Retrieval of higher order statistical moments from full-waveform LiDAR data for tree species classification

Abstract

Tree inventories, inter alia, need information on the tree species. Previous studies aimed at distinguishing tree species based on three dimensional tree structure metrics derived from airborne laser scanning (ALS) point clouds or based on features from full-waveform (FW) laser scanning data provided by today's sensors. Classifications based on FW features mainly use echo amplitude, pulse energy (hereafter referred to as energy) and width, which are typically retrieved by waveform decomposition, often performed using the symmetric Gaussian distribution function.

However, for forested areas, the symmetry of the echo shape is potentially modified by multiple scattering and the distribution of scattering elements (e.g. leaves). We assess the potential of processing full-waveform ALS data such that the third and fourth statistical moments, i.e. the echo skewness and the echo kurtosis, can be retrieved in addition to the amplitude, energy and FWHM. We propose a waveform decomposition approach using the skew normal distribution (SND) function, which enables the modelling of skewed echoes. We investigate the difference of tree-crown aggregated SND derived FW features between seven tree species (969 individual trees) in a temperate mixed forest with the aim of detecting the most descriptive echo features. The such derived FW features are tested for species classification.

The results reveal that the largest differences across the tree species are in the mean energy of the first echoes (15 out of 21 species pairs show differences), followed by the mean amplitude of the first echoes and the mean skewness of all echoes originating from a single crown (14 out of 21 species pairs show differences against each other for both features).

The differentiation of coniferous and deciduous trees benefits from the features derived from the SND decomposition compared to the use of echo amplitude only (0.39 vs. 0.61 in Cohen's κ). As the classification accuracy of the three dominant tree species within the test site only shows a small increase (0.20 vs. 0.26 in Cohen's κ) by adding FW features, we propose the use of such features in combination with features from multispectral data for this purpose.

The SND decomposition is comparable to the Gaussian decomposition regarding the decomposition accuracy (RMSE = 4.45 vs. RMSE = 3.50) and computational cost. Hence, we propose the default use of the SND decomposition, as the SND is a more flexible function, allowing for the modelling of normally distributed echoes, as well as the fitting of skewed echoes, while no limitations regarding the direction of the skewness are introduced. We attribute the difficulties in the tree species classification to the relatively wide ranges of the crown aggregated features within one species, which for some features results in a considerable overlap of the feature ranges across the species.