Agricultural and Forest Meteorology 171–172 (2013) 104–114 Contents lists available at SciVerse ScienceDirect Agricultural and Forest Meteorology jou rn al h om epa ge: www.elsevier.com/locate/agrformet Classification of tree species based on structural features derived from high density LiDAR data Jili Li a,∗ , Baoxin Hu a , Thomas L. Noland b a Department of Earth and Space Science and Engineering, York University, Toronto, Ontario M3J 1P3, Canada b Ontario Ministry of Natural Resources, Ontario Forest Research Institute, 1235 Queen St. East, Sault Ste. Marie, Ontario P6A 2E5, Canada a r t i c l e i n f o Article history: Received 22 May 2012 Received in revised form 12 November 2012 Accepted 17 November 2012 Keywords: Remote sensing LiDAR Forestry Species classification Point pattern Genetic algorithm Linear discriminant analysis a b s t r a c t Automated tree species classification using high density airborne light detection and ranging (LiDAR) data will support more precise forest inventory but further research is required to improve the associated methods. Most existing methods rely on geometric and vertical distribution features, which often do not accurately represent the internal foliage and branch patterns of an individual tree. Our study objective was to develop novel algorithms to characterize internal structures of an individual tree crown and to test their effectiveness for use in classifying tree species. We derived several LiDAR features to describe the three-dimensional texture, foliage clustering degree relative to tree envelop, foliage clustering scale, and gap distribution of an individual tree in both horizontal and vertical directions. Features were selected using a genetic algorithm and then tree species were classified using linear discriminant analysis based on the selected features. The four species, sugar maple (Acer saccharum Marsh.), trembling aspen (Populus tremuloides Michx.), jack pine (Pinus banksiana Lamb.) and eastern white pine (Pinus strobus L.), were classified with an overall accuracy of 77.5% and a Kappa coefficient of 0.7. The results demonstrate the significance of the derived structural features as aids to classify tree species. Our investigation also showed a positive linear correlation (R 2 = 0.88) between LiDAR point density and species classification accuracy. © 2012 Elsevier B.V. All rights reserved. 1. Introduction Accurate characterization of forest species and their spatial distribution is critical for sustainable forest management and for ecological and environmental protection. The estimation of biomass, carbon content, species diversity, and condition of a for- est community requires precise individual-tree information by species. Since the early stages of remote sensing, its use as a cost- effective way to classify forest species has been an ongoing topic of discussion (Franklin, 1994; Martin et al., 1998; Franklin et al., 2000; Key et al., 2001). Existing methods for classifying forest species from remote sensing data are mostly based on the spectral information from forest canopies. Despite vegetation cover classifi- cation successes at the stand- and landscape-level, the accuracy of individual-tree classification remains low. This is attributed mainly to: (1) the effects of several spectral and spatial factors on sur- face reflectance values of forest canopies and (2) limitation in the spectral and spatial configurations of image sensors. The availability of light detection and ranging (LiDAR) instru- ments to measure three-dimensional (3-D) positions of tree elements, such as foliage and branches, provides an opportunity ∗ Corresponding author. Tel.: +1 4167362100x20647. E-mail address: ljili2008@gmail.com (J. Li). to significantly improve forest species classification accuracy. Tree species differ in their foliage distribution and branching patterns, resulting in divergent architectures. For example, most aspen (Pop- ulous spp.) foliage is clustered near the top of the stem, while maple (Acer spp.) foliage is more evenly distributed along the stem. As a result, the structural features of a tree characterized using many LiDAR data points within the crown may increase an interpreter’s ability to accurately identify tree species. The challenge is to extract and select key diagnostic features among numerous species from the huge volume of LiDAR data. For the past decade, researchers have examined the potential to use data from airborne LiDAR to classify forest stand types or individual species (Brandtberg et al., 2003; Holmgren and Persson, 2004; Moffiet et al., 2005; Brenan and Webster, 2006; Reitberger et al., 2008; Ørka et al., 2009; Kim et al., 2011). Several LiDAR fea- tures have been extracted to describe crown structural properties of individual trees, such as crown shape and vertical foliage dis- tribution. Crown top sharpness and symmetry have been the most commonly derived features related to crown shape. These features were usually calculated based on the parameters of a 3-D surface model fitted to the LiDAR points within a given tree (e.g., Holmgren and Persson, 2004; Holmgren et al., 2008; Reitberger et al., 2008). Because they may be too inflexible to correctly model crown shape (Reitberger et al., 2008), concerns have been raised about whether LiDAR-derived surface models represent crown shape features 0168-1923/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.agrformet.2012.11.012