A TLS point cloud dataset with detailed annotations for leaf-on and leaf-off seasons across multiple tree species - HuashuTrees

Terrestrial laser scanning (TLS) captures accurate 3D tree geometry, benefiting forest management and ecology. Deep learning has enhanced TLS point cloud processing but requires extensive training data with semantic annotations. Existing datasets typically focus on tree segmentation or species classification, lacking detailed leaf-wood separation, which is essential for carbon stock and photosynthesis modelling. Real-world TLS data with leaf-wood labels remain scarce, come from limited regions (e.g., Cameroon, Finland, USA, India), and are mostly leaf-on only. To fill this gap, we introduce HuashuTrees, a public multi-species bi-seasonal TLS dataset from deciduous broadleaf forests in Nanjing, China (leaf-off: January 2025; leaf-on: July 2025). Other than plot-level point clouds, it includes 318 single-tree point clouds from 159 trees across seven species, generated after registration, denoising, segmentation, leaf-wood separation, and manual correction. Technical validation confirms high consistency for leaf-wood annotation, with a mean agreement of 96.34% between the final dataset labels and independent manual re-classification. Additionally, TLS-derived diameter at breast height (DBH) shows strong agreement with field measurements (RMSE = 0.356 cm, R² = 0.978). HuashuTrees supports algorithm development for leaf-wood separation, segmentation, species classification, structure inversion, 3D reconstruction, and multi-temporal monitoring.