DeepGrass: A transformer-based approach for enhanced grassland management through Sentinel-2 time series spectral signature and vegetation indices forecasting

Grasslands are vital ecosystems that offer numerous ecological services, including water regulation, carbon storage, climate mitigation, and pollination. They support diverse flora and fauna, serving as essential habitats for many species, and are fundamental to livestock grazing, thereby contributing to food production and rural economies. Accurate forecasting of grassland productivity aids in setting appropriate stocking rates, ensuring sustainable grazing practices, and assists in planning for forage availability and water resources, enhancing overall management efficiency and environmental sustainability. In this study, we developed DeepGrass, an encoder-decoder transformer model designed to forecast all 10 Sentinel-2 spectral bands (SBs) and 28 vegetation indices (VIs) over the coming 15 days at field-level. The model was trained and validated using data from grasslands across Sweden, spanning three phenological seasons (2020−2022) for a total of 565,614 grassland fields, with additional inference testing on 2023 data (for a total of 87,738 grassland fields in Sweden and 44,764 in Finland) to simulate real-world and generalization applicability. DeepGrass achieved competitive results with respect to other state-of-the-art approaches, focused solely on NDVI forecasting, and demonstrated consistent results across all VIs and SBs, for example: NDVI with RMSE ranging from 0.028 ( t + 1 ) to 0.105 ( t + 15 ); NDMI with RMSE ranging from 0.038 ( t + 1 ) to 0.115 ( t + 15 ); Red band with RMSE ranging from 0.0075 ( t + 1 ) to 0.0205 ( t + 15 ); and NIR band with RMSE ranging from 0.028 ( t + 1 ) to 0.073 ( t + 15 ). Notably, VIs derived from forecasted SBs exhibited performance comparable to that of directly forecasted VIs. This finding implies that, in practice, all desired VIs can be computed a posteriori from the forecasted SBs without necessitating separate model retraining for each specific VI. Additionally, the model trained on field-average time series performed reasonably well on pixel-level data, suggesting computational efficiency without compromising accuracy. The availability of satellite observations in the input time series was crucial, especially for long-term forecasts, emphasizing the need to assess cloud coverage to evaluate model output reliability. This advanced forecasting approach provides a valuable tool for monitoring grassland dynamics. Accurate forecasting of SBs and VIs is vital for effective grassland management, enabling informed decisions for grazing, mowing, and conservation practices.