Detection of fallen trees in ALS point clouds by learning the Normalized Cut similarity function from simulated samples
- 1Dept. of Geoinformatics, Munich University of Applied Sciences, 80333 Munich, Germany
- 2Photogrammetry and Remote Sensing, Technische Universität München, 80333 Munich, Germany
- 3Dept. for Research and Documentation, Bavarian Forest National Park, 94481 Grafenau, Germany
Keywords: Fallen tree detection, graph cuts, virtual sample generation
Abstract. Fallen trees participate in several important forest processes, which motivates the need for information about their spatial distribution in forest ecosystems. Several studies have shown that airborne LiDAR is a valuable tool for obtaining such information. In this paper, we propose an integrated method of detecting fallen trees from ALS point clouds based on merging small segments into entire fallen stems via the Normalized Cut algorithm. A new approach to specifying the segment similarity function for the clustering algorithm is introduced, where the attribute weights are learned from labeled data instead of being determined manually. We notice the relationship between Normalized Cut’s similarity function and a class of regression models, which leads us to the idea of approximating the task of learning the similarity function with the simpler task of learning a classifier. Moreover, we set up a virtual fallen tree generation scheme to simulate complex forest scenarios with multiple overlapping fallen stems. The classifier trained on this simulated data yields a similarity function for Normalized Cut. Tests on two sample plots from the Bavarian Forest National Park with manually labeled reference data show that the trained function leads to high-quality segmentations. Our results indicate that the proposed data-driven approach can be a successful alternative to time consuming trial-and-error or grid search methods of finding good feature weights for graph cut algorithms. Also, the methodology can be generalized to other applications of graph cut clustering in remote sensing.