Food security decisions require lead time. A government agency that learns about a significant crop shortfall two weeks before harvest can do very little. The same agency with 60 days of advance warning can trigger procurement, adjust import arrangements, activate subsidy mechanisms, and communicate with affected farmers in time to have a meaningful impact.
The agricultural value of satellite intelligence is measured in days of advance notice.
The data inputs
For the Danube Basin study covering winter wheat across Hungary, Romania, northern Serbia, and eastern Austria, we combined three primary data streams.
Sentinel-2 multispectral imagery at 10 metre resolution provided vegetation indices measured throughout the growing season. ERA5 soil moisture reanalysis from the European Centre for Medium Range Weather Forecasts provided the soil condition context that is the primary driver of drought related yield loss. Sentinel-3 land surface temperature data provided heat stress indicators during critical grain filling periods.
The causal attribution step
Standard statistical models for crop yield forecasting produce an anomaly score. They can tell you that vegetation in a given district is below normal. They cannot tell you why.
Our causal attribution layer adds this dimension. For each district showing a yield stress signal, the system determines whether the primary cause is soil moisture deficit, heat stress, or vegetation index patterns consistent with disease. This matters because the correct response to drought is different from the correct response to disease, and activating the wrong intervention wastes resources and may make outcomes worse.
Results
Against the 2025 harvest outcomes reported by national statistics offices, the district level forecasts we produced in mid April showed a mean prediction error of 4.3 percent across 340 districts. For comparison, the European Commission's agricultural monitoring bulletin showed mean errors of 7.1 percent for the same area in the same year.
Of the twelve districts where our error exceeded 10 percent, nine had been explicitly flagged by our epistemic confidence system as high uncertainty cases. The system correctly identified where it was less confident. This is the epistemic honesty property in operation.
The practical consequence is that a food security agency acting on our April forecast would have had accurate 60 day advance warning of the districts facing significant yield deficits.