Classifying species of individual trees by intensity and structure features derived from airborne laser scanner data

Abstract

The objective of this study was to identify candidate features derived from airborne laser scanner (ALS) data suitable to discriminate between coniferous and deciduous tree species. Both features related to structure and intensity were considered. The study was conducted on 197 Norway spruce and 180 birch trees (leaves on conditions) in a boreal forest reserve in Norway. The ALS sensor used was capable of recording multiple echoes. The point density was 6.6 m^− 2. Laser echoes located within the vertical projection of the tree crowns, which were assumed to be circular and defined according to field measurements, were attributed to three categories: “first echoes of many”, “single echoes”, or “last echoes of many echoes”. They were denoted FIRST, SINGLE, and LAST, respectively. In tree species classification using ALS data features should be independent of tree heights. We found that many features were dependent on tree height and that this dependency influenced selection of candidate features. When we accounted for this dependency, it was revealed that FIRST and SINGLE echoes were located higher and LAST echoes lower in the birch crowns than in spruce crowns. The intensity features of the FIRST echoes differed more between species than corresponding features of the other echo categories. For the FIRST echoes the intensity values tended to be higher for birch than spruce. When using the various features for species classification, maximum overall classification accuracies of 77% and 73% were obtained for structural and intensity features, respectively. Combining candidate features related to structure and intensity resulted in an overall classification accuracy of 88%.

Introduction

In recent years, high resolution sampling density airborne laser scanning (ALS) has become readily available, providing x, y, z point datasets with 5–20 height measurements per square meter. Such data are useful for terrain, vegetation, and forest mapping. From these dense point clouds, individual trees can be identified by means of various segmentation procedures. These procedures extract the outline of the tree crowns. Individual tree segmentation is often done by using an ALS-derived canopy height model (e.g. Hyyppä et al., 2001, Persson et al., 2002, Solberg et al., 2006), but also other methods are used, like for example clustering (Morsdorf et al., 2004). When the outline of a tree crown is defined, laser echoes inside the segment can be tied to the tree and information about the tree such as stem position, height, and stem diameter can be derived (e.g. Persson et al., 2002, Solberg et al., 2006). This high resolution tree information can form a basis for forest planning by aggregating information to management units (Hyyppä et al., 2001).

Tree species is another parameter that may be derived from laser echoes inside individual tree segments. Species classification on a individual tree level using ALS-derived features has been accomplished in boreal forest in Scandinavia (Holmgren et al., 2008, Holmgren and Persson, 2004, Liang et al., 2007), in mixed coniferous and deciduous forest in central Europe (Heurich, 2006, Reitberger et al., 2008), in deciduous forest in western Virginia (Brandtberg, 2007, Brandtberg et al., 2003), and in sub-tropical forest in Queensland, Australia (Moffiet et al., 2005). Individual tree species information could also be found using high spatial resolution images (e.g. Brandtberg, 2002, Carleer and Wolff, 2004, Key et al., 2001, Olofsson et al., 2006). However, acquisition of both ALS data and imagery will increase inventory costs. Furthermore, because ALS provides more accurate estimates of biomass and height compared to image remote sensing methods (Hyde et al., 2006, Hyyppä and Hyyppä, 1999), the possibilities of utilize ALS data also to discriminate between tree species are of interest in order to control data acquisition cost.

Structural features of the tree crowns can be derived from ALS height measurements and such features might be considered for tree species classification. The basic idea behind using structural features for tree species classification is that different species have different crown properties such as crown shape, reflectivity, and location of biomass. For example, crown shapes for spruce trees tend to be conical, whereas more spherical or rounded shapes are found for deciduous trees. Deciduous trees also tend to allocate more biomass higher in the crown. The structural differences of tree crowns will influence on the recorded laser echoes. When a laser echo is recorded, the elapsed time between emission and receipt of a significant amount of returned energy is converted to range. Since the position and orientation of the platform are known by Global Navigation Satellite Systems (GNSS) and Inertial Navigation System (INS), the position of the target can be calculated. To trigger a laser echo from a tree crown or any other surface, the properties of the surface hit by the laser pulse is of importance. One example is the high rate of success in detecting power lines. A power line covers just a small portion of a laser footprint, but is still detectable in an ALS dataset because of the high reflectivity of power lines. On the other hand, a tree crown surface represented by branches and leaves often covering the entire laser footprint has lower reflectivity and a different structure. The laser pulse will therefore tend to penetrate into the canopy before a significant echo is recorded by the sensor (Gaveau & Hill, 2003). Thus, different crown properties affect the distribution of laser echoes within and on the surface of the tree crowns. This may lead to distinct echo height distributions for separate species. Therefore, it might be useful for automated species classification based on ALS data to identify which structural features derived from the echo height distribution that are most suited to distinguish species.

In addition to the spatial coordinates of laser echoes, most ALS systems measure the intensity of the backscattered laser signal (Wehr & Lohr, 1999). For pulse lasers, intensity often represents the peak amplitude of the returned pulse. It is expected that this value could assist species classification. Already in 1985, Schreier et al. (1985) demonstrated classification of individual trees into conifers and broadleaves partly based on airborne laser intensity. Since then the use of laser intensity has been little explored. This is mainly because of lack of methods for radiometric calibration of intensity values (Kaasalainen et al., 2005). However, recently some authors have tested intensity features for tree species classification (Brandtberg, 2007, Brandtberg et al., 2003, Holmgren et al., 2008, Holmgren and Persson, 2004, Moffiet et al., 2005, Reitberger et al., 2008) and for discerning age classes (Farid et al., 2006a, Farid et al., 2006b) as well as land-cover classes (Brennan & Webster, 2006) where deciduous and coniferous forest were treated as separate classes. Despite the lack of calibration methods, intensity features derived from ALS data may improve classification (e.g. Brandtberg et al., 2003, Holmgren et al., 2008). As methods for calibration of the intensity mature, the usefulness of intensity used for individual tree species classification may increase.

In species classification, features derived from the laser height distribution, such as the mean height of the laser echoes, could be used directly in the classification algorithm (e.g. Brennan & Webster, 2006) or as a scaled feature, for example normalized with tree height (Holmgren & Persson, 2004). In individual tree classification, independence of tree height is important, especially in forests where tree height distributions differ between species. To ensure this independence features should be scaled. Brandtberg (2007) normalized the 3D point cloud using estimated tree height to ensure independence. Holmgren and Persson (2004) used relative height features, i.e., laser height features divided by the laser estimated height, to separate Norway spruce and Scots pine. It should be noted, however, that it has so far not been tested if scaling methods really produce independence of tree height. Robust scaling may be important for practical applications covering large areas. In large forested landscapes, species-specific height distributions will vary in the landscape according to soil properties, management history, and a number of other factors. Hence, selection of robust and unbiased classification features is important.

The aim of this study was to identify candidate ALS-derived features suitable for classification of spruce and birch. In order to reach our aim, we (1) conducted an analysis of differences in (1a) structural- and (1b) intensity features between spruce and birch trees, and (2) tested the classification performance of candidate features.