Tree Detection in Point Clouds Using Geometric Algorithms

The compelling role of tree identification and measurement spans ecological and socio-economic domains and emphasizes its importance for environmental studies and forest management. The accuracy of tree detection and parameter estimation is crucial, which has led to the use of advanced technological methods in recent research. In this study, geometric algorithms for tree detection in point clouds from terrestrial laser scanning (TLS) are evaluated to contribute to forest inventory and geographic information systems. Conventional tree measurement methods are based on manual inspection, which, despite its widespread use, has disadvantages such as high cost, labor and human error, which reduces accuracy. Our study explores geometric algorithms for automated and precise solutions. The circle fitting method automates the detection of tree trunks in horizontal cross-sections at certain heights and proves its efficiency in processing point cloud data. However, in certain cases, the method is affected by the irregular shapes of tree trunks that deviate from a circular shape, resulting in inaccurate estimates of both tree position and diameter at breast height DBH. The circular Hough transform, which is known to refine and eliminate unwanted shapes, is beneficial for circle detection and noise reduction in point clouds. It improves tree detection compared to manual methods, especially in terms of processing speed and error reduction, but is limited in complete denoising. The Random Sample Consensus (RANSAC) algorithm closes this gap and excels in removing outliers and accurately detecting cylindrical shapes of tree trunks. The basic methodology of the RANSAC algorithm involves applying ellipses to incomplete datasets and fitting lines to collections of 3D points to solve problems in locating cylinders in range data. We also investigate the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm, which is effective in detecting stems and eliminating irrelevant data. DBSCAN divides the data into clusters based on point density and removes regions of low density and noise but requires a minimum number of cluster points. It is effective in identifying and segmenting stems in complex high-density forest stands and complements another algorithm, such as Hough circle fitting, to better remove noise and avoid the impact on stem detection accuracy by the DBSCAN method in some cases. Our analysis highlights the utility of geometric algorithms in detecting trees and improving measurements in point clouds. These algorithms refine the shapes and filter the noise, which contributes to a more accurate estimation of tree parameters. However, each method has its advantages and limitations, with the choice of algorithm depending on specific requirements such as the type of point cloud data, the desired accuracy and the application purpose. In summary, our study provides a comprehensive evaluation of geometric algorithms for tree detection in point clouds and demonstrates the potential for the development of more sophisticated algorithms and methods. These results make an important contribution to forest inventory, especially in the application of terrestrial laser scanning in forestry.