LAUSR

Abstract Tree species identification is important in forest management. The multispectral lidar Titan of Teledyne Optech Inc. can improve tree species separation by providing classification features computed from the three-channel intensities, ratios and normalized differences. However, the value of features used in classification algorithms (e.g., random forest, RF) may vary with tree size. The focus of the present study is to show how tree height influences the 3D and intensity features, how this relationship may affect the species classification accuracy, and how different classification strategies may circumvent this problem. Six needleleaf species (Pinus resinosa, Pinus strobus, Pinus sylvestris, Larix laricina, Picea abies and Picea glauca), found in plantations of different ages, were sampled to train classifiers. Some features yielded a good discriminatory power for species identification, despite their relation to tree height (r2 up to 0.6). Two classification strategies—a) using only size-invariant features (SIF) and b) training separate classifiers per tree height strata (HSC)—were compared to a standard classification (STD: all features, without height stratification). The accuracy of the SIF approach was lowest, useful variables being removed due to their relationship to tree height. The HSC provided only a minor improvement over the STD results.