LiDAR waveform features for tree species classification and their sensitivity to tree- and acquisition related parameters

Abstract

Abstract This study looked to investigate the basic properties and species classification performance of waveform (WF) features in LiDAR data for the main tree species in Finland. We conducted experimental research using four LiDAR datasets and individual Scots pine, Norway spruce, and silver/downy birch trees. The classification performance and the importance of features were evaluated for dominant/co-dominant trees (N = 9930), and compared to a subset of trees (N = 3630) that had well separable crowns. We used data obtained from two discrete-return (DR) Leica ALS60 sensors, in which the first echo triggers the recording of a WF sequence. Using experience with simulated WF data, we defined a set of informative and technically simple WF features. Quadratic discriminant analysis was applied to classify tree species and also to discover the most important WF features. The WF features outperformed the DR intensity data (0.57–0.75 vs. 0.74–0.86 in Cohen's kappa). The total backscattered energy (E) of single returning WF sequences was the most important feature. Early summer data outperformed late summer data, and we observed differences in the noise characteristics of individual sensors. We performed analyses of the best-performing mean WF features, using linear mixed-effects modeling. The non-quantified variation in tree structure explained 13–65% of within-species feature variance, while dataset–species interaction (i.e. the dependence of species effect on LiDAR dataset), tree height, age, site type, and scan zenith angle accounted for only 2–25%. The residual variance of each feature was 13–84% and depended on the number of pulses. This dependence was weakest for E, implying good performance at low pulse densities. Further analysis showed how intra-species variations in crown and branch morphology and vigor were logically linked with WF features. We conclude that improvements in species classification may be obtained by a stratification according to tree height and by acquiring data during early summer. The inherent between-tree and within-species variation in geometric-optical properties sets an upper limit for classification performance.