Abstract
Knowledge of forest structures—and of dead wood in particular—is fundamental to understanding, managing, and preserving the biodiversity of our forests. Lidar is a valuable technology for the area-wide mapping of trees in 3D because of its capability to penetrate vegetation. In essence, this technique enables the detection of single trees and their properties in all forest layers. This paper highlights a successful mapping of tree species—subdivided into conifers and broadleaf trees—and standing dead wood in a large forest 924 km2 in size. As a novelty, we calibrate the critical stopping criterion of the tree segmentation based on a normalized cut with regard to coniferous and broadleaf trees. The experiments were conducted in Šumava National Park and Bavarian Forest National Park. For both parks, lidar data were acquired at a point density of 55 points/m2. Aerial multispectral imagery was captured for Šumava National Park at a ground sample distance (GSD) of 17 cm and for Bavarian Forest National Park at 9.5 cm GSD. Classification of the two tree groups and standing dead wood—located in areas of pest infestation—is based on a diverse set of features (geometric, intensity-based, 3D shape contexts, multispectral-based) and well-known classifiers (Random forest and logistic regression). We show that the effect of under- and oversegmentation can be reduced by the modified normalized cut segmentation, thereby improving the precision by 13%. Conifers, broadleaf trees, and standing dead trees are classified with overall accuracies better than 90%. All in all, this experiment demonstrates the feasibility of large-scale and high-accuracy mapping of single conifers, broadleaf trees, and standing dead trees using lidar and aerial imagery.
Highlights
Forest inventories based on remote sensing data vary in terms of scale, sensors, and calculated forest structural attributes
We were able to demonstrate that tree segmentation based on normalized cut can be successfully performed with a stopping criterion calibrated for tree groups ’conifers’ and ’broadleaf trees’, thereby significantly reducing the effect of under- and oversegmentation
This concept to calibrate parameter NCutmax regarding the tree groups bears the drawback that the segmentation needs to be performed manifold in the case of multiple tree species
Summary
Forest inventories based on remote sensing data vary in terms of scale, sensors, and calculated forest structural attributes. Latifi and Heurich [1] reported that lidar point clouds advantageously fused with optical imagery are the most prominent choices for inventory of forest structural variables at the landscape and regional scales. Aside from area-based methods, tree-level approaches estimate forest inventory parameters utilizing segmented single trees. A huge set of raster-based approaches are available that use the canopy height model (CHM) as basic information [3]. Aside from these low-level methods, more sophisticated approaches operate on the point cloud instead of using the CHM. The study successfully shows a significant improvement of the single tree detection (up to 20%) in the lower forest layers
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.