Yield Forecasts with AI: how satellite data is changing agriculture

A farmer looks over his freshly tilled field. The soil has been prepared, the seed has been sown - now it's time to wait. Months pass before the harvest, during which wind, weather and other environmental influences determine success or failure. Even after many years of experience, forecasting yields is often like looking into a crystal ball. This uncertainty is not a sign of a lack of expertise, but an expression of the enormous complexity of modern agriculture.

Target data collected on the ground serve as comparative values. Agricultural machinery equipped with special sensors measures the yield directly at harvest. Based on this information, the scientists can see how valid the algorithm's forecast was. The realisation after numerous test runs: The AI system achieves a high level of agreement with measured yields by utilising the satellite images and downstream data fusion.

For wheat cultivation in the USA, for example, an R² score of 0.92 is achieved. The R² value shows how closely a prediction matches the actual yield. The closer the value is to 1, the better. A value of 0.92 therefore corresponds to a congruence with the yield measured in the respective field of around 92 per cent.

There is therefore an error tolerance not only in the system itself, but also in the comparative data used. This results in the hypothetical scenario that the algorithm is closer to the actual yield than the sensors of the agricultural machinery. AI forecasts also remain probabilistic. Factors such as extreme weather events or inaccurate sensors can influence accuracy, which is why uncertainties are modelled and taken into account.

The predictability of expected yields results in direct added value for the agricultural sector. Logistics become easier to plan and the entire harvest chain becomes more efficient. In retrospect, weak points in the fields can be identified and analysed and appropriate countermeasures can be taken. Beyond the agricultural sector, the aggregated forecasts support decisions in the areas of food security, trade strategy and climate adaptation.

The modular architecture of the technology also allows it to be used beyond agriculture: from forest condition monitoring and biodiversity analyses to urban heat mapping or extreme weather forecasts. The technology platform developed at DFKI is transferable in order to visualise environmental changes and recognise trends at an early stage.

The project is a prime example of basic research with concrete application relevance. Scientists at DFKI have worked closely with companies to utilise machine learning in areas where it can offer concrete solutions to key challenges - from agriculture to climate adaptation.