Optimization
Save the food already grown with a forecast made in season.
We think food security plans should give losing less food the same weight as growing more. Food lost after harvest has already used the land, water and money it took to grow. FAO estimates that 13.3% of food was lost between harvest and retail in 2023, and the UN's 2026 report on food security estimates that around 645 million people faced hunger in 2025. We would forecast harvests during the season and plan where food is stored and moved.
How can AI improve food security and agricultural planning?
AI helps when it turns satellite, weather and market data into forecasts during the season that are checked against official statistics. We would combine crop models with satellite imagery to estimate yields by district, link them to stock and price data, and use optimization to plan where to store and move food at the lowest combined cost and loss. Agronomists review each forecast, and a named official signs off decisions.
Food is lost while people go hungry
FAO's loss estimate covers food lost on farms and in transport, storage, wholesale and processing. UNEP adds that 1.05 billion tonnes of food were wasted in 2022, almost one fifth of the food available to consumers. The 2026 food security report estimates that 2.69 billion people could not afford a healthy diet in 2025.
Timing matters as much as harvest size. A harvest forecast that arrives during the season leaves more options for imports, reserves and transport than one that arrives after the harvest.
What we would model, and with which data
Satellite imagery shows the crop while it grows, before official harvest statistics exist.
- FAOSTAT, which FAO describes as free and open access to food and agriculture data from 245 countries and territories since 1961.
- FAO's Global Information and Early Warning System, which monitors the food security situation in every country, including crop prospects and food prices.
- Free satellite imagery from the Copernicus Sentinel missions, to track crop growth through the season.
- Weather records and seasonal forecasts from national weather services.
- Stock, storage and transport records from food agencies, cooperatives and companies, added under agreement.
- Wholesale and retail prices from market monitoring systems.
How the model is built and checked
A crop model simulates how a crop grows day by day from weather, soil and farm practice, and estimates its yield. It cannot see everything that happens in a field, so satellite observations correct it during the season. A supply model adds stocks, imports and expected demand by area and month, and shows where shortfalls are likely. Optimization then plans storage and transport, choosing which warehouse to draw from and which route to use. Back-testing runs past seasons and compares the estimates with official harvest statistics.
What decision-makers get
A forecast is worth most to the people who store and move food.
- Agriculture ministries get district yield forecasts during the season, with ranges and early warning of likely shortfalls.
- Food reserve agencies get stock plans that say how much to buy, where to store it and when to release it.
- Agribusinesses, food processors and input suppliers get supply forecasts for their sourcing areas, and crop insurers get an independent yield estimate for each area.
- Distributors and retailers see the routes and storage points where loss is highest.
A first project picks one crop in one region, or one food reserve. We would build the model from FAO data, satellite imagery and published statistics, back-test it on past seasons and hand over a scorecard. The ministry or company then decides whether to continue.
How it runs safely
Farm-level data belongs to farmers and cooperatives.
- It is used only with their agreement, and published outputs are aggregated so no single farm can be identified.
- Forecasts that can move prices are released on a published schedule, to everyone at once.
- Fairness checks look at whether forecast errors are larger in smallholder areas, where data is thinner, and those areas get wider ranges rather than false precision.
- Agronomists and officials review every forecast before it informs a decision, and we log every run.
Open-source building blocks
Open crop models let anyone check the assumptions behind a forecast.
- PCSE, the Python Crop Simulation Environment, with the WOFOST, LINGRA and LINTUL3 crop and grassland models.
- DSSAT, the Decision Support System for Agrotechnology Transfer, whose cropping system model covers more than 45 crops.
- AquaCrop-OSPy, an open-source Python crop-water model based on AquaCrop-OS. Its authors note it is not an official FAO implementation.
- Open Data Cube, for analyzing satellite data through time at continental scale.
- eo-learn, a Python framework for processing Earth observation data for machine learning.
- OR-Tools, for storage and transport planning.
Has extendfuture done this for a ministry or food company?
No. Optimization is a new extendfuture offering. Our production work so far is AI for companies, with accuracy measured as a number, human review and an audit trail.
Can this work where farm data is scarce?
Yes, with wider ranges. Satellite imagery and FAO statistics cover every country, so a first model can run before local data arrives. Field surveys then narrow the ranges.
Does this replace official crop statistics?
No. Official statistics remain the record. The model gives an earlier estimate during the season and is checked against the official figures when they arrive.
Do you need farmers' personal data?
No. District forecasts run on satellite imagery, weather and aggregate statistics. Farm records are used only with the farmer's or cooperative's agreement.
Talk to the people who build.
One call. An honest read on what AI can do for this, and the number it has to beat.