Improving Chinese Fir Plantations DBH Inversion Accuracy Using Ensemble Learning Models Base on UAV-LiDAR
Diameter at breast height (DBH) is a fundamental measurement indicator in forest resource surveys. This study explores the use of uncrewed aerial vehicle light detection and ranging (UAV-LiDAR) for individual tree DBH inversion in Chinese fir (<italic>Cunninghamia lanceolata</italic>) pl...
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10964709/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Diameter at breast height (DBH) is a fundamental measurement indicator in forest resource surveys. This study explores the use of uncrewed aerial vehicle light detection and ranging (UAV-LiDAR) for individual tree DBH inversion in Chinese fir (<italic>Cunninghamia lanceolata</italic>) plantations with complex terrain and rich understory vegetation, and compares the results with those from Backpack-LiDAR. First, individual tree segmentation was performed on the UAV-LiDAR point cloud data from the study area, and individual tree point cloud feature variables were extracted by applying different height thresholds. Then, three types of models—statistical model multiple linear regression (MLR), traditional machine learning models including K-nearest neighbor regression and support vector regression, and ensemble learning models including random forest, extreme gradient boosting, and categorical boosting (CatBoost)—were employed for DBH inversion. The results show that: 1) Using data above a 5-m height threshold effectively reduces interference from understory vegetation; 2) Key feature variables, such as canopy volume (V), tree height (Hmax), the interquartile range of cumulative height percentiles (AIHiq), and canopy area (S), significantly affect DBH inversion, with V contributing 25% to the feature importance; 3) Ensemble learning models, particularly CatBoost, outperform the other models, achieving an <italic>R</italic><sup>2</sup> of 0.81% —14.1% higher than MLR; 4) DBH inversion closely matches field observed data, and UAV-LiDAR performs better than Backpack-LiDAR in complex forest environments. This study provides an efficient and cost-effective approach to forest resource monitoring. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |