An Unsupervised Approach to Build a Training Dataset for Individual Tree Crown Delineation Using Airborne Lidar and Field Observations

Most studies exploring individual tree crown (ITC) delineation in remote sensing manually label ITCs to train machine learning algorithms, which is a subjective and time-consuming task. In this study, we tested an approach to tune the segmentation parameters for tree counting and ITC delineation in different forest types (boreal, temperate and tropical) based on overlapping field data and canopy height models (CHM) extracted from LiDAR data. This was guided by the hypothesis that parameters for estimating crown size and crown structure vary by forest type, but generalize within forest types, meaning parameter estimation from scarce field data will enable scalable ITC generation from abundant LiDAR data. We parameterized an unsupervised segmentation model and compared the resulting number of crown segments with the observed number of trees within 576 field plots distributed in different forest types. The best results varied according to forest type where the sites are located. For sites in boreal areas, we reached a bias of -0.27 trees, while sites in temperate areas, reached a bias of -2.72 trees. The worst result was observed in tropical forest areas with 11.1 trees of bias. This approach can be used to choose the segmentation parameters to be applied to a much larger collection of sites with LiDAR but no field data available, enabling the creation of a large reference dataset of ITCs that can be used to train deep learning methods on optical data with supervised models.