3D Tree Dimensionality Assessment Using Photogrammetry and Small Unmanned Aerial Vehicles

Abstract

Detailed, precise, three-dimensional (3D) representations of individual trees are a prerequisite for an accurate assessment of tree competition, growth, and morphological plasticity. Until recently, our ability to measure the dimensionality, spatial arrangement, shape of trees, and shape of tree components with precision has been constrained by technological and logistical limitations and cost. Traditional methods of forest biometrics provide only partial measurements and are labor intensive. Active remote technologies such as LiDAR operated from airborne platforms provide only partial crown reconstructions. The use of terrestrial LiDAR is laborious, has portability limitations and high cost. In this work we capitalized on recent improvements in the capabilities and availability of small unmanned aerial vehicles (UAVs), light and inexpensive cameras, and developed an affordable method for obtaining precise and comprehensive 3D models of trees and small groups of trees. The method employs slow-moving UAVs that acquire images along predefined trajectories near and around targeted trees, and computer vision-based approaches that process the images to obtain detailed tree reconstructions. After we confirmed the potential of the methodology via simulation we evaluated several UAV platforms, strategies for image acquisition, and image processing algorithms. We present an original, step-by-step workflow which utilizes open source programs and original software. We anticipate that future development and applications of our method will improve our understanding of forest self-organization emerging from the competition among trees, and will lead to a refined generation of individual-tree-based forest models.

Fig 1. SIFT-based scene keypoint detection and matching on two overlapping images.

Top: Original images; Middle: 1464 (left) and 1477 (right) keypoints with arrows denoting orientation and radii scale; Bottom: 157 keypoint pairs, matched by color and number.

Fig 2. Removal of lens distortion.

Demonstration of a. original, vs. b. OpenCV-calibrated lateral tree image obtained with a UAV-based GoPro camera at an above-ground altitude of 18 meters. Horizontal red line drawn to illustrate form of horizon in each version of the image.

Fig 3. 3D reconstruction in simulation.

a. Perspective view of point cloud acquired with terrestrial LiDAR and camera locations (red spheres) used to obtain virtual images of the scene. b. Scene reconstruction obtained by processing of the images.

Fig 4. Custom built UAV hexacopter used to collect imagery data in this study.

Fig 5. Different UAV trajectories tested for image acquisition.

a. circular, at constant height; b. ‘stacked circles’, each at different above-ground height, for tall trees (height more than 20 m); c. spiral, for trees with complex geometry; d. vertical meandering, targeting tree sectors; e. clover, for trees with wide, ellipsoidal tree crowns; f. ‘spring-hemisphere’, designed for trees with flat-top, asymmetrical crowns; g. ‘nested circles’, centered on the tree; and h. ‘jagged saucer’, designed for trees with dense foliage but low crown compaction ratio.

Fig 6. Visualization of designed and accomplished UAV trajectories.

a. and c. circular and clover templates as seen in Mission Planner with yellow lines showing the flight paths, green balloons indicating waypoints, and red balloons the center of targeted trees. b. and d. perspective scene view in Google Earth, with yellow pins indicating camera locations along each trajectory at the moment images were captured.

Fig 7. Accuracy and completeness of reconstruction for a Pinus ponderosa tree.

This analysis is based on synthetic imagery simulated using visualization of terrestrial LiDAR point clouds and two camera trajectories. Percentage of collocated filled voxels is used as reconstruction completeness criterion.

Fig 8. Orthographic horizontal view of reconstructed point cloud and UAV-based oblique perspective image.

Colored arrows denote corresponding tree crown components.

Fig 9. Illustration of comprehensive tree reconstructions (right column) and reference UAV-based images (left column).

Fig 10. Demonstration of artifacts in the 3D tree reconstruction pertaining to a single UAV image.

a. Initial reconstruction, positioned facing the camera with a band of white-colored points belonging to sky background near the top, and light colored points to the sides belonging to fallow land background, b. Side view, with camera position to the left and sky points in oval and land points in rectangle, and c. Trimmed reconstruction positioned facing the camera.

Fig 11. Comparison between real and reconstructed trajectory.

Nadir view of reconstructed tree with camera GPS locations at image frame acquisition moments (yellow circles) and VSfM-calculated locations (red dots). Frame frequency 2Hz, GPS fixes at 1Hz. Inset at the lower left shows lateral view of the reconstructed tree.