.png){kind=icon}
.png){kind=icon}
Return
A multi-task collaborative method for individual tree segmentation and multi-dimensional attribute extraction in mountainous areas using deep learning technologies
.png){kind=icon}
OA
AI
DOI：10.1016/j.jag.2026.105539.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
•
A novel multi-task collaborative framework is proposed for ITS and ITAE.
•
A two-stage segmentation method is developed to mitigate under-segmentation.
•
Transfer learning facilitates cross-task feature sharing and improves attribute accuracy.
•
Key attributes (position, height, crown width, and forest type) are extracted for each tree.
Abstract
Accurate mapping and monitoring of forest structure and composition in complex mountainous regions remain challenging because of crown occlusion, crown adhesion, and pronounced topographic variation. Moreover, current monitoring methods often focus on single-attribute retrieval and lack an integrated framework for collaborative multidimensional attribute extraction. To address these challenges, this study proposes a multi-task collaborative method for individual tree segmentation and integrated multidimensional attribute extraction in mountainous forests using deep learning and unmanned aerial vehicle (UAV) RGB imagery. A two-stage segmentation method combining a U-Net++ deep learning model with a marker-controlled watershed algorithm was developed, followed by a sequential transfer learning strategy to facilitate knowledge transfer among individual tree segmentation, tree height prediction, and forest type classification. The proposed method was evaluated over a 5 km × 5 km area within the Wanglang National Nature Reserve (WNNR) using manually delineated tree crowns and canopy height models (CHM) derived from UAV LiDAR data. The results showed that the two-stage segmentation method combined the semantic representation capability of deep learning with the geometric partitioning ability of the watershed algorithm, thereby reducing under-segmentation caused by crown occlusion and adhesion in mountainous forests. The omission rate decreased from 9.12 % to 4.67 %, while the F1-score increased from 0.9298 to 0.9537. Sequential transfer learning also improved individual-tree attribute prediction, with the RMSE of tree height prediction decreasing from 7.1504 m to 3.4825 m and the overall accuracy of forest type classification increasing from 81.68 % to 92.51 %. The proposed method enables integrated mapping of individual-tree location, crown width, tree height, and forest type in complex mountainous forests, providing detailed information for precision forest management and forest carbon assessment.
Keywords:
Individual Tree Segmentation
Tree Height Estimation
Forest Type Classification
Multi-Task Collaboration
Deep Learning
Journal
IF:
8.6
Papers:
5.1K
Citations:
2.4W