International Journal of Applied Earth Observation and Geoinformation 18 (2012) 101–110 Contents lists available at SciVerse ScienceDirect International Journal of Applied Earth Observation and Geoinformation jo u r n al hom epage: www.elsevier.com/locate/jag Investigating multiple data sources for tree species classification in temperate forest and use for single tree delineation Johannes Heinzel ∗ , Barbara Koch Department of Remote Sensing and Landscape Information Systems, Albert-Ludwig-University Freiburg, Tennenbacherstraße 4, 79106 Freiburg, Germany a r t i c l e i n f o Article history: Received 16 February 2011 Accepted 27 January 2012 Keywords: Full-waveform LiDAR Multispectral Support vector machines Tree species Single tree delineation a b s t r a c t Despite numerous studies existing for tree species classification the difficult situation in dense and mixed temperate forest is still a challenging task. This study attempts to extend the existing limitations by investigating comprehensive sets of different types of features derived from multiple data sources. These sets include features from full-waveform LiDAR, LiDAR height metrics, texture, hyperspectral data and colour infrared (CIR) images. Support vector machines (SVM) are used as an appropriate classifier to handle the high dimensional feature space and an internal ranking method allows the determination of the most important parameters. In addition, for species discrimination, focus is put on single tree applicable scale. While most experiences within these scales derive from boreal forests and are often restricted to two or three species, we concentrate on more complex temperate forests. The four main species pine (Pinus sylvestris), spruce (Picea abies), oak (Quercus petraea) and beech (Fagus sylvatica) are classified with an accuracy of 89.7%, 88.7%, 83.1% and 90.7%, respectively. Instead of directly classifying delineated single trees a raster cell based classification is conducted. This overcomes problems with erroneous polygons of merged tree crowns, which occur frequently within dense deciduous or mixed canopies. Lastly, we further test the possibility to correct these failures by combining species classification with single tree delineation. © 2012 Elsevier B.V. All rights reserved. 1. Introduction High resolution remote sensing data and modern processing techniques play an increasing role within the discipline of precision forestry. According to Bare (2003), the goal of precision forestry can be defined as the development of “tools and processes that increase the precision of forest data to support better decisions about forests, their services and products”. In this context, methods using high resolution remote sensing data often aim to provide monitoring tools on a single tree scale. Automated algorithms are being devel- oped which intend the extraction of single trees and their related parameters. One of the most important and basic parameters is the tree species. If correctly determined, the knowledge of high reso- lution species classes could improve the precision of economic or environmental information and go beyond common site-specific scale (Moskal et al., 2009). During the last decade, airborne LiDAR has become established as an important instrument in forestry application (Edson, 2011). In particular, full-waveform LiDAR allows the extraction of unique information from within the canopy (Wagner et al., 2008). Several ∗ Corresponding author. Tel.: +49 761 2038690; fax: +49 761 2033701. E-mail addresses: johannes.heinzel@felis.uni-freiburg.de, johannes.heinzel@gmx.de (J. Heinzel). approaches exist that use LiDAR derived features for tree species classification. Some authors utilize the intensity information of the reflected signal (Kim et al., 2009; Reitberger et al., 2008), while others compute density or height metrics which refer to the dis- tribution of laser beam reflections within the canopy (Holmgren and Persson, 2004; Ørka et al., 2009). LiDAR is also a favoured data source for single tree delineation and a number of methods have already been developed (Falkowski et al., 2006; Hyyppä et al., 2001; Koch et al., 2006). Besides LiDAR, passive data sources are also used to derive information on single tree scale. These sources include colour infrared (CIR) aerial images and hyperspectral data. CIR images play a traditional role in species classification (Leckie et al., 2005b; Meyer et al., 1996) due to the typical reflectance characteris- tics of vegetation in the near infrared (Hildebrandt, 1996). Attempts have also been made to derive geometric information from these images as a basis for single tree delineation (Brandtberg and Walter, 1998; Leckie et al., 2005a). Within recent studies, like those from Heikkinen et al. (2010) and Lucas et al. (2008), airborne captured hyperspectral data is also used for species classification. Hyper- spectral data allows the extraction of reflection characteristics of different species, but within a broader spectral range together with a higher level of spectral detail than CIR images. A particularly high number of studies dealing with species clas- sification on single tree level are conducted on test sites located in boreal forests (Holmgren and Persson, 2004; Ørka et al., 2009). 0303-2434/$ – see front matter © 2012 Elsevier B.V. All rights reserved. doi:10.1016/j.jag.2012.01.025