Floods, landslides, earthquakes

Three hazard families, three AI fits

The three hazards that define risk in Nepal — floods, landslides, earthquakes — sit at very different points on the “what can AI do” curve.

Floods are the strongest fit. A flood forecasting model takes rainfall observations (from gauges and satellite), upstream river-level data, terrain models, and historical flood records, and outputs a probabilistic forecast of where and when water will breach a bank. Google’s Flood Hub and the open-source DELFT-FEWS family work in Nepal today; the Department of Hydrology and Meteorology (DHM) runs an operational early-warning system on parts of the Koshi, Karnali, and Bagmati basins. The hardest part is no longer the model. It is the last mile — getting a warning into the hands of a fisherman in Saptari before the water arrives.

Landslides are a moderate fit. Susceptibility maps — which slopes are at risk — are produced well by AI today, using terrain, geology, land cover, rainfall history, and the location of past slides. Sentinel-2 imagery before and after monsoons lets a model detect new slides at scale, where field surveys would take months. Forecasting a specific slide is much harder; the science of predicting precisely when a slope will give way is still rough, regardless of model size.

Earthquakes are the weakest fit. Earthquake forecasting in the sense of “the next quake on this fault, in this month, of magnitude M” is genuinely not a solved problem, by any technology, anywhere. What AI can do is help with secondary tasks: rapid damage assessment from satellite imagery in the hours after a quake, identifying which neighbourhoods need search teams first, estimating shake-induced landslide locations, optimising relief logistics. Most of the high-value AI work for earthquakes is post-event, not prediction.

The last-mile problem

The 2021 Melamchi flood is a study in how the model can work and the country can still lose. DHM had upstream rainfall data. The risk picture was visible to anyone watching the gauges. But the warning, when it went out, did not propagate fast enough or specifically enough through the chain that ends with a shopkeeper in the bazaar getting time to move stock to high ground. The model was right. The institutional pipe leaked.

This is not an AI problem in the narrow sense. It is a systems problem with AI as one component:

• Detection — model says “high risk of flooding at this location, this window.”
• Decision — DHM, the local administration, and the army agree to issue.
• Dissemination — SMS blast, FM radio, sirens, ward-level WhatsApp groups, jhola-carrying volunteers in the gaupalika.
• Action — the receiving citizen actually moves, on time.

A perfect model and a broken dissemination layer save no one. The marginal improvement worth the most in Nepali disaster AI is often not in the model — it is in moving the warning from “DHM bulletin” to “everyone whose name is on the ward register, in their pocket, in their language, within five minutes.”

What is already running

For floods, DHM’s early-warning system covers parts of Nepal’s major river basins, integrating rainfall, river-level, and snowmelt observations. ICIMOD, the UK Met Office, and the World Meteorological Organization have all contributed to the operational stack over the years. The system is not glamorous and is not yet country-complete; both are fixable.

For landslides, ICIMOD’s Hindu Kush–Himalaya geoportal hosts susceptibility maps used by NGOs and the government. The Department of Mines and Geology maintains the geological substrate these maps build on.

For earthquakes, the National Society for Earthquake Technology (NSET), in cooperation with international research groups, has been Nepal’s institutional memory of seismic risk since well before 2015. NSET’s risk modelling, school-retrofit programme, and post-event response protocols are mature.

What is missing in each case is scale and integration. We have models on three river basins out of dozens, landslide maps that are not updated between monsoons, and an earthquake response protocol whose primary information bottleneck is still a phone tree.

Three AI projects that would change things

If asked to point to three disaster-AI projects that would meaningfully improve outcomes, in this order:

1. A country-complete, district-resolution flood early-warning system, with a tested dissemination layer that reaches mobile users in their language inside five minutes of a warning issuing. Extend coverage from a few basins to all of them.

2. A continuously-updated post-monsoon landslide inventory, derived automatically from Sentinel-2 imagery, with new slides flagged for field verification within days. This unblocks every road-clearing, relief, and infrastructure decision that follows.

3. A pre-positioned post-earthquake damage-assessment pipeline, trained on imagery and damage labels from past quakes (Gorkha, but also Sichuan, Haiti, Türkiye), ready to deliver rapid neighbourhood-level damage estimates in the hours after an event. The team and tooling sits ready; the next quake activates it.

None of these is a research project. Each is a deployment project, and all are achievable inside three years.

Check your understanding

Quick check

—

In disaster early warning for Nepal, what is usually the *most* important place to improve — once a working forecasting model exists?

1. AMake the model larger.
2. BThe “last mile” — getting the warning into the hands of the right citizens, in their language, in time to act.
3. CSwitch from open-source to proprietary models.
4. DAdd more cloud GPUs.

Check answer→

What comes next

The last section in this chapter widens the lens — to the satellite imagery that makes all of this possible. Nepal is a mountain country, and a fifth-generation satellite stack now exists that can see every glacier, forest, river, and town in the country every five days. The next section is about what that imagery is good for, and where it stumbles.