UrbanTrunkSegNet: An Enhanced Deep Learning Approach to Individual Tree Trunk Segmentation for Urban Forest Inventory

Résumé fourni par la source

Urban tree trunk segmentation is a crucial part of accurate and scalable urban forest inventories. It allows for various downstream tasks such as tree health monitoring, trunk diameter estimates, and species identification. Segmenting trunks in street-level images presents challenges ranging from complex urban backgrounds to varying illumination conditions, occlusions, and structural diversity. To improve or mitigate these challenges for tree trunk segmentation, we developed UrbanTrunkSegNet, a hybrid deep learning model combining global feature modeling capabilities of a Vision Transformer (ViT), a portion of the U-Net++ decoder architecture to improve spatial detail, and leveraging multi-scale feature fusion and skip connections to improve boundary accuracy in complex urban settings. UrbanTrunkSegNet was evaluated using the Tree Binary Segmentation (TBS) and Urban Street Segmentation Trunk (USST) benchmark datasets. Across 8,007 images, UrbanTrunkSegNet achieved: Dice $=0.96$, IoU $=0.90$, and Accuracy $=0.95$ on TBS; and Dice $=0.90$, $\mathrm{IoU}=0.87$, and Accuracy $=0.89$ on USST. High segmentation performance and ability to generalize to complex urban scenes were confirmed with these results. Visual analysis confirmed high robustness against occlusions and background interference during segmentations and the ability to reliably segment multi-trunk structures with few false positives. These results confirm that UrbanTrunkSegNet is a reliable solution to automatically segment tree trunks within street-level images, with very good segmentation performance.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.