Many biophysical forest properties such as wood volume and leaf area index (LAI) require prior knowledge on either photosynthetic or non-photosynthetic components. Laser scanning appears to be a helpful technique in nondestructively quantifying forest structures, as it can acquire an accurate three-dimensional point cloud of objects. In this study, we propose an unsupervised geometry-based method named Dynamic Segment Merging (DSM) to identify non-photosynthetic components of trees by semantically segmenting tree point clouds, and examining the linear shape prior of each resulting segment. We tested our method using one single tree dataset and four plot-level datasets, and compared our results to a supervised machine learning method. We further demonstrated that by using an optimal neighborhood selection method that involves multi-scale analysis, the results were improved. Our results showed that the overall accuracy ranged from 81.8% to 92.0% with an average value of 87.7%. The supervised machine learning method had an average overall accuracy of 86.4% for all datasets, on account of a collection of manually delineated representative training data. Our study indicates that separating tree photosynthetic and non-photosynthetic components from laser scanning data can be achieved in a fully unsupervised manner without the need of training data and user intervention.