Improving individual tree segmentation from aerial imagery using slicing-aided hyper inference and derivation of tree metrics

Trees are among the fundamental components of urban and forest ecosystems, contributing to carbon storage, microclimate regulation and the maintenance of ecological balance. Accurate extraction of individual trees in urban areas, forest stands and trees outside forests is therefore important for biomass estimation, carbon stock monitoring, forest inventory updating, and sustainable urban and land management. However, semantic segmentation of trees, instance-level delineation of individual trees and derivation of tree metrics from the resulting segments remain difficult to automate from aerial imagery because of overlapping crowns, variations in shadow and illumination, and background complexity. This thesis proposes an approach that combines Slicing Aided Hyper Inference (SAHI) with SAM 3, the latest version of the Segment Anything Model (SAM) family of vision foundation models. Semantic segmentation of trees was performed first, followed by instance-level segmentation of individual trees and tree canopies, and tree metrics were finally derived from the resulting segments. These stages were carried out in two separate configurations, with SAM 3 used alone and in combination with SAHI, so that the contribution of SAHI to segmentation performance could be assessed through direct comparison on the OAM-TCD dataset. The results were evaluated using precision, recall, F1 score and Intersection over Union (IoU). For semantic segmentation of tree cover, integrating SAHI sliced inference into SAM 3 increased IoU from 67.2% to 72.5%, the F1 score from 80.4% to 84.1% and recall from 74.6% to 86.5%. For individual tree crown detection outside canopy regions, recall increased from 23.4% to 63.2%, the F1 score from 36.0% to 52.8% and mAP50 from 20.84 to 47.90.