Reading the steppe from space, checked against the ground

Our approach

We stand on published science — including our own

Estimating how much grass grows across 1.5 million square kilometres, and which winters will turn deadly for herds, is a very hard scientific problem. Nobody solves it alone. Our pipeline is built on decades of peer-reviewed research on Mongolian rangelands, from the agencies that pioneered winter disaster (dzud) risk mapping to the field ecologists who described how pastures degrade and recover.

That literature is not just background reading. Specific published findings are wired directly into how our models are built, which warning signs our indices weight, and how honestly we grade our own accuracy. Where the research community has learned a hard lesson, such as the finding that models often look far more accurate than they really are when tested carelessly, we adopt the stricter standard.

The foundation

One national grid, fed by satellites and field crews

Everything we publish starts from the same foundation: a single grid laid over the whole country. For every cell and every year since 2000, we assemble more than 150 indicators — how green the vegetation is and when it greens up, how much rain and snow fell, how hot the summer and how brutal the winter, the shape of the terrain, the soil beneath, and how many animals graze nearby.

The satellite backbone comes from three long-running public programmes — MODIS, Landsat, and Sentinel — processed through Google Earth Engine, alongside global climate and soil datasets. Livestock numbers flow in from the National Statistics Office's open data service. On top of that sits what makes the models trustworthy: ground truth, measured by hand.

• ALAMGaCPasture biomass plots. Field teams from the land administration agency clip, dry, and weigh grass at monitoring sites across the country — the scale against which our biomass model is calibrated.
• NAMEMRangeland productivity survey & dzud methodology. The meteorological agency's long-running degradation survey and its operational winter-risk parameters anchor our other two products.
• NSOLivestock census & mortality records. The annual animal census and per-district loss tables — the hard numbers our dzud index is judged against.

01 Pasture biomass

How much grass is out there?

The first question herders, planners, and insurers all ask is the simplest: how much forage does each district have this year? Satellites can't weigh grass — but they can see how green it is, when it greens up, and how that compares to the rain and heat that produced it.

Our biomass model learns the connection between those satellite signals and the real thing: thousands of plots where ALAMGaC field crews have clipped, dried, and weighed the grass by hand, year after year. A machine-learning model trained on those pairs can then estimate forage for every cell of the national grid — including the vast areas no field crew could ever reach.

Graded on places it has never seen

A model tested on the same sites it learned from will flatter itself. So we grade ours the hard way: it must predict biomass for monitoring sites that were entirely held out of its training, in regions it has never seen. Published comparisons show this stricter test can halve a model's apparent accuracy — which is exactly why we use it. The numbers we report are the honest ones.

From biomass we derive a second number herders recognise: carrying capacity — how many animals, in standard sheep units, a district's pasture can sustain without being eaten past recovery.

02 Rangeland productivity

Is the pasture degrading — and can it come back?

Rangelands don't fade smoothly from good to bad. Ecologists describe them with state-and-transition models: a pasture sits in one of several distinct states, and pressure — drought, sustained overgrazing, trampling — can tip it from one state into the next. Early transitions reverse on their own when pressure eases. Later ones cross a threshold where the palatable grasses are gone, and no single good summer brings them back.

Mongolia is unusually well prepared for this kind of analysis. NAMEM's national monitoring network has rated tens of thousands of sites on a five-class health scale for over a decade, and Mongolian and international ecologists have catalogued the country's rangeland types and their transition pathways. Our classifier learns to read those same five classes from the satellite record — comparing each site not against the whole country, but against healthy examples of its own ecological type, so a naturally sparse desert-steppe pasture isn't mislabelled as a degraded mountain meadow.

The prize is coverage: field crews can rate thousands of points a year; the model extends their standard to every cell of the grid, flagging where land is slipping toward a threshold while recovery is still cheap.

03 Dzud risk

Which districts face the hardest winter?

A dzud is Mongolia's signature disaster: a winter so severe that livestock die in the millions. The deadliest pattern is well documented in the research — a dry summer that leaves animals thin and pastures bare, followed by deep snow and long cold that lock away whatever forage remains. It is a two-season story, and any credible risk measure has to read both seasons.

Our dzud indicator follows the parameter framework used operationally by NAMEM, refined with corrections from published vulnerability research: summer pasture condition and drought carried into the winter, snow depth and cover, sustained extreme cold, and the pressure of herd sizes on the pasture that must feed them. These signals are combined into a single index and ranked into five risk classes for every district.

Judged against real losses

An index that has never been checked is an opinion. Ours is validated against the National Statistics Office's official livestock-loss records: in the catastrophic winter of 2023–24, the provinces our index placed in its highest risk band were among those that in fact lost the most animals. The agreement is strong but honestly imperfect — winter weather is only part of why herds survive or don't.

Straight talk

What we don't claim

Trust in a monitoring platform comes from knowing its edges. Ours are these:

• The dzud index is an assessment, not a forecast. It scores conditions as they are observed. Turning it into a true early-warning forecast is ongoing work with the relevant agencies.
• Winter losses are never weather alone. Hay reserves, herd condition, household wealth, and emergency response all shape outcomes — and none of them are visible from orbit. A district ranked lower-risk can still suffer.
• Our maps are landscape-scale. They describe districts and landscapes, not individual winter camps or a single family's pasture.
• Models inherit their ground truth. Where field measurements are sparse — remotest deserts, high mountains — our estimates carry more uncertainty, and we flag it rather than hide it.

Reading list

Research we build on

A non-exhaustive selection of the published work that most directly shaped GreenZone's models and indices.

1. Purevjav, A.-O., Avirmed, T., Wilcox, S. W. & Barrett, C. B. (2025). “Climate Rather than Overgrazing Explains Most Rangeland Primary Productivity Change in Mongolia.” Science, 389.Why our pipeline separates climate signals from grazing pressure instead of conflating them.
2. Chuluun, T., Altanbagana, M., Ojima, D., Tsolmon, R. & Suvdantsetseg, B. (2017). “Vulnerability of Pastoral Social-Ecological Systems in Mongolia.” In Rethinking Resilience, Adaptation and Transformation in a Time of Change, Springer.Source of key refinements to our dzud index, including carrying last summer's conditions into the winter score.
3. Rao, M. P., Davi, N. K., D'Arrigo, R. D., et al. (2015). “Dzuds, droughts, and livestock mortality in Mongolia.” Environmental Research Letters, 10, 074012.The benchmark study linking summer drought and winter severity to livestock mortality — the pattern our index is built to detect.
4. Tachiiri, K., Shinoda, M., Klinkenberg, B. & Morinaga, Y. (2008). “Assessing Mongolian snow disaster risk using livestock and satellite data.” Journal of Arid Environments, 72.Early evidence that satellite snow and vegetation data can anticipate dzud losses.
5. Nandintsetseg, B., et al. (2018). Dzud risk as combined natural hazard and livestock vulnerability.Frames dzud risk as hazard × vulnerability — the structure behind our component-based index.
6. Fernández-Giménez, M. E., Batkhishig, B. & Batbuyan, B. (2012). Lessons from the Dzud: Adaptation and Resilience in Mongolian Pastoral Social-Ecological Systems. World Bank.Grounds our honesty about the human factors — reserves, mobility, response — that no satellite index can see.
7. Densambuu, B., Sainnemekh, S., Bestelmeyer, B. & Budbaatar, U. (2018). National Report on the Rangeland Health of Mongolia: Second Assessment. Green Gold–Animal Health Project, SDC.The national catalogue of rangeland types and their state-and-transition pathways — the ecological backbone of our degradation classifier.
8. Muro, J., et al. (2022). “Predicting plant biomass and species richness in temperate grasslands across regions, time and land management with remote sensing and deep learning.” Remote Sensing of Environment.Why we grade our biomass model on regions it has never seen, and report the stricter number.
9. Meyer, H. & Pebesma, E. (2021). “Predicting into unknown space? Estimating the area of applicability of spatial prediction models.” Methods in Ecology and Evolution.The method behind flagging where our predictions are reliable — and where they are stretched.
10. Brown, C. F., et al. (2025). “AlphaEarth Foundations: an embedding field model for accurate and efficient global mapping from sparse label data.” arXiv:2507.22291.One of the modern satellite representations our recent model generations draw on.