By Cornelis Valk (The Netherlands)
In this study a method is presented for evaluating performance of T-LiDAR tree separation algorithms, which is tested by selecting, implementing, running and analysing three tree separation algorithms. First a literature study was conducted to obtain an overview of the available algorithms and to make a selection representing different algorithm classes. The three selected algorithms are: Voxel-based label connected components (LCC), Voxel-based normalized cut (Ncut) and Point-based region growing (Pbased). The performance evaluation method uses four test datasets to which the tree separation algorithms are applied, and for which a number of ground truth trees were manually extracted. The performance evaluation method analyses the tree separation results using a number of performance measures, some expressing general performance or shape similarity, some providing more detailed information about local variation in algorithm performance. A performance survey was completed by 9 people knowledgeable about LiDAR to provide unbiased performance indications for comparison with the performance evaluation method results. The F-score was found to be a reliable measure of overall performance. No relation was found between shape-similarity measures and the performance as estimated visually by the survey respondents. The measures providing insight in local variation of performance proved useful for detecting flaws in and weaknesses of the algorithms. Overall LCC was the best algorithm for tree separation, though an F-score between 0.35 and 0.78 indicates that a lot needs to improve. Despite the fact that Ncut was not correctly implemented, it obtained a second place. Pbased is clearly not suited for tree separation from T-LiDAR scans, though its results can be used for tree height estimation in structurally simple forests. For all three algorithms areas of improvement were identified and suggestions for future development are discussed.