WorldCereal

Knowing where and what crops are grown is fundamental to food-security monitoring, yet timely crop maps that remain consistent across regions and growing seasons are difficult to produce. WorldCereal addresses this challenge through an open, modular Earth-observation workflow that generates seasonal 10 m crop information, including cropland extent and selected crop types. AI enters the workflow at the classification stage: satellite time series and other geospatial inputs are transformed into crop maps using geospatial foundation models finetuned on harmonised reference data. The system combines a Reference Data Module for storing and standardising reference data, openEO-based processing workflows for training and inference, and user-facing tools for product generation and access. In its current evolution, WorldCereal’s use of geospatial foundation models such as Presto improve generalisation beyond well-represented countries and years while keeping retraining feasible in constrained computing environments. Users can apply infrastructure-provided workflows and, where appropriate, retrain classifiers with public or contributed reference data.

WorldCereal was initiated as an ESA-funded effort to demonstrate that seasonal, field-scale crop mapping could be made more timely, open and reproducible at large scale. The service builds on collaboration between Earth-observation scientists, machine-learning specialists, software engineers, data curators and agricultural-monitoring users. A key enabling condition is the combination of harmonised in-situ reference data with cloud-based processing workflows exposed through open standards such as openEO.

The operational requirement is not only high classification accuracy, but also transferability across geographies, seasons and user contexts. Reference data are valuable but geographically and temporally uneven, and the service must support mapping in areas and years that are not always well represented in the training set. This has motivated continued work on more generalisable model components, including lightweight foundation models. Such models are relevant only if they fit the operational constraints of the service: users may need to retrain models without specialised GPU infrastructure, and workflows must remain portable across cloud backends.

The integration of newer model components is therefore treated as an operational engineering question rather than a model-substitution exercise. The team first defines the service requirements and hypotheses, then determines how the model can be adapted to the WorldCereal processing chain, and finally evaluates performance under realistic spatial, temporal and visual conditions. This helps distinguish improvements that are useful in production from benchmark gains that may not translate into trustworthy maps.

The wider service is supported by open-source code, documented workflows, a processing hub, training activities and stakeholder engagement through the WorldCereal consortium and advisory structures. These mechanisms help make the workflow accessible beyond the core development team while preserving traceability and scientific oversight.

WorldCereal’s main outcome is an operational pathway for producing timely, high-resolution crop information in a way that is open, reproducible and adaptable to regional needs. The workflow supports seasonal crop monitoring by combining harmonised reference data, satellite observations, cloud processing and machine-learning classification. It also provides user-facing mechanisms for generating products and adapting workflows to local monitoring questions, helping bridge scientific methods and operational food-security applications.

The experiments with lightweight foundation models add a second outcome: a better tested route for improving generalisation in operational remote-sensing systems. Benchmarking showed that pre-trained representations can improve performance particularly in more difficult transfer settings, while also confirming that not every additional adaptation step is useful. The evaluation therefore goes beyond headline accuracy scores and considers where models succeed or fail across countries, years and visual map quality.

The practical lessons are that model performance must be judged together with operational constraints: portability, compute requirements, compatibility with existing processing chains, reference-data coverage, transparency and user trust. Lightweight models are attractive because they can support retraining in moderate-resource settings, but they still require validation and expert interpretation before being used for decisions. More broadly, WorldCereal demonstrates how open Earth-observation infrastructure, curated reference data and AI can be combined into a transferable crop-monitoring service, while making clear that responsible use depends on documentation, uncertainty-aware interpretation and continued engagement with agricultural-monitoring users.

https://esa-worldcereal.org/en

Van Tricht, K., Degerickx, J., Gilliams, S., Zanaga, D., Battude, M., Grosu, A., … & Szantoi, Z. (2023). WorldCereal: a dynamic open-source system for global-scale, seasonal, and reproducible crop and irrigation mapping. Earth System Science Data, 15(12), 5491-5515. https://essd.copernicus.org/articles/15/5491/2023/essd-15-5491-2023.pdf

Butsko, C., Van Tricht, K., Tseng, G., Milli, G., Rolnick, D., Cartuyvels, R., Reshef, I.B., Szantoi, Z. and Kerner, H., 2025. Deploying geospatial foundation models in the real world: lessons from worldcereal. arXiv preprint arXiv:2508.00858. https://arxiv.org/abs/2508.00858