Multitemporal and Multispectral Drone Data for Classifying Tree Species in an Austrian Riparian Forest

The climate crisis is threatening native forests all over Europe and a change in the composition of species can be expected in the future. Diversity of tree species is one key aspect of resilience against the climatic stress caused by the climate crisis. To maintain one of the last huge floodplains in Europe – the Danube Floodplain – for future generations, it is necessary to monitor the change of the species distribution and the development of eco-systems. Remote sensing is widely used for establishing a constant monitoring of forests, including tree species classification (TSC). Currently, Unmanned Aerial Vehicles (UAVs) offer very high-resolution data together with temporal flexibility and cost efficiency which can be used in the management practice of forests and national parks in particular. However, due to the extensive diversity inherent in different forest types and tree species, the results obtained in the state-of-art research in TSC via very high-resolution optical data cannot be generalised. As there is still a gap in research in the field of TSC in riparian forests, this study aims at filling this gap with preliminary results of TSC in a riparian forest, namely the Danube Floodplain National Park (Austria). Therefore, three drone flights were conducted during October, September, and May spanning the years 2021 and 2022 together with a simultaneous collection of reference data in the field. Tree crowns were delineated manually in two different ways: point-buffered and exact delineation of the crown shape. Multiple object-based Random Forest models were performed, comparing mono- and multitemporal data as well as two different spatial resolutions (3.0 cm and 6.4 cm) and the two different levels of detail of the delineation of tree crowns. Highest Overall Accuracy (OA) for 12 different tree species and one dead wood class could be reached by the multitemporal model at 82.1 % (kappa = 80.8 %) with the higher spatial resolution (3.0 cm) and the exact delineation of the reference data. Producer’s Accuracy (PA) and User’s Accuracy (UA) varied between 50 % and 100 % for different classes. Promising results from this study showed that the presented method can be used for precise monitoring of tree species diversity in the Danube Floodplain National Park. Further improvement could be reached by merging data from different sensors.