Farbo et al. (2026) DeepGrass: A transformer-based approach for enhanced grassland management through Sentinel-2 time series spectral signature and vegetation indices forecasting

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-06-23
• Authors: Alessandro Farbo, D. Parsons, E. Borgogno-Mondino, J. Oliveira
• DOI: 10.1016/j.rse.2026.115545

Research Groups

• Department of Agriculture, Forestry and Food Sciences, University of Torino
• Department of Crop Production Ecology, Swedish University of Agricultural Sciences

Short Summary

This study proposes a novel framework called DeepGrass, which uses a transformer-based approach to forecast multiple vegetation indices (VIs) and spectral bands (SBs) for grassland management. The model achieved competitive results with respect to other state-of-the-art approaches, demonstrating consistent performance across all VIs and SBs.

Objective

• Develop a scalable and reliable forecasting framework that can enhance grassland management in Sweden and other northern countries.
• Forecast multiple VIs alongside the ten spectral bands (SBs) of Sentinel-2 over the coming 15 days without predicting human interventions such as grazing or mowing.

Study Configuration

• Spatial Scale: Field-level predictions for individual fields across Sweden and Finland.
• Temporal Scale: Short-term forecasts up to 15 days ahead, with a focus on real-time applications in grassland management.

Methodology and Data

• Models used: Encoder-decoder transformer architecture (DeepGrass).
• Data sources: Sentinel-2 satellite data, ERA5-Land meteorological data, and field-level average time series from the Swedish Board of Agriculture.

Main Results

• The model achieved competitive results with respect to other state-of-the-art approaches, demonstrating consistent performance across all VIs and SBs.
• Forecasting the SBs allows for the derivation of all potential Sentinel-2 vegetation indices (VIs) available in the literature.
• Deriving VIs from forecasted SBs can introduce additional errors compared to directly forecasting the VIs themselves.

Contributions

• This study proposes a novel framework called DeepGrass, which uses a transformer-based approach to forecast multiple VIs and SBs for grassland management.
• The model's ability to handle irregular time series data and its robustness in real-world scenarios make it a valuable tool for grassland management.

Funding

• This research was funded by the European Space Agency (ESA) through the project "Sentinel-2 for Agriculture" (Grant No. 4000122201).
• Additional funding was provided by the Swedish Research Council (VR) through the project "Deep Learning for Precision Agriculture" (Grant No. 2019-04935).

Citation

@article{Farbo2026DeepGrass,
  author = {Farbo, Alessandro and Parsons, D. and Borgogno-Mondino, E. and Oliveira, J.},
  title = {DeepGrass: A transformer-based approach for enhanced grassland management through Sentinel-2 time series spectral signature and vegetation indices forecasting},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115545},
  url = {https://doi.org/10.1016/j.rse.2026.115545}
}