Comparison of Single Tree Species Classification Using Very Dense ALS Data or Dual-Wave ALS Data

In this study, we compared the classification accuracy at plot-level of three tree species in a boreal forest using a deep learning (DL) network applied to very high resolution (VHR) ALS data acquired with ~593 points/m
²
, or a linear discriminant analysis (LDA) method applied to dual-wavelength (DW) ALS data, acquired with ~80 points/m
²
.The methods were applied to the single trees, which were aggregated to the plot-level (10 m radius) where the majority class was used for assessing the accuracy. The overall best results were obtained with LDA, relying on DW data, with an overall accuracy (OA) of 82.1% (n=28 plots), while for the DL method using VHR data, the best OA=75.0% (n=28 plots). We acknowledge that the point density of DW data (used for LDA) is already relatively high, and the benefit of using additional spectral information is therefore higher in our study than the value of increasing the point density and identify the tree species from geometric properties (DL approach).We conclude that both dense mono-wavelength ALS data and DW ALS data contain enough information that tree species can be classified at the singletree level in this boreal forest test site in Sweden. With the point densities in this study, the classifications were more accurate with DW data, while the DL approach may be improved with a refined segmentation and pre-processing approach.