Individual Tree-Crown Delineation and Treetop Detection in High-Spatial-Resolution Aerial Imagery

The cost of forest sampling can be reduced substantially by the ability to estimate forest and tree parameters directly from aerial photographs. However, in order to do so it is necessary to be able to accurately identify individual treetops and then to define the region in the vicinity of the treetop that encompasses the crown extent. These two steps commonly have been treated independently. In this paper, we derive individual tree-crown boundaries and treetop locations under a unified framework. We applied a two-stage approach with edge detection followed by markercontrolled watershed segmentation. A Laplacian of Gaussian edge detection method at the smallest effective scale was employed to mask out the background. An eight-connectivity scheme was used to label the remaining tree objects in the edge map. Subsequently, treetops are modeled based on both radiometry and geometry. More specifically, treetops are assumed to be represented by local radiation maxima and also to be located near the center of the tree-crown. As a result, a marker image was created from the derived treetop to guide a watershed segmentation to further differentiate touching and clumping trees and to produce a segmented image comprised of individual tree crowns. Our methods were developed on a 256by 256-pixel CASI image of a commercially thinned trial forest. A promising agreement between our automatic methods and manual delineation results was achieved in counting the number of trees as well as in delineating tree crowns. Introduction Modern forest management requires that forest resources be efficiently managed, not only for timber production, but also for such purposes as maintaining biodiversity and meeting wildlife, environmental, and recreational needs. Accordingly, there is an increasing need for detailed knowledge of forest stands, which are the basic units for forest management. Inevitably, stand measurement involves the measurement of individual trees within the stand. The traditional method for deriving stand information is to utilize sampling designs with transects, random or systematically selected plots, so that the final stand parameters can be derived based on statistical extrapolation methods. By utilizing remote sensing data, we can reduce the amount of field sampling; hence, information gathering becomes more cost-effective. In the 1940s, manual interpretation of mediumand largescale aerial imagery for forestry emerged (Brandtberg, 1999). Individual Tree-Crown Delineation and Treetop Detection in High-Spatial-Resolution Aerial Imagery Le Wang, Peng Gong, and Gregory S. Biging Since then, field inventory in combination with aerial photointerpretation has played an important role in forest data collection. The visual aerial photointerpretation method is labor-intensive, time-consuming, and dependent on the interpreter’s experience. Thus, there is merit for developing an automated or semiautomated aerial photo measurement method for forest trees. With the increasing availability of large-scale and highresolution imagery, a new round of research on computerbased photointerpretation of trees was recently initiated (Gong et al., 1999). Various algorithms have been developed for automatic individual tree recognition. They can be grouped into four major types: local maximum (LM)-based methods (Blazquez, 1989; Dralle and Rudemo, 1996), contour-based (CB) methods (Pinz et al., 1993; Gougeon, 1995), template-matching (TM)-based methods (Pollock, 1996; Tarp-Johansen, 2002), and 3D-model-based methods (Sheng et al., 2001; Gong et al., 2002). The LM method makes the assumption that the peak of the tree-crown reflectance is located at or very close to the treetop (Brandtberg and Walter, 1998). Therefore, by filtering the image to find the local maximum, treetops are finally detected. An image-smoothing step can be introduced in this method to reduce the noise effect (Dralle and Rudemo, 1996). In addition, LM methods can be combined with an advanced region-based analysis of the image objects (Pinz et al., 1993). Although this method has the merit of being fast and simple, it performs poorly when undesirable background phenomena and varying illumination conditions exist in the image. The TM method includes a model generation and a template-matching procedure (Pollock, 1996). Intuitively, a series of models are built to characterize what a tree looks like at different locations in an image by taking into consideration both the trees’ geometric and radiometric properties. Once this knowledge is gained, a moving-window correlation procedure is implemented to search for the locus of best matching where trees are most likely to be. From another perspective, the CB method attempts to find the delimiter between tree crowns and their background. Briefly, the main strategy here is either to follow the intensity valleys underlying the image (Gougeon, 1995) or to detect the crown boundary with edge-detection methods (Brandtberg and Walter, 1998). For the valley-following method, a set of rules is predefined before the actual crown following takes place. Few people endeavor to use the edge-detection method P H OTO G R A M M E T R I C E N G I N E E R I N G & R E M OT E S E N S I N G March 2004 3 5 1 Center for Assessment and Monitoring of Forest and Environmental Resources (CAMFER), 145 Mulford Hall, University of California, Berkeley, CA 94720-3114 ([email protected]) Photogrammetric Engineering & Remote Sensing Vol. 70, No. 3, March 2004, pp. 351–357. 0099-1112/04/7003–0351/$3.00/0 © 2004 American Society for Photogrammetry and Remote Sensing 01-135.qxd 2/4/04 11:45 AM Page 351