Abstract
• Advantages of both CHM-based and point-based methods are combined. • Crown width scales are estimated automatically based on gradient magnitudes. • Multi-scales segmented results are fused and optimized gradually. • Neighboring trees are separated based on probability density function splitting. To obtain satisfying results of individual tree detection from LiDAR points, parameters using traditional methods usually need to be adjusted by trials and errors. When encountering complex forest environments, the detection accuracy cannot be satisfied. To resolve this, a multi-level self-adaptive individual tree detection method was presented in this paper. The proposed method can be seen as a hybrid model, which combined the strength of both raster-based and point-based methods. Raster-based strategy was first used for achieving initial trees detection results, while the point-based strategy was adopted for optimizing the clustered trees. In the proposed method, crown width scales were estimated automatically. Meanwhile, multi-scales segmented results were fused together to take advantage of segmented results of both larger and small scales. Six different coniferous forests were adopted for testing. Experimental result shows that this study achieved the lowest omission and commission errors comparing with other three classical approaches. Meanwhile, the average F1 score in this paper is 0.84, which is much highest out of other methods.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have