Assessment of tree detection and segmentation pipelines for terrestrial laser scanning dataset of orange orchards

Abstract

Abstract. This paper discusses the challenges faced by current tree segmentation pipelines in accurately detecting and performing coarse-to-fine segmentation of individual trees from terrestrial laser scanning (TLS) point clouds acquired in fruit-bearing crops such as orange orchards. Most pipelines for tree detection and individual tree segmentation were originally developed for forest environments, particularly boreal and temperate forests. Consequently, tropical forests and trees with more complex structures pose a challenge. For instance, orange, coffee, and lime crops present dense and overlapping canopies, which differ from those of boreal and managed forests. Our discussion is supported by a study aiming to detect and segment trees in an orange orchard using raster-based, point cloud-based, and hybrid algorithms. The results highlight the advantages and disadvantages in performance across the pipelines. Although stem detection is generally a more stable and accurate method for identifying tree positions, we concluded that approaches based on canopy height models (CHM) for tree detection and raster-based segmentation tend to provide more comprehensive results for orange crop trees. These methods offer better performance in cases where the canopy structure is complex, compared to those that rely on stem detection and clustering segmentation techniques.