In this paper, we propose a new method for specifying individual tree species based on depth and curvature image creation from point cloud captured by terrestrial laser scanner and Convolutional Neural Network (CNN). Given a point cloud of an individual tree, the proposed method first extracts the subset of points corresponding to a trunk at breast-height. Then branches and leaves are removed from the extracted points by RANSAC -based circle fitting, and the depth image is created by globally fitting a cubic polynomial surface to the remaining trunk points. Furthermore, principal curvatures are estimated at each scanned point by locally fitting a quadratic surface to its neighbouring points. Depth images clearly capture the bark texture involved by its split and tear-off, but its computation is unstable and may fail to acquire bark shape in the resulting images. In contrast, curvature estimation enables stable computation of surface concavity and convexity, and thus it can well represent local geometry of bark texture in the curvature images. In comparison to the depth image, the curvature image enables accurate classification for slanted trees with many branches and leaves. We also evaluated the effectiveness of a multi-modal approach for species classification in which depth and curvature images are analysed together using CNN and support vector machine. We verified the superior performance of our proposed method for point cloud of Japanese cedar and cypress trees.
展开▼
