Model-Based Conifer-Crown Surface Reconstruction from High-Resolution Aerial Images

Knowledge of tree-crown parameters such as height, shape, and crown closure is desimble in forest and ecological studies, but those pammeters are difficult to measure on the ground. The stereoscopic capability of high-resolution aerial images provides a method for crown-surface reconstruction. However. 'existing digital photogrammet~packages, designed to map terrain surfaces. cannot accuratelv extmct tree-crown surfaces. particularb for-conifer-crowns d t h steep vertical profilis. ' In this paper, we integrate crown features derived from images with stereo matching, and develop a model-based approach for reconstructing conifer-crown sqfaces. The model is based on the fact that most conifer crowns are a form of solid geometry. We model a conifer crown as a genemlized hemi-ellipsoid, establish the optimal tree model using a geometric equation, and apply the optimal tree model to guide a conventional pyramidal image matching in crown-surface reconstruction. The effectiveness of the approach is illustmted using an example of a redwood tree on 1:2,400-scale aerial photographs. Introduction A description of three-dimensional ( 3 ~ ) crown shape is useful in estimating the amount of foliage and the photosynthetic activity of trees. In forest inventory, parameters about crowns sometimes are collected such as diameter, height, closure, etc. Crown width (i.e., crown diameter) and height are important inputs to forest models (Deutschman et al., 1997), and are critical in modeling forest fires (Keane et al., 1999). It is a time-consuming and labor-intensive process to measure crown diameter and height in the field, let alone measure the 3D crown surface. This led us to develop photo-ecometrics techniques for forest inventory (Gong et al., 1999). Aerial photography provides a practical means for tree measurement. Large-scale aerial photographs have long been used for measuring such parameters. Andrews (1936) derived tree height from aerial images in a stereo oair as earlv as in the 1930s. Tree ~ height reading&om 1:1,000-kale aeridphotographs using a stereoplotter were found even more accurate than field measurements using tapes and clinometers (Kovats, 1997). Moessner (1949) developed a crown-density scale as a reference to interpret crown closure. Sayn-Wittgenstein (1961) applied crown characteristics (e.g., crown density, size, and marginal and apex shape) to tree species recognition. Crown information is estimated most easily from aerial photos. However, the approaches are mainly based on visual interpretation. As a consequence, the process is less efficient, is subjective, and is error-prone (Biging et al., 1991). Precise measurement of crowns from aerial photographs requires relatively accurate crown surface data. Due to the perspective view of aerial photographs, crown closure is overestimated when a stand is located far away from the principal points of the photographs. Theoretically, the parameters derived from orthophotos are free from the displacement influences. 3D crown surface data are needed to generate orthophotos from perspective photos. Another requirement for crown-surface data comes from automated photointerpretation for forestry applications. As both the computing power and spatial resolution of remotely sensed data improves, more attention is being paid to individual tree-based photointerpretation (Gougeon. 1993; Gougeon, 1995; Pollock, 1996; Larsen, 1998). Although high-resolution photos capture information on individual trees, the displacement of a crown surface is large when measured in pixels. Current work in this field has focused on tree delineation using spectral information from monocular images. When forest photointerpretation results are input into a geographic information system (GIS), the coordinate data contain errors due to the perspective view of aerial photography. To eliminate the geometric errors, we need to use the 3D crown coordinates to orthorectify the aerial photographs. Obviously, the marriage of crown-surface and spectral information will substantially benefit tree delineation, and crown-surface data can be used to produce orthographic tree maps for input into a GIs. Literature on crown-surface reconstruction is rather limited. Laser range detection (LIDAR: ~ ~ g h t Detecting ~ n d ~anging), radar interferometry, and photogrammetry are three major techniques for surface reconstruction. Airborne laser scanning systems detect range using laser pulses in the visible or near infrared wavelengths. To measure crown-surface heights, a LIDAR records multiple echoes; the first one is supposed to be reflected from the outer surface of a crown while the last from the ground after penetrating through the canopy. Crown-surface heights can be derived by subtracting the first range reading from the last. The penetration ability of a laser pulse is critical to the quality of surface reconstruction. It was found that, with near vertical incident angles of laser systems, 20 to 40 percent penetration rates were exCenter for Assessment and Monitoring of Forest and Environmental Resources, and Department of Environmental Science, Policy, and Management, 151 Hilgard Hall, University of California, Berkeley, CA 94720-3110 ([email protected]). P. Gong is also with International Institute for Earth System Science, Nanjing University, Nanjing 210093, China. Photogrammetric Engineering & Remote Sensing VO~. 67, NO. 8, August 2001, pp. 957-965. 0099-lll2/01/6708-957$3.00/0 O 2001 American Society for Photogrammetry