A Sensitivity analysis for a novel individual tree segmentation algorithm using 3D lidar point cloud data

LiDAR sampling or full-area coverage is deemed as favorable means to achieve timely and robust characterizations of vertically distributed forest attributes. So far, two main strategies on the use of LiDAR data in forestry are reported: area-based method (ABA) and individual tree method (ITC). Recently, a novel 3D segmentation approach has been developed for extracting single trees from LIDAR data. It is an integrated approach, which delineates tree crowns by using the watershed algorithm and stem detection followed by applying 3D normalized cuts segmentation. However, all the parameters for the modules used in the whole strategy including watershed and normalized cut segmentations are empirically determined and the key parameter settings are to be optimized. Additionally, the robustness of the 3D tree segmentation approach needs to be examined in order to give a better understanding of the algorithm mechanism. This paper highlights a study for sensitivity analysis of 3D single tree detection from small-footprint airborne LiDAR data by varying key parameters used in the segmentation procedure, such as segmentation threshold, minimal number of voxels allowed in a segment, voxel size, merged size of watershed segments, etc. The aim of the study is to find out the optimal combinations of key parameter values towards the performance of single tree detection via sensitivity analysis. It could help us to gain more insights into different models used in the whole procedure and is also of benefit to the further improvement and development in such models. Test data used in this work were captured with the Riegl LMS-Q680i scanner at four-fold point densities of 5pts/m, 10pts/m 15pts/m and 20pts/m under leaf-off condition. The study results proved the robustness and efficiency of the 3D segmentation approach in different airborne LiDAR data with respect to the single tree detection. Datasets whose point density is larger than 10 pts/m seem to hardly much contribute to the improvement to the performance of 3D tree detection. The performance of the approach might be further revealed and improved by optimizing the determination of key parameters’ values with respect to different data properties.