Grassland yield estimation based on Sentinel-2 time series and comparison to process-based modelling and statistics in southern Germany

Kurzfassung

Managed grasslands dominate the landscape in southern Germany, as in many regions of Europe. They are used for livestock fodder production and provide a variety of further ecosystem services, such as carbon storage, water filtration and the provision of habitats. Consequently, grasslands play a major role for mitigating climate change and biodiversity loss. In southern Germany, grasslands are constantly managed, often highly intensive. They are mown up to six times per year with varying mowing dates and partly also grazed. The management intensity strongly determines grassland productivity and the provision of other ecosystem services. There it is often a trade-off between high yields and ecosystem services relevant for ecology and nature conservation. However, grassland management as well as productivity and other ecosystem services are usually unknown. Grassland harvests are usually not weighed and traded and, therefore, spatially explicit data on grassland yields are lacking. Earth observation has the potential to inform on grassland yields as extensive and reproducible monitoring of grassland biomass is enabled. However, in mown grasslands, grassland biomass is characterized by recurring removal and regrowth during the growing season. Multi-temporal biomass estimates as well as information on mowing dates are therefore required to estimate annual grassland yields. In this study, we built and optimized an empirical extreme gradient boosting model based on Sentinel-2 data and above-ground biomass field measurements acquired in 2019-2021 in southern Germany to estimate grassland biomass. We included mowing dates of am existing multi-annual grassland mowing product in the modelling process to support the biomass estimations. Apart from the time since the last mowing event or since the start of the growing season, Sentinel-2 reflectance bands from the green, red, blue, near infrared and short-wave infrared domains as well as the Enhanced Vegetation Index, the tasseled cap wetness index and the acquisition date were used as model input features. Through applying the model to Sentinel-2 time series of 2019, multi-temporal biomass was mapped. Afterwards, we combined the multi-temporal biomass estimates with the mowing dates and calculated the accumulated annual yield. During this step, we aggregated the pixel-based biomass information to field level by using the 95% percentile of the biomass derived on the field area. We investigated the yield maps for the study region in southern Germany and compared the results to grassland yield products of different sources. On the one hand, grassland yields resulting from the bio-geochemical process-based model LandscapeDNDC, optimized for the study region, were used for comparison. On the other hand, yield maps were derived from reference values from the Bavarian State Institute for Agriculture (LfL), coupled with mowing information and data on the stocking rate, as used by administrative authorities. In addition, the relationship of yield to influencing factors, such as mowing frequency, temperature, and precipitation, was examined. The satellite data-based extreme gradient boosting model achieved a R² of 0.97 (RMSE = 0.18 t/ha) during training and a R² of 0.68 (RMSE = 0.19 t/ha) with an independent testing data set. The most important input features were a short-wave infrared reflectance band, the tasseled cap wetness index and the time since the last mowing event or start of season. The estimated annual yields based on the Earth observation approach reached plausible results of 3-10 t/ha for the study region in southern Germany. The LandscapeDNDC approach and the approach based on reference values resulted in annual grassland yields of around 4-9 t/ha. The modelled yields showed the largest value range for the Earth observation approach, the overall yields were the smallest compared with the yields of the other two approaches. Yield maps show patterns of annual grassland yields with similarities and differences between the three approaches and rather follow the mowing frequency than patterns of climatic conditions. The mowing frequency was also the influencing factor with the highest correlation coefficient (Pearson r = 0.81 for the Earth observation approach) for all three yield estimation approaches. The largest differences between the estimated yields from the different approaches were found for grasslands mown one to two times per year. The study demonstrates the capability of estimating annual grassland yields based on Earth observation data and field measurements. The comparison to two other approaches, which are either used within the scientific community, or used by administrative authorities, aids in evaluating the results and showing advantages and limitations of the various yield estimation approaches. Using Earth observation enables grassland yield monitoring without large amounts of input data and calibration as needed for bio-geochemical models, capturing intra- and interannual variations in contrast to approaches solely based on reference values.