Tree Genera Classification by Ensemble Classification of Small-Footprint Airborne LiDAR

Tree genera information is useful in environmental applications such as forest management, forestry, urban planning, and the maintenance of utility transmission line infrastructure. The ability of small foot print airborne LiDAR (Light Detection and Ranging) to acquire 3D information provides a promising way of studying vertical forest structures. This provides an extra dimension of information compared to the traditional 2D remote sensing data. However, the techniques for processing this type of data are relatively recent and have becoming an innovative research direction. The existing perspective for processing LiDAR data for tree species classification involve calculating the statistics attributes of the vertical point profile for individual trees. This method however does not explicitly utilize the geometric information of the tree form such as shapes of the tree crown and geometric features that are derivable inside of the tree crown. Therefore, the aim of this dissertation research is to derive geometric features from individual tree crowns and use these features for genera classification. The second goal of this research is to improve classification results by combining the newly developed features with the conventional vertical point profile features through ensemble classification system. Final goal of this research is to design a classification system to cope with the situation where the number of classes in the validation data exceeds the number of classes in the training data. 24 geometric features were initially derived and six of them are selected for the classification of pine, poplar and maple. Average classification accuracy of 88.3% is achieved by using this method. When the geometric features are combined with vertical profile features by ensemble classification system, the average classification accuracy increased to 91.2%. While the individual performance of geometric classifier and vertical classifier is 88.0% and 88.8% respectively for the classification of pine, poplar and maple. Lastly, when samples that do not belong to pine, poplar and maple are added to the validation data, the classification accuracy dropped to 72.8% by using randomly selected samples for training. However, through diversified sampling technique, the classification accuracy increased to 93.8%.