Individual tree extraction through 3D promptable segmentation networks
Abstract The extraction of individual trees from three‐dimensional (3D) forest point clouds plays a pivotal role in forest inventory updates, forest resource management, growth and yield estimation of trees, etc. Existing end‐to‐end deep learning‐based methods for extracting individual trees typical...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wiley
2025-08-01
|
| Series: | Methods in Ecology and Evolution |
| Subjects: | |
| Online Access: | https://doi.org/10.1111/2041-210X.70057 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract The extraction of individual trees from three‐dimensional (3D) forest point clouds plays a pivotal role in forest inventory updates, forest resource management, growth and yield estimation of trees, etc. Existing end‐to‐end deep learning‐based methods for extracting individual trees typically rely on extracting instance‐sensitive features and clustering techniques. In this paper, inspired by the Segment Anything Model (SAM) and prompt‐driven paradigm, we propose a novel approach to forest point cloud instance segmentation, called the 3D Promptable Segmentation Network (3DPS‐Net). This network generates object masks using prompt points, thereby enabling the extraction of individual trees. We have designed two testing modes: prompt testing and automated testing. Prompt testing allows for interactive manual tree segmentation, providing a flexible and controllable analysis tool, while automated testing supports the autonomous identification and segmentation of individual trees within the input dataset. Through accuracy evaluation experiments conducted on two forest point cloud datasets (the ForestSemantic dataset and the FOR‐instance dataset), we have demonstrated the integrity and reliability of the individual tree instances predicted by 3DPS‐Net. Our proposed network exhibits robust performance with overstory trees. It requires only 1.5 seconds to predict a 3D individual tree mask from a prompt point, indicating significant potential for real‐time processing applications. This innovative approach introduces a new methodology for individual tree extraction using deep learning, thereby contributing to the advancement and future progress of this field. |
|---|---|
| ISSN: | 2041-210X |