The accuracy of image-based individual tree crown detection and delineation across vegetation types

Abstract. Australia's terrestrial ecosystems are critical to the global carbon cycle, yet they face numerous environmental pressures such as forest degradation and biodiversity loss. Accurate monitoring of vegetation dynamics is crucial to mitigating these challenges and informing sustainable management strategies. Individual Tree Segmentation (ITS) methods, powered by deep learning, enable large-scale mapping of tree crowns, which is vital for assessing above-ground biomass and carbon stocks across vast landscapes. Despite their promise, inconsistencies in algorithmic performance arise due to varying vegetation types, point cloud densities, and dataset-specific characteristics, which limit the generalizability of supervised models.
This study evaluates the performance of different ITS and Canopy Height Model (CHM) algorithms for generating large tree crown datasets using LiDAR-derived data from across Australia. We applied these methods to 37 representative airborne LiDAR point clouds across 15 vegetation classes, representing a range of ecosystems from rangelands to tropical forests.
Our analysis reveals that the effectiveness of tree detection and crown delineation varies significantly across vegetation types and point cloud densities. The Pit-Free CHM algorithm generally outperforms others, yielding higher match rates in the delineation of tree crowns. Additionally, the DalPonte ITS algorithm provides the most accurate results, especially in sparsely vegetated areas such as rangelands, which are critical for mapping and monitoring. In contrast, closed-canopy forests present challenges, particularly due to crown clumping and multi-layered vegetation structures. This study highlights the importance of selecting the appropriate ITS and CHM methods for different vegetation types and emphasizes the need for algorithm optimization in complex environments, such as tropical and eucalypt forests. Ultimately, these findings provide valuable insights into enhancing large-scale vegetation monitoring and improving model generalization for tree crown detection.