Detection of mowing events from combined Sentinel-1, Sentinel-2, and Landsat 8 time series with machine learning
The intensity of land use in permanent grasslands affects both biodiversity and important ecosystem services. Optical satellite systems have already proven to be suitable for area-wide detection of proxies of grassland management intensity, namely mowing events. However, clouds lead to considerable gaps in time series, resulting in an underestimation of the total number of events. SAR systems like Sentinel-1 (S1) can overcome this limitation, yet the information obtained is more complex to interpret. To test the synergy and complementarity of both sensor types for mowing detection, we computed high-density SAR and optical time series over three test sites in Germany covering detailed reference data on grassland management. For the growing periods in 2018 and 2019, we tested two binary, supervised machine learning algorithms, a convolutional neural network (CNN) and support vector machines (SVM), classifying sliding windows into mown and not mown. S1 VH/VV backscatter ratio, as well as Sentinel-2 (S2) and Landsat 8 (L8) normalized difference vegetation index (NDVI), were used as input features. Both models show promising results in detecting mowing events, where SVM performed slightly better. Overall, the approach shows a high potential for routinely mapping grassland management intensity over large areas in heterogeneous environments.