Towards national-scale characterization of grassland use intensity from integrated Sentinel-2 and Landsat time series

The increased availability of systematically acquired high spatial and temporal resolution optical imagery improves the characterization of dynamic land surface processes such as agriculture. The use of time series phenology can help overcome limitations of conventional classification-based mapping approaches encountered when, for example, attempting to characterize grassland use intensity. In Europe, permanent grasslands account for more than one third of all agricultural land and a considerable share of the EU Common Agricultural Policy (CAP) budget is devoted to grasslands. The frequency and timing of mowing events is an important proxy for grassland use intensity and methods that allow characterizing grassland use intensity at the parcel level and over large areas are urgently needed. Here we present a novel algorithm that allows detecting and quantifying the number and timing of mowing events in central European grasslands. The algorithm utilizes all imagery acquired by Sentinel-2 MSI and Landsat-8 OLI for one entire year as available from the NASA Harmonized Landsat-Sentinel dataset. Cloud-free observations from both sensors are first synthesized through compositing at 10-day interval. Machine learning algorithms are then used to derive a grassland stratum. The intra-annual growing season profiles of NDVI values are subsequently assessed and compared to an idealized growing season trajectory. Residuals between the idealized trajectory and a polynomial model fit to the observed NDVI values are then evaluated to detect potential mowing events. We demonstrate and evaluate the performance of our algorithm and utilize its large area analysis capabilities by mapping the frequency and timing of grassland mowing events in 2016 on the national-scale across Germany. Our results suggest that 25% of the grassland area is not used for mowing. Validation results however suggest a relatively high omission error of the algorithm for areas that only experienced a single mowing event. The date ranges of detected mowing events compare overall well to a sample of interpreted time series points and to farm level reports on mowing dates. The mapped mowing patterns depict typical management regimes across Germany. Overall, our results exemplify the value of multi-sensor time series applications for characterizing land use intensity across large areas.