Monitoring herbaceous biomass and restoration of semi-natural grasslands (Iris Luik)

“Monitoring herbaceous biomass and restoration of semi-natural grasslands using machine learning on Sentinel-1 and Sentinel-2 imagery” by Iris Luik The Open-Earth Monitor Global Workshop 2026 "Next-Generation EO & GeoAI Built on Openness and Distributed Data" 7-9 October 2026, Barcelona, Spain https://earthmonitor.org/global-workshop-2026/ Semi-natural grasslands are critical ecosystems that provide a range of essential services, with their role as habitats for diverse species being among the most significant. However, over the past century, semi-natural grasslands that once covered vast areas across Europe have largely been transformed into intensively managed agricultural lands, abandoned, or converted into forests. These large-scale land-use changes have led to considerable biodiversity loss, making the conservation and restoration of semi-natural grasslands an important component of sustainable landscape management. We utilized 97 in-situ herbaceous biomass samples collected during the summer of 2019 from alvar grasslands in Western Estonia, all restored between 2015 and 2019. Samples were collected from 20 × 20 cm plots nested within 2 × 2 m botanical plots. Sentinel-1 and Sentinel-2 imagery from the same period was used, with median band values and derived indices (e.g., NDVI, BSI, SAVI, VH/VV) included as predictors. Random Forest models were developed using Sentinel-1 and Sentinel-2 spectral bands and derived indices as predictors. Model robustness was evaluated using 5-fold cross-validation. Two approaches for linking field and satellite data were tested: point sampling and a 3 × 3 kernel mean, with point sampling performing slightly better. The model achieved an RMSE of 98 ± 54 g/m², an MAE of 71 ± 30 g/m², and an R² of 0.32 ± 0.08, reflecting the high spatial variability of semi-natural grasslands. SHAP analysis identified SAVI, NDVI, and the vegetation red edge band B8A as the most important predictors, while Sentinel-1 variables contributed less to model performance. These results highlight the dominant role of optical data in herbaceous biomass estimation and demonstrate that simple point-based sampling can outperform spatial averaging approaches. The proposed methodology provides a practical and scalable solution for monitoring grassland restoration.