Volume IV-2/W4
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-2/W4, 35-42, 2017
https://doi.org/10.5194/isprs-annals-IV-2-W4-35-2017
© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 4.0 License.
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-2/W4, 35-42, 2017
https://doi.org/10.5194/isprs-annals-IV-2-W4-35-2017
© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 4.0 License.

  12 Sep 2017

12 Sep 2017

DETECTION OF SINGLE TREE STEMS IN FORESTED AREAS FROM HIGH DENSITY ALS POINT CLOUDS USING 3D SHAPE DESCRIPTORS

N. Amiri1,2, P. Polewski1, W. Yao1, P. Krzystek1, and A. K. Skidmore2 N. Amiri et al.
  • 1Department of Geoinformatics, Munich University of Applied Sciences, Munich, Germany
  • 2Department of Natural Resources, Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente, Enschede, the Netherlands

Keywords: High density ALS, 3D point clouds, Tree trunk, Classification, Forestry

Abstract. Airborne Laser Scanning (ALS) is a widespread method for forest mapping and management purposes. While common ALS techniques provide valuable information about the forest canopy and intermediate layers, the point density near the ground may be poor due to dense overstory conditions. The current study highlights a new method for detecting stems of single trees in 3D point clouds obtained from high density ALS with a density of 300 points/m2. Compared to standard ALS data, due to lower flight height (150–200 m) this elevated point density leads to more laser reflections from tree stems. In this work, we propose a three-tiered method which works on the point, segment and object levels. First, for each point we calculate the likelihood that it belongs to a tree stem, derived from the radiometric and geometric features of its neighboring points. In the next step, we construct short stem segments based on high-probability stem points, and classify the segments by considering the distribution of points around them as well as their spatial orientation, which encodes the prior knowledge that trees are mainly vertically aligned due to gravity. Finally, we apply hierarchical clustering on the positively classified segments to obtain point sets corresponding to single stems, and perform ℓ1-based orthogonal distance regression to robustly fit lines through each stem point set. The ℓ1-based method is less sensitive to outliers compared to the least square approaches. From the fitted lines, the planimetric tree positions can then be derived. Experiments were performed on two plots from the Hochficht forest in Oberösterreich region located in Austria.We marked a total of 196 reference stems in the point clouds of both plots by visual interpretation. The evaluation of the automatically detected stems showed a classification precision of 0.86 and 0.85, respectively for Plot 1 and 2, with recall values of 0.7 and 0.67.