Watershed vs. Region Growing for Individual Tree Segmentation from Airborne LiDAR: An Urban Case Study in Bologna

Abstract

We present a LiDAR-based pipeline for the segmentation and structural characterisation of tall urban vegetation in Bologna. Starting from airborne LiDAR point clouds previously classified as high vegetation by a Random Forest model, we implement and compare two individual-tree segmentation strategies: a watershed algorithm applied to the Canopy Height Model and a point-wise region growing algorithm operating directly on the three-dimensional cloud. Over an area of 12 tiles covering the Talea district, the watershed method detects a larger number of trees (7589), dominated by short trees with a height peak around 5 meters, whereas region growing yields fewer trees (6432) but retains the highest returns, yielding individuals above 40 meters, reflecting the absence of a smoothing step in the latter. We further confront the segmentation results with the municipal Open Data catalogue Alberi in manutenzione: on a sample tile roughly half of the LiDAR-detected trees turn out to be absent from the catalogue, several catalogued positions correspond to locations where no tree is observed, and the majority of the height records date back about two decades, which makes the dataset unsuitable as a ground truth reference. From the segmented trees we extract height and crown radius and use them to compute preliminary vegetation indicators, namely above-ground biomass and carbon storage through allometric relations and annual pollen production through species-specific inflorescence coefficients. All results are collected in an interactive per-tree metadata map. The pipeline is modular and can be recalibrated as improved classifications, field campaigns, or complementary sensing technologies become available, providing a scalable basis for data-driven urban greenery planning.