Efficient individual tree identification from multiple source point cloud

Detection and identification of individual trees are becoming increasingly important for forest related applications. Current, researches mostly concentrate on the segmentation of canopy height models (CHMs), which only produce rough tree delineation and numbers. Considering that different trees have changeable economic value and growing speed, this approach proposed a robust workflow that detects not only the location and shape of single trees but also the basic tree type. First, the multiple source point clouds are combined and classified to build the CHMs, based on the geometric information. Then, the tree tops are extracted based on the mathematical morphology operations. The tree delineation are then grown based on the tree tops and the competition among adjacent tress is considered. Finally, a multi-view projection based classification network was developed to identify the tree type, which meanwhile indicate the connecting trees that are hardly separated in CHMs. Experiments results on various data source demonstrate that the proposed approach can produce significant improvements on the detection results than existing researches.