Assessing optimization strategies for unsupervised individual tree crown detection and delineation to support continental-scale inventories: role of vegetation type and point cloud data density

To monitor forest degradation, deforestation, reforestation, and above-ground biomass, individually delineated tree crowns are key satellite-derived products. Unsupervised individual tree segmentation (ITS) algorithms generate crown delineations from Light Detection And Ranging (LiDAR)-derived point clouds or rasters representing canopies. Supervised deep learning models trained with these delineations can then be used to segment billions of crowns using high-resolution imagery at continental scales. However, diverse vegetation types, variable LiDAR point cloud densities, and algorithm misparameterization introduce errors, undermining the generalizability of supervised models trained on these outputs. The aim of this study is to identify the most suitable CHM generation and unsupervised canopy segmentation approach across diverse canopy height distributions and point cloud densities. To this end, we evaluate detection and delineation performances of (a) three common canopy height models (i.e., (i) point-to-raster; (ii) triangulated irregular network; and (iii) pit-free), (b) four ITS algorithms (i.e., (i) Watershed; (ii) DalPonte; (iii) Li; and (iv) Silva), across (c) 15 vegetation types using (d) three point-cloud density classes, comparing accuracy obtained using default versus heuristically optimized parameters. We found that the DalPonte algorithm achieved the highest delineation accuracy in 84% of cases, while the Watershed algorithm, despite notable optimization gains and its most frequent use in literature, ranked as the second least accurate method. Vegetation type substantially influenced accuracies, with semi-arid rangelands and woodlands performing best and dense, multi-layered forests the worst. High-density point clouds provided the best results, but heuristically optimized methods performed well with low-density datasets (<3 pts/m 2 ). Optimizations generally improved detection and delineation accuracies, with +51% match ratio, -25% oversegmentation, and negligible increases in undersegmented crowns. The largest gains occurred in low-density point clouds, with +92% detection accuracy and four times more accurate delineations. Finally, this research identified the most accurate optimized algorithms for specific vegetation types and point cloud densities, providing a foundation for large-scale crown delineations to train highly generalizable deep learning models. • 3 CHM and 4 ITS algorithms run across 39 diverse LiDAR point clouds • Accuracy evaluated for 15 vegetation and 3 LiDAR density classes • DalPonte leads in 84 % of cases for top crown delineation accuracy • Optimizations boost < 3 pts/m 2 point clouds: +92 % detection, 4× accurate delineations • Optimized image-based ITS enables cost-effective, diverse DL training data creation