基于区域合并的Mean Shift算法识别单木研究

Individual Tree Detection by Mean Shift Algorithm Based on Region Merging

为准确识别森林单木,采用区域合并的MeanShift算法对机载点云进行单木分割。首先,以点云三维特征空间为特征向量,选择核带宽度及收敛阈值,采用MeanShift算法对点云进行初始过分割;其次,以过分割点簇为对象,选择分割尺度、平滑度和紧凑度参数,采用基于区域邻接图的最优层次合并方法对点簇进行合并。最后,剔除3.5m高度以下和异常点云,以点云中心点为单木位置,计算森林密度。实验结果表明,基于区域合并的MeanShift算法能够检测到89%以上的单木,单木识别精度达91.6%,避免了生成CHM的初始误差。

The Mean Shift algorithm based on region merging was used to identify the individual tree information by the airborne point cloud segmentation. With the feature space of three-dimensional point cloud as the feature vector and selection of kernel band width and convergence threshold, the Mean Shift algorithm was adopted for initial over segmentation of point cloud. The segmentation scale, smoothness and compactness of point clusters were selected, and the optimal hierarchical merging method based on the region adjacency graph was adopted to merge the point clusters. The forest density was calculated by removing the point cloud below 3.5m height and abnormal point cloud and taking the point cloud center as the individual tree position. The experimental results showed that this approach could detect more than 89% of individual trees with the identification accuracy of 91.6%, avoiding the initial error of CHM generation.