Mapping private urban trees in Marseille with LiDAR and machine learning

Overview & Synopsis

Urban trees provide essential ecosystem services, particularly in Mediterranean cities facing recurrent heatwaves, drought, and pollution. While public inventories are increasingly available, vegetation on parcels outside the public right-of-way remains largely undocumented, which hinders city-wide estimates of canopy, cooling capacity, and carbon storage and biases actions toward streets and squares only. To address this evidence gap in Marseille, we assemble a multi-source data fusion that combines airborne LiDAR (IGN), terrestrial mobile LiDAR (Greehill), and IGN BD ORTHO ® RGB/IRC orthophotography with machine learning to map dominant genera and canopy structure across both municipal and parcel contexts. Airborne and terrestrial LiDAR show high consistency for canopy height (ordinary least squares, R 2 = 0.78 on n = 60,583 pairs), and a Random Forest baseline reaches ≈80 % overall accuracy for five common genera (Pinus, Platanus, Celtis, Tilia and Cupressus); we also report complementary agreement metrics (macro-averaged F1, Jaccard/IoU, and Cohen's κ) to ensure robust comparison. Building on these metrics, we quantify and compare structural and taxonomic differences between vegetation managed in the public domain and vegetation on parcels,