Combining graph-cut clustering with object-based stem detection for tree segmentation in highly dense airborne lidar point clouds

Single tree detection has been a major research topic when it comes to support of collecting automatic field inventory using lidar. All previous methods show under- and over-segmentation effects because the associated control parameters have a limited scope. This paper describes a novel integrated single tree segmentation using a graph-cut clustering method that is supported by automatic stem detection. The key idea is to replace the static stopping criterion, which is usually defined by trial and error or by a sensitivity analysis, here with a query for whether a stem position has been provided by the stem detection in the remaining cluster to be partitioned. The stem detection automatically detects tree stems by identifying vertical lines based on a hierarchical classification procedure. We evaluate both stem detection and integrated single tree segmentation on mixed temperate forest plots captured in a leaf-on situation. The highly dense airborne lidar data was acquired with an average point density of more than 200 points/m2. The forest areas, which are located in Eastern Austria, are characterized by a stem density of around 1000 stems/ha and a tree age between 15 and 63 years. We test our algorithms with reference data measured both by visual interpretation of the laser point clouds using a conventional field campaign. In our experiments, we confirm that the automatic stem detection technique successfully locates stems if the lidar point density of the stems is at least five points/m. The experimental results show that stem detection alone can detect more than 80% of the stems, with a precision of better than 70%. Moreover, we prove that the integration of the stem detection renders the graph-cut segmentation effectively independent of the stopping criterion and improves the overall accuracy of the tree segmentation as well. In terms of F-scores, the overall improvement is up to 15% and 6% for reference data from visual inspection and field measurements, respectively. Compared with the results of two existing tree segmentation methods that were applied to the same datasets, the accuracy improvement is also demonstrated by F-scores increased up to 22% and 5%, respectively. Very interestingly, the integrated tree segmentation considerably enhances the detection accuracy, especially in mixed and deciduous forest areas by more than 10% in the case of reference data provided by visual inspection.