LAUSR

Abstract Determining tree species composition in natural forests is essential for effective forest management. Species classification at the individual tree level requires fine-scale traits which can be derived through terrestrial laser scanning (TLS) point clouds. A generalizable species classification framework also needs to decouple seasonal foliage variation from deciduous species, for which wood filtering is applicable. Different machine learning and deep learning models are feasible for wood filtering and species classification. We investigated 13 machine learning and deep learning classifiers for 9 species, and 15 classifiers for filtering wood points from TLS plot scans. Each classifier was evaluated using the criteria of mean Intersection over Union accuracy (mIoU), training stability and time cost. On average, deep learning classifiers outperformed machine learning classifiers by 10% and 5% in terms of wood and species classification mIoU, respectively. PointNet++ provided the best species classifier, with the highest mIoU (0.906), stability, and moderate time cost. Among wood classifiers, UNet achieved the top mIoU (0.839) while ResNet-50 was recommended for rapid trial and error testing. Across the classifications, the factors of input resolution, attributes and features were also analyzed. Hot zones of species classification with PointNet++ were visualized to indicate how AI interpret species traits.