Source-linked AI summary
CedarCypress3D: an annotated UAV-LiDAR dataset of individual trees in planted cedar and cypress forests
Katsuto Shimizu, Fumiaki Kitahara, Tomohiro Nishizono, Hideki Saito, Masayoshi Takahashi, Shingo Obata, Shunsuke Tei, Naoyuki Furuya, Tomoya Goto, Eiji Kodani, Yusuke Yamada
TL;DR
Annotated UAV-LiDAR datasets for individual-tree analysis remain limited in temperate forests, despite their value for forest measurement and machine-learning development. CedarCypress3D addresses this gap with manually annotated UAV-LiDAR data, field measurements, and selected terrestrial LiDAR data, achieving close agreement between UAV-LiDAR-derived and field-survey tree heights.
Problem
Publicly available annotated UAV-LiDAR datasets have limited geographic and forest-type coverage, while individual-tree measurements support forest management and machine-learning development.
Method
The study constructs a manually annotated dataset from Japanese cedar and cypress plantations, combining UAV-LiDAR point clouds, field surveys, selected terrestrial LiDAR, and tree-level instance and semantic labels.
Results
UAV-LiDAR tree-height RMSE was 1.24 m for Japanese cedar and 0.96 m for Japanese cypress, with all field-surveyed trees matched to annotated point clouds.
Takeaways & Limitations
CedarCypress3D supports development and validation of individual-tree instance and semantic segmentation, tree-attribute prediction, and multi-platform LiDAR analysis.
Takeaways & Limitations
Terrestrial LiDAR-derived tree heights should not be used as a reference source because the sensor is less suitable for measuring tree height.
Abstract
from arXiv · showhide
Individual tree measurements derived from Light Detection and Ranging (LiDAR) mounted on Unmanned Aerial Vehicles (UAV) provide valuable information for forest inventory, ecosystem monitoring, and sustainable forest management. Recent advancements in machine learning have increased the demand for annotated datasets to develop and evaluate point cloud-based approaches, especially for individual tree segmentation. However, publicly available annotated UAV-LiDAR datasets in temperate forests are limited. In this article, we present CedarCypress3D, a manually annotated UAV-LiDAR dataset collected in Japanese cedar (Cryptomeria japonica) and Japanese cypress (Chamaecyparis obtusa) plantations in Japan. The dataset consists of UAV-LiDAR point clouds and field survey measurements from 34 circular plots across two sites with different topographic characteristics, along with terrestrial LiDAR point clouds available for a subset of 22 plots. A total of 1,627 trees were measured in the census field survey and manually annotated to match the corresponding trees in the UAV-LiDAR point clouds. For the subset of plots with terrestrial LiDAR data, semantic labels (i.e., stem and non-stem) were additionally assigned to tree points in the UAV-LiDAR data. CedarCypress3D provides high-quality annotated UAV-LiDAR data for developing and evaluating individual tree instance segmentation and semantic segmentation methods in temperate planted forests. The dataset can also support research on tree attribute prediction and multi-platform LiDAR analysis. The dataset is publicly available at https://doi.org/10.5281/zenodo.22168721.
1 Introduction
CedarCypress3D addresses limited annotated UAV-LiDAR coverage in temperate planted forests by providing manually labeled Japanese cedar and cypress data for individual-tree analysis.
- Motivation: UAV-LiDAR provides dense three-dimensional forest data for individual-tree detection, segmentation, and attribute estimation.Its applications include predicting DBH, crown base height, and crown structural metrics.
- Research gap: Publicly available annotated UAV-LiDAR datasets remain limited in geographic and forest-type coverage.Existing datasets include tropical and temperate forests, but relatively few cover the needed range of settings.
- Contribution: CedarCypress3D contains manually annotated UAV-LiDAR and terrestrial LiDAR point clouds with corresponding census field measurements.Individual-tree instance labels are provided, while stem and non-stem semantic labels are available for selected UAV-LiDAR plots.
- Contribution: The dataset supports development and evaluation of individual-tree instance segmentation, semantic segmentation, detection, and tree-attribute prediction algorithms.It is designed as ready-to-use data for point-cloud-based individual-tree analysis.
2 Materials and Methods
The dataset was collected across 34 circular plots at two Japanese plantation sites using field surveys, UAV-LiDAR, and terrestrial LiDAR, followed by co-registration and manual tree annotation.
- Study area: 34 circular plots were established across Saiki and Kokonoe, where Japanese cedar and Japanese cypress were the dominant planted conifers.The sites are in Oita prefecture, western Japan, with different stand-age ranges.
- Field survey: Field surveys recorded DBH, tree height, canopy base height, tree form class, species, tree locations, and plot-center coordinates.Circular plots had a radius of 12 m, and all planted trees were measured regardless of DBH.
- UAV-LiDAR measurements: Different UAV-LiDAR systems acquired the two site datasets, including an average density of 5,802 points/m2 across Saiki plots.The Saiki and Kokonoe campaigns used different sensors, UAV platforms, flight periods, and acquisition settings.
- Co-registration: Terrestrial LiDAR was collected for 22 of 34 plots using nine scans per plot, then co-registered with UAV-LiDAR data.Field tree locations were aligned with UAV-derived treetops before annotation, and terrestrial data were registered using a similar procedure.
- Manual annotation: UAV-LiDAR clouds were manually segmented into individual trees, reviewed by a third annotator, and linked to field-survey tree IDs by nearest-neighbor matching.Marker-controlled watershed segmentation assisted annotation at Kokonoe but was not applied at Saiki.
- Semantic labeling: Stem and non-stem semantic labels were manually assigned to tree points in the 22 plots with corresponding terrestrial LiDAR data.Labels covered 244 trees at Saiki and 624 trees at Kokonoe.
3 Dataset description
CedarCypress3D distributes LAS point clouds and CSV field data with tree-level identifiers, point classifications, and conditional stem/non-stem component labels.
- Data formats: UAV-LiDAR and terrestrial LiDAR point clouds are provided in LAS format, while field survey measurements are provided in CSV format.The terrestrial point clouds are available only for a subset of plots.
- Identifiers: The treeID attribute links annotated point-cloud trees to corresponding field-survey records.Non-tree points and trees outside the plot are assigned treeID = 0.
- Point classification: The Classification attribute distinguishes outside-plot, ground, and other point categories used in the dataset.Outside-plot points include buffer points and crowns from trees rooted outside the plot boundary.
- Tree components: The treeComponent attribute encodes unlabeled, stem, and non-stem components as 0, 1, and 2, respectively.Stem and non-stem labels are available only for tree points in plots with corresponding terrestrial LiDAR measurements; other points remain 0.
4 Dataset validation
Validation showed good agreement between field-survey measurements and annotated UAV-LiDAR data, while terrestrial LiDAR was less reliable for tree-height measurement. Remaining uncertainties involve overlapping crowns, missing field records, attribute calculations, and terrestrial-LiDAR limitations.
- UAV-LiDAR validation: 1.24 m and 0.96 m RMSEs were obtained for UAV-LiDAR tree heights in Japanese cedar and Japanese cypress, respectively, after excluding dead trees.All field-surveyed trees matched annotated point clouds, with a mean location difference of 0.51 m.
- Terrestrial-LiDAR validation: -2.94 m bias and 3.85 m RMSE were observed for terrestrial-LiDAR tree heights, attributed to canopy occlusion and the sensor characteristics.Figure 8 provides visual comparisons of normalized UAV-LiDAR and terrestrial-LiDAR point clouds by height bin.
- Annotation uncertainty: Overlapping branches and leaves, especially in Japanese cypress, complicate individual-tree separation, while suppressed trees may be partly or completely occluded.Stem-versus-non-stem semantic labels may also contain uncertainty because separating stems from branches and leaves is difficult.
- Field-survey correspondence: Some trees in Saiki plot 13 lacked corresponding field-survey records, so new tree IDs were assigned and their height and DBH were stored separately.The discrepancies were likely caused by omissions during the field survey.
- Field-attribute uncertainty: Field-survey stem volumes may be uncertain because they were calculated with species-specific equations, while stand ages came from potentially mismatched management records.Reported RMSEs between harvested and calculated stem volumes for the two species ranged from 2.5–13.5%.
- Terrestrial-LiDAR scope: Terrestrial-LiDAR tree heights should not be used as a reference source, although the data can generally provide accurate DBH measurements.Noise points were observed in some plots and may influence analyses.
5 Dataset usage
CedarCypress3D is intended for individual-tree analysis using annotated UAV-LiDAR, field measurements, and selected terrestrial-LiDAR data. Its labels and paired measurements support segmentation, tree-attribute prediction, and cross-platform studies.
- CedarCypress3D supports development and validation of individual-tree instance-segmentation methods using ready-to-use manually annotated UAV-LiDAR data.
- Semantic stem and non-stem labels are available for 22 of 34 plots, supporting semantic-segmentation research.
- Field-survey reference measurements support models predicting DBH, tree height, and stem volume from individual-tree data.
- Paired UAV-LiDAR, terrestrial-LiDAR, and field-survey data facilitate comparative studies of sensing platforms for individual-tree analysis.