Multisensor data to derive peatland vegetation communities using a fixed-wing unmanned aerial vehicle

The restoration of drained peatlands to re-establish biodiverse peat-forming plant communities and typical habitats is a long-term process. To document this process, monitoring concepts must be found that are as operational, cost-effective and non-invasive as possible for the new sensitive ecosystems. The monitoring of the developing plant communities using multisensory Unmanned Aerial Vehicle (UAV) data has great potential. We investigated two fen sites in north-eastern Germany that were rewetted in the late 1990s. The areas were flown with a Fixed-Wing UAV and three sensors (RGB, multispectral, thermal). A multisensor dataset consisting of the sensor data, plant height and five spectral indices was classified with a random forest algorithm. The classification accuracies were 87.1 and 89.0% for 10 and 11 classes for the respective sites. Furthermore, the band importance was analyzed using the Gini index. Plant height in combination with the multispectral information were the most important variables. This study underlines the suitability of UAVs for monitoring spectrally similar vegetation because of their capability to generate 3D surface models using structure-from-motion methods and to carry different sensor systems. In addition, they can be used on a regional scale and mostly independent of cloud cover. Secondary results of the study are that the band importance can be determined by Random Forest algorithms even with few trees (e.g. 10 trees), because the order remains very stable independently of the number of trees.