, young geology in the Himalaya.
- Intense, concentrated monsoon rainfall.
- Rapid rural road building that cuts slopes without adequate drainage.
- Sparse gauge networks in high, remote terrain.
- Communication gaps in the villages most at risk.
Honest limits
- False alarms erode trust. A system that warns every rainy day will be ignored by week three.
- Small events are invisible to coarse satellite data.
- Warnings without evacuation plans don't save lives.
- Models trained on past monsoons may underestimate future extremes.
What would actually help
- Better inventories, including community-reported landslides with photos and GPS.
- Regional thresholds updated each season with new events.
- Impact-based messages for road users and villages: "Avoid the X–Y road section tonight." See Impact-Based Forecasting.
- Integration with existing flood and weather warnings, so people get one message, not three.