Comparison of individual tree detection methods in forests with different structure and species composition using airborne LiDAR data

Introduction and Objective: ALS data has been widely used to describe the 3D canopy structure due to collection of 3D point clouds with high spatial resolution and accuracy. Individual tree detection using LiDAR data is an efficient method for extracting the structural characteristics of the forest such as tree height, diameters at breast height and above-ground biomass. Accurate estimation of these parameters is of key importance in the management of forest resources.
Material and Methods: So far, different methods have been developed to detect single trees, which have different limitations and capabilities and show different reactions to changes in forest tree species and canopy vertical structure. Therefore, in this research, a quantitative approach was developed in order to evaluate the scientific-technical performance of six single tree detection algorithms from ALS data. These methods include two raster-based methods, two point-based methods and two combined methods. Due to lack of access to LiDAR and ground data from Iran's forests, the unique dataset of the NEWFOR project was used, which covers different forests of the Alpine region with a variety of tree species and different canopy structures.
Results: The results showed that the vertical structure of the canopy plays a significant role in detection accuracy of ITD methods and its effect is greater than the composition of forest tree species. The highest detection rate is related to the combined method in single-layered coniferous forests with a value of 0.91 and the lowest detection rate is related to raster-based method in multi-layered mixed forests with a value of 0.45. The detection rates of the studied methods in the highest height layer vary from 66% to 91% and in the lowest height layer (2-5 meters) from 13% to 52%.
Conclusion: Although the understory trees cannot be extracted with the same accuracy as the dominant trees, the results showed that the hybrid method of marker-controlled watershed segmentation with K-means clustering algorithm was able to detect 91% of the trees in the highest height layer and the highest number of understory trees with a detection rate of 52% in the lowest height layer. This method has the highest values of detection rate, detection accuracy and overall accuracy with values of 0.83, 0.91 and 0.87 respectively, the lowest amount of commission and omission errors with values of 9.03% and 17.36% respectively and also the best horizontal and height accuracy with values of 1.33 m and 0.87 m, respectively.