Satellite imagery for a mountain country

What the imagery is good for

Five practical applications in Nepal, in roughly increasing technical difficulty:

Land cover and land use classification. Forest, cropland, settlement, water, snow — a basic classification model can label every pixel of the country, updated seasonally. ICIMOD has been doing exactly this for the Hindu Kush–Himalaya for years. The Nepal portion is freely available and is the backing layer for many downstream projects.

Crop area estimation. Distinguishing paddy from wheat from maize at scale, by combining the seasonality of each crop’s spectral signature with the time-of-year. This feeds the yield estimation models from the agriculture section.

Glacial lake inventory and change detection. Sentinel-2 lets a model identify every glacial lake in the country, and detect when one grows by more than a defined threshold — a primary signal for glacial lake outburst flood (GLOF) risk. ICIMOD maintains an active inventory; the AI gain is mostly in change detection between updates.

Built-up area expansion. Where is Pokhara growing? Which slopes around Tansen are being settled where they should not be? An imagery-based settlement model gives municipal planners a current view that paper records do not.

Damage detection after a disaster. Before-and-after imagery of an earthquake or flood zone, segmented automatically, can identify destroyed buildings, blocked roads, and changed river courses within hours of cloud cover clearing. This is the core of the post-quake assessment pipeline mentioned in the previous section.

What makes Nepal harder than it looks

Three specific geographical features make satellite AI in Nepal harder than the global average — not impossible, but harder than tutorials suggest.

Clouds. A satellite that passes overhead during monsoon sees clouds, not crops. For three months of the year, large parts of the middle hills and Tarai are reliably obscured. The fix is to combine optical imagery (which clouds block) with synthetic aperture radar (which clouds don’t) from Sentinel-1. Radar imagery is harder to interpret and has its own artefacts, but it is the only way to see through monsoon.

Steep terrain. Mountains complicate every satellite measurement. The same patch of land looks different depending on the sun angle, the slope, and the aspect — and “the same patch” on the satellite is a flat projection of what is actually a tilted surface. Models trained on flat agricultural plains (the US Midwest, the Indo-Gangetic plain) make subtle but persistent errors when imported into a Nepali context. Topographic correction is a real and necessary step.

Snow and shadow. A snow-covered slope looks bright; a shadowed slope looks dark. A model can confuse “snow” with “cloud” and “shadow” with “water” in ways that flatter terrain never produces. Standard tools handle this but require local tuning.

What a small team can do this year

A two-person team with a laptop, a Google Earth Engine account, and access to ICIMOD’s published datasets can ship, this year, any of the following:

1. A municipal land-use change dashboard for a specific province, updated monthly from Sentinel-2.
2. An automated post-monsoon landslide tally for a chosen river basin, with field-verifiable polygons.
3. A glacial lake change-detection alerting tool for the high Himalaya, with weekly bulletins.
4. A flood-extent rapid mapper triggered by DHM warnings, producing inundation polygons within hours.

None of these requires foundation-model-scale compute. None requires data that does not already exist. Each is a credible portfolio item, a credible donor proposal, and — if the team is willing to do the boring integration work — a credible piece of public-good infrastructure.

Check your understanding

Quick check

—

Why is satellite imagery analysis in Nepal harder than tutorials made for flatter agricultural regions might suggest?

1. ANepali satellites use a different wavelength than US satellites.
2. BSteep terrain causes the same patch of land to look different from different sun angles, monsoon clouds block optical sensors for months, and snow/shadow confuse standard classifiers — all of which require additional correction work that flat-plain models skip.
3. CThere is no satellite imagery available for Nepal.
4. DNepali law forbids the use of foreign satellite data.

Check answer→

What comes next

We have walked through three of the country’s defining substrates — language, money, and land. The fourth is people. Chapter 5 turns to health, identity, and education — the three public-good areas where AI in Nepal has the most direct effect on whether life gets noticeably better or not.