Multimodal UAV data for individual tree species classification in a complexly structured, species-rich riparian forest

Riparian forests worldwide are pressured by many factors, including changes in climatic conditions and direct human interference. As a result, only very few riparian forests remain in Europe, and their role as a biodiversity hotspot and refugium for many species is more significant than ever, given the rapid decline in biodiversity worldwide. Methods for precisely monitoring biodiversity are an essential part of providing data for the management of these areas. Yet, vertical structural diversity, complex growing conditions, and a broad mix of tree species within one of Europe's biggest remaining riparian forests, the Nationalpark Donau-Auen (Austria), challenge current tree species mapping techniques. This study aims to contribute to an enhancement of monitoring practices as a key element for the management of the national park area. Therefore, the potential of flexible data acquisition by unmanned aerial vehicles (UAVs) for reliable mapping of tree species diversity is explored. The driving factors for the separability of tree species in the complexly structured test site are identified. A diverse set of multimodal UAV data, including multispectral imagery, photogrammetric point clouds, and LiDAR data from multiple timesteps, was acquired. A reference dataset was collected during intensive fieldwork. Individual tree segments were extracted manually on the one hand for optimized representation and automatically by a large-scale mean shift (LSMS) algorithm on the other hand. Multiple Random Forest models were trained using multitemporal and multisource data fusion from UAVs as input for both manual and automatic segmentation techniques. It was shown that the resulting overall accuracy (OA) for 14 tree species and the class ‘dead wood’ in the complexly structured test site varied greatly with the segmentation technique used. An OA of 86% could be reached for manually delineated trees, whereas the use of the area-wide automatic LSMS segmentation used for operational mapping yielded 78.3% OA. For manual segmentation, the effect of different sensor types was found to be small compared with the added value of multitemporal datasets, regardless of the sensor type used. Both multispectral and LiDAR sensors yielded an OA of around 73% for the best monotemporal timestep during the vegetation period. The findings of this study highlight the feasibility of using UAV-acquired data to map trees, even in richly structured and species-diverse forests.