Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Online Access: | https://doi.org/10.5281/zenodo.18113771 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <p><span lang="EN-US">Canopy segmentation is a crucial step in obtaining canopy parameters through forestry remote sensing. </span><span lang="EN-US">However, tree canopy segmentation based on UAV imagery still suffers from issues such as adjacent canopy edges sticking together, insufficient edge detail, and background interference. To address the aforementioned issues, this researh conducted comparative experiments using four models—U-Net, PSPNet, HRNet, and DeepLab V3+—on a tree canopy segmentation dataset, verifying the superiority of the U-Net model in tree canopy segmentation tasks. Then we further propose a semantic segmentation network based on the U-Net model, termed BKFE-Unet. This network incorporates a Differential Border Attention Module and a Key Feature Enhancement Module. The Differential Border Attention Module enhances canopy edge information through differential computation, mitigating issues such as adjacent canopy edge adhesion and insufficient edge detail, and the Key Feature Enhancement Module integrates Ghost convolution, which stacks Ghost features to simultaneously achieve the filtering of redundant information and the enhancement of key semantic features in canopy objects, thereby reducing background interference. The proposed BKFE-Unet model demonstrates outstanding performance on the UAV canopy segmentation dataset, achieving mPA, Accuracy, mIoU, and F1 scores of 92.62%, 96.30%, 86.5%, and 0.922, respectively. These results represent significant improvements over the baseline U-Net model.</span></p>