Urban trees play a vital role in sustainable urban planning, biodiversity, conservation and climate resilience. The expanding availability of Light Detection and Ranging (LiDAR) data offers new opportunities for three-dimensional analysis of urban vegetation. However, the development of deep learning methods for tree species classification remains constrained by the lack of annotated datasets at the individual-tree level. This study proposes a robust and reproducible workflow for generating annotated urban tree point cloud datasets by integrating high density airborne LiDAR data with field inventory information, followed by field validation to ensure reliability. The methodology includes preprocessing of both point cloud and field inventory data: noise and outlier removal, vegetation filtering, elevation normalizing using a Digital Terrain Model, and extraction of acquisition dates for temporal consistency. Field inventory records were cleaned by removing duplicates, harmonizing species names, and eliminating incomplete or spatially inconsistent entries. Individual trees were then segmented from the point clouds using a combined DBSCAN and Watershed approach, and matched to inventory records through a Nearest Neighbor method. A field verification confirmed that the retained trees had not changed between the inventory and study dates. The workflow was applied to create a dataset of 152 individual urban trees across six species. The results demonstrate its effectiveness for integrating multi-sources geospatial data and producing high-quality annotated datasets, which will be made publicly available to support open science, reproducibility, and future applications in urban tree species classification, smart cities and urban digital twins.