Grassland Use Intensity Classification Using Intra-Annual Sentinel-1 and-2 Time Series and Environmental Variables

Detailed spatial data on grassland use intensity is needed in several European policy areas for various applications, e.g., agricultural management, supporting nature conservation programs, improving biodiversity strategies, etc. Multisensory remote sensing is an efficient tool to collect information on grassland parameters. However, there is still a lack of studies on how to process, combine, and implement large radar and optical image datasets in a joint observation framework to map grassland types on large heterogeneous study areas. In our study, we assessed the usefulness of 2521 Sentinel-1 and 586 Sentinel-2 satellite images and topographic data for mapping grassland use intensity. We focused on the distinction between intensively and extensively managed permanent grassland in a large heterogeneous study area in Slovenia. We provided dense Satellite Image Time Series (SITS) for 2017, 2018 and 2019 to identify important differences, e.g., management practices, between the two grassland types analysed. We also investigated the effectiveness of combining two different remote-sensing products, the optical Normalised Difference Vegetation Index (NDVI) and radar coherence. Grassland types were distinguished using an object-based approach and the Random Forest classification. With the use of SITS only, the models achieved poor performance in the case of cloudy years (2018). However, the performance improved with additional features (environmental variables). The feature selection method based on Mean Decrease Accuracy (MDA) provided a deeper insight into the high-dimensional multisensory SITS. It helped select the most relevant features (acquisition dates, environmental variables) that distinguish between intensive and extensive grassland types. The addition of environmental variables improved the overall classification accuracy by 7-15%, while the feature selection additionally improved the final overall classification accuracy (using all available features) by 2-3%. Although the reference dataset was limited (1259 training samples), the final overall classification accuracy was above 88% in all years analysed. The results show that the proposed Random Forest classification using combined multisensor data and environmental variables can provide better and more stable information on grasslands than single optical or radar data SITS on large heterogeneous areas. Therefore, a combined approach is recommended to distinguish different grassland types.