Crop Stress and Yield Early Detection Use Case
The new European Green Deal has recently highlighted the importance of accelerating the digital and green transitions to achieve a climate-neutral, sustainable economy. European agriculture is continually affected by extreme weather, which is likely to become more severe and frequent in the near future. The way crops respond to adverse weather conditions largely depends on their development stage. Systems for timely monitoring of crop phenology are essential for understanding and evaluating the impact of climate change on crop production.
Despite the development of crop maps and crop yield forecasting at the European scale, integrating Earth observation and weather/climate data is needed to capture the effects of increasing weather extremes on agricultural phenology. This use case aims to monitor crop phenology at a country or continental scale by using Sentinel-2/1 data, fused with climate data (e.g., ERA5), and publicly available European crop-type maps derived from farmers’ declarations.
Data sources used
The EO datasets used include Sentinel‑2, Daymet, ERA5 climate reanalysis products, and MODIS surface reflectance products, as well as publicly available European crop data.
Methodology
This use case leverages existing geospatial foundation models such as Prithvi and Terramind, along with the TerratorcSpatiotemporal datasets from Earth observation and climate sensors are retrieved over the growing season for crops such as wheat and maize. Spatiotemporal datasets from Earth observation and climate sensors are retrieved over the growing season for crops such as wheat and maize. These datasets are preprocessed to ensure cloud-free acquisitions and alignment with agriculture-specific areas. The retrieved patches are used as inputs for the foundation models employed to produce embeddings.
Embeddings
To generate embeddings, Sentinel-2 patches are encoded with several foundation models such as Prithvi and Terramind. These pretrained GeoFM encoders (Prithvi and Terramind) produce a global embedding vector that captures the crop conditions and phenology activity in a compact representation. These embeddings are finally utilised to track seasonal phenological activity and to model in-season and end-of-season crop yield using several machine learning algorithms, such as Random Forests.
Expected Impact
These GeoFMs provide compressed, general-purpose representations (embeddings) of EO datasets that encode spectral, spatial, and contextual information, thereby reducing data storage and compute requirements for continental-wide crop monitoring.
This use case is expected to support a range of actors in the agricultural community:
- farmers and agricultural organisations (e.g., improved monitoring/forecasting of field damage assessment);
- the public sector responsible for governing the transition of agriculture;
- the private sector, including the agricultural technology and machinery industry, seed companies and agribusiness retailers, the agrochemical industry, and the insurance sector for risk management;
- environmental agencies conducting crop forecasting activities.
Future work
Future work will extend this analysis to include additional crops and explore methods to adaptively fuse multimodal GeoFM embeddings to support scalable yield and phenology research. Also of interest will be investigating (i) cross-domain adaptation learning, and (ii) the transferability of embeddings across different regions and scales from field or parcels to the regional level, such as NUTS3.
References
- K. Adriko, R. Sedona, L. Seguini, M. Riedel, G. Cavallaro and C. Paris, (2025) “From MODIS to Sentinel-2: A Regional Comparative Analysis of Crop-Yield Prediction With Matched Spatiotemporal Data,” in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 18, pp. 27663-27683