6533b835fe1ef96bd129f0eb

RESEARCH PRODUCT

Automatic recognition of tree species from 3D point clouds of forest plots

subject

description

The objective of the thesis is the automatic recognition of tree species from Terrestrial LiDAR data. This information is essential for forest inventory. As an answer, we propose different recognition methods based on the 3D geometric texture of the bark.These methods use the following processing steps: a preprocessing step, a segmentation step, a feature extraction step and a final classification step. They are based on the 3D data or on depth images built from 3D point clouds of tree trunks using a reference surface.We have investigated and tested several segmentation approaches on depth images representing the geometric texture of the bark. These approaches have the disadvantages of over segmentation and are quite sensitive to noises. For this reason, we propose a new 3D point cloud segmentation approach inspired by the watershed technique that we have called «Burst Wind Segmentation». Our approach succeed in extracting in most cases the characteristic scars that are next compared to those stored in a dictionary («ScarBook») in order to determine the tree species.A large variety of characteristics is extracted from the regions segmented by the different methods proposed. These characteristics are the roughness, the global shape of the segmented regions, the saliency and the curvature of the contour, the distribution of the contour points, the distribution of the shape according to the different orientations.Finally, for the classification of the visual characteristics, the Random Forest method by Leo Breiman and Adèle Cutler is used in a two steps approach: selection of the most important variables and cross classification with the selected variables.The bark of the tree changes with the trunk diameter. We have thus studied different natural variability criteria and we have tested our approaches on a test set that includes this variability. The accuracy rate is over 96% for all the proposed segmentation approaches but the best result is obtained with the «Burst Wind Segmentation» one due to the fact that this approach can better extract the scars, it uses a dictionary of scars for recognition, and it has been evaluated on a greater variety of shapes, curvatures, saliency and roughness.