UAV BASED MULTI SEASONAL DECIDUOUS TREE SPECIES ANALYSIS IN THE HAINICH NATIONAL PARK USING MULTI TEMPORAL AND POINT CLOUD CURVATURE FEATURES

Abstract. Low cost UAV systems are a flexible and mobile platform for very detailed spatial high-resolution point cloud and surface height mapping projects. This study investigates the potential of the DJI Phantom 4 Pro 3D point clouds and derived crown surface height information in combination with RGB spectral information for mapping of deciduous tree species in the Hainich national park area. RGB image data was captured in August, early October and November 2018 to create a multi seasonal spectral dataset for a 100 ha test area. The flight campaigns were controlled from the Hainich flux tower platform in 40 m height owned and operated by University of Göttingen in the central part of the park area. Absolut georeferencing accuracy of the datasets was improved using 7 DGPS measured control points within the stand structure on small forest clearings. Image files and ground control points were processed to a dense point cloud model with 2.6 billion points (approximately 200k points per tree crown object) using the Agisoft Metashape cluster processing environment. Additionally, a digital surface model and a true ortho image mosaic with 3 cm spatial resolution was generated. For the differentiation of deciduous tree species, a reference data set with coordinates for the tree species Fagus sylvatica (beech), Fraxinus excelsior (ash), Acer pseudoplatanus (sycamore maple), Carpinus betulus (hornbeam) and dead trees and early defoliated trees was defined. The study site is however dominated by Fagus sylvatica and Fraxinus excelsior. We studied two different groups of features: tree crown surface height variability parameters using point cloud densities, point cloud height variance, local standard deviation of gaussian curvature, standard deviation of local point cloud roughness and multi temporal normalised spectral features using multi seasonal uncalibrated UAV RGB data. Analysis of feature separability showed that very high-resolution point cloud surface curvature properties with small neighbourhood radii can differentiate some tree species types but we also found multitemporal spectral ratios based on RGB data to be very successful in differentiating the main tree species.

Results of this work show that super fine very dense point cloud models and derived roughness measures of mixed forest stand surfaces hold valuable information for deciduous species discrimination and will likely also be very useful for morphological analysis of tree crown types.