Tree Species Identification in Mixed Baltic Forest Using LiDAR and Multispectral Data

This work describes the task of inventorying Baltic mixed forests at an individual tree level. The development of a practicable methodology for semi-automated identification of tree species was targeted. Data acquisition equipment and preprocessing software, explored forest area, processing approaches, obtained classification results as well as newly developed software are described. To resolve the core problem - tree species identification - a classification approach is proposed for processing multi-spectral imagery data from the vicinity of tree tops. A multi-class classifier is designed from multi-spectral data of interactively selected trees included in initial design (training) sets for two conifer and three deciduous species of interest. An approach for the stabilization of the classification results is proposed, based on improving the representativeness of the design sets by selection of trees from different locations, dismissing trees with overlapping crowns and anomalies, followed by the calculation of a spectral dissimilarity parameter of the design sets and dismissing the sets of trees of any species which are too similar. The best classification results were obtained using a two-stage procedure. In the first stage, species clusters were created by adding randomly selected trees from the whole analyzed forest area. Final classification of all trees was done by using a Bayes classifier designed on the basis of cluster properties. A procedure for increasing robustness of the clustering stage is proposed, based on performing multiple clustering attempts, each using a randomly sampled subset of a chosen design set for the classifier design, and making a decision about the class of each tree by the majority vote from the results of these attempts. This classification algorithm was tested against the set of trees, for which information was available from field work. It is shown that a mean classification error below 3% can be achieved and the maximum error rate was decreased substantially by applying the proposed approach for selection of representative design sets.