Mapping North Korean Agriculture from Satellite Imagery

Commissioned by the Ministry of Unification, with Prof. David Lobell and Prof. Jennifer Burney at Stanford. North Korea has no public field labels, so standard crop mapping cannot be trained there.

We combine two crop classifiers trained in other countries, keep the pixels where they agree, and self-train a small transformer on those labels. It maps rice, maize and wheat/barley at 10 m resolution across North and South Hwanghae for 2019–2025, and the area estimates are checked against official statistics.

Funding: Ministry of Unification · with Stanford University
Period: 2026
Team: Prof. Jihee Kim (PI, KAIST), with Prof. David Lobell and Prof. Jennifer Burney (Stanford)
My role: Crop classification, area estimation and county-level analysis.

Four-step method: two models guess, keep where both agree, train our model, map every field — Method: two existing classifiers label North Korean pixels, we keep only where they agree, train a small transformer on those labels, and map all cropland at 10 m.

WHY IT IS HARD

There are no public field labels for North Korea, so a crop classifier cannot be trained there directly, and models trained in other countries transfer imperfectly.

WHAT WE DID

Two existing classifiers, one trained in China and one in the US and France, label Sentinel-2 pixels
We keep the pixels where both agree as training labels
A small transformer is trained on those labels and refined by self-training
The model maps rice, maize and wheat/barley at 10 m across North and South Hwanghae, 2019–2025

HOW WE CHECK IT

With no field ground truth, accuracy is measured as agreement with independent estimates: national statistics, the Rural Development Administration, and other published crop maps.