Research on waveform separation of trees and shrubs in spaceborne LiDAR GEDI based on spectral clustering

To address the issue that spaceborne LiDAR waveforms are difficult to directly use for parameter inversion of a single vegetation type, this study aims to propose and validate an automatic separation method for tree and shrub lidar waveforms, in order to obtain independent waveform data for trees and shrubs.

Based on the data from the spaceborne LiDAR (Global ecosystem dynamics investigation, GEDI), this study performed Gaussian decomposition on the GEDI L1B field data after removing ground waveforms. The obtained Gaussian echoes were then clustered using spectral clustering methods to separately obtain arbor echoes and shrub echoes. Based on the principle of echo simulation, combined with ground measurements, a standardized reference dataset of arbor and shrub waveforms was simulated to verify and evaluate the clustering accuracy.

In the areas with slopes of [0°, 15°), [15°, 30°), and≥30°, the correlation coefficients between the tree waveform separation results and the simulation results were 0.89, 0.84, and 0.76, respectively, with root mean square errors of 0.021, 0.032, and 0.056, respectively. The correlation coefficients between the shrub waveform separation results and the simulation results were 0.82, 0.78, and 0.72, respectively, with root mean square errors of 0.024, 0.044, and 0.077, respectively. In the areas with canopy densities of [0, 30%), [30%, 60%), and [60%, 100%], the correlation coefficients between the tree waveform separation results and the simulation results were 0.79, 0.82, and 0.88, respectively, with root mean square errors of 0.055, 0.042, and 0.017, respectively. The correlation coefficients between the shrub waveform separation results and the simulation results were 0.85, 0.81, and 0.73, respectively, with root mean square errors of 0.031, 0.039, and 0.053, respectively. In the areas with shrub coverages of [0, 30%), [30%, 60%), and [60%, 100%], the correlation coefficients between the tree waveform separation results and the simulation results were 0.86, 0.82, and 0.81, respectively, with root mean square errors of 0.028, 0.029, and 0.033, respectively. The correlation coefficients between the shrub waveform separation results and the simulation results were 0.76, 0.84, and 0.88, respectively, with root mean square errors of 0.061, 0.041, and 0.028, respectively.

This paper proposes a tree-shrub waveform separation method based on spectral clustering, which eliminates the need for data annotation and elevation threshold presetting. This method can provide reliable technical support for analyzing the vertical structure of forests and monitoring carbon stock dynamics at the global scale.