3D segmentation of forest structure using an adaptive mean shift based procedure

Abstract

Plant communities display a vertical structure based on the size and growth pattern of the dominant species. To a large extent, this pattern, called vertical stratification, depends on the climatic zone. Vertical structure analysis consists in detecting the number of layers and their limits within a forest stand. So far, there is a lack of robust approaches applied to airborne laser scanning (ALS) data that properly segment the different strata of forests having complex structures. In this study, we propose a procedure to characterize vertical forest stratification based on the mean shift (MS) algorithm. The MS is a non-linear filter that searches for local density maxima (modes). It is a non-parametric and unsupervised approach, which only requires a single criterion, the kernel bandwidth. Since the forest point cloud is a multi-modal distribution, the MS is used to find the modes which are supposed to be the barycenters of vegetation features. Once achieved, the modes are grouped together according to height range and the corresponding ALS points are assigned to each vegetation strata. Due to their complex pattern, using a single scale over the whole space is not recommended for the analysis of such environments. On this basis, the modes are computed using a variable kernel bandwidth according to the forest pattern. To depict such a pattern, we propose a new technique that segments the main forest layers at the plot level: overstory, understory, and surface vegetation. The procedure has been carried out on 45 plots of a Portuguese forest mainly composed of eucalyptus (Eucalyptus globulus) and pine (Pinus pinaster) trees that can be strongly populated by understory and surface vegetation.