Methods for spatially explicit estimation of NATURA 2000 grassland forage quality using satellites

Spatially explicit mapping of grassland forage quality is of major interest for sustainable grazing management of NATURA 2000 areas, especially if those are large or have limited accessibility. Therefore, this study is concerned with the estimation of crude protein (CP) and organic acid detergent fiber (oADF) content at regional scale using Sentinel-2 and Landsat 8 remote sensing data. Field data were collected in the Grafenwoehr military training area in Bavaria, Germany. Different combinations of predictor variables were applied using cross-validated random forest regression, linear regression with lasso penalty and linear regression with ridge penalty models. The red-edge band of Sentinel-2, centered at 705 nm, as well as the shortwave infrared bands of both sensors and related vegetation indices contributed the most to the respective models. Linear regression with lasso penalty and Sentinel-2 data performed consistently better, compared to the other models. The results (CP (10.1 - 23.1%): max R2 0.53, RMSE 1.78%; oADF (22.7 - 39.5%): max R2 0.72, RMSE 2.3%) demonstrate the potential of remote sensing as an information tool in supporting the conservation management of grassland areas with limited access.