Exploring the potential of land surface phenology and seasonal cloud free composites of one year of Sentinel-2 imagery for tree species mapping in a mountainous region

Abstract Optical satellite imagery with high temporal and spatial resolution, such as acquired by Sentinel-2, is increasingly becoming available and is used to derive maps of tree species. Such mapping products are required in the scope of operational and sustainable forest management. Existing studies that employ Sentinel-2 imagery have already evaluated different classification algorithms but are often confined to areas smaller than a single Sentinel-2 scene. In this study, the area of interest (a large part of the Province of Tyrol (Austria)) is covered by two Sentinel-2 tiles, of which approximately 5000 km² are forested. In order to deal with seasonal metrics under recurrent cloud cover conditions, we exploit one year of Sentinel-2 imagery by using land surface phenology (LSP) and seasonal cloud-free composites for mapping five different tree species groups (Broadleaved-, Larch- (Larix), Pine- (Pinus), Dwarf Pine- (Pinus mugo) and Spruce/Fir (Abies alba/Picea abies) stands). Although a regular multitemporal classification setup based on three cloud-free images reached an overall accuracy of around 84.4 % and outperformed monotemporal setups by around 10 % points, the availability of single cloud-free images was limited in the mountainous region. Thus, alternative approaches, using combined measures for the entire time series of Sentinel-2 imagery, i.e. three-monthly temporal reflectance composites and phenological metrics, were tested and could even improve overall accuracy by 1–2 % points. As a conclusion, we agree with previous studies that multitemporal imagery can help improving the mapping accuracy. However, leveraging satellite image time series for large-scale mapping of tree species should not only rely on high-quality cloud-free single images and should strongly be supported by i.e. seasonal composites or multi-image metrics. Therefore, development and provisioning of such datasets should be fostered.