WHU-STree: A Multi-modal Benchmark Dataset for Street Tree Inventory
ISPRS Journal of Photogrammetry and Remote Sensing, 2026
Paper
The study presents WHU-STree, a pioneering multi-modal dataset designed to address critical gaps in urban street tree inventory research. By integrating synchronized point clouds and high-resolution images collected across two geographically distinct cities, WHU-STree offers a comprehensive resource for advancing automated urban forestry management. The dataset encompasses 21,007 annotated tree instances with species labels for 50 tree species and morphological parameters, enabling simultaneous support for tasks such as individual tree segmentation, species classification, and 3D morphological analysis. Its cross-city collection strategy, combined with rich annotations and multi-modal integration, sets a new benchmark for evaluating algorithmic generalization and robustness in real-world urban environments.
- 2026-07-21: The WHU-STree-NJ dataset will be used for the Data Processing Competition of the 10th China LiDAR Conference. We have also added
corrected_hdi.zip, which contains manually correctedtraj.csvfiles based on visual inspection of the data alignment. The ZIP archive is included directly in this GitHub repository for convenient download. Each file is named[road_id]_[ply_id]_traj.csv. - 2026-03-04: The reference data and Shenyang subset have been supplemented in the download links, enabling access to the final complete version of the dataset.
- 2026-02-04: WHU-STree is accepted by ISPRS J!
Fill in this Data Request to obtain the dataset. We provide download links for both Baidu Netdisk and Google Drive.
As the download links have changed multiple times, please ensure that you have obtained the final complete dataset. We welcome any questions or suggestions and would be grateful for your valuable feedback and guidance.
The newly added corrected_hdi.zip can be directly downloaded from this repository. Furthermore, visual comparisons before and after trajectory correction are provided to demonstrate the improvement in point cloud and image alignment.
Using the WHU-STree-NJ, we organized an expanded competition for the 10th China LiDAR Conference.
Compared with the previous competition, the task has been extended from individual tree segmentation to joint individual tree segmentation and tree species classification, aiming to advance automated urban vegetation understanding based on mobile mapping data.
For detailed competition specifications, please refer to the official conference website.
Utilizing a portion of the dataset, we organized an individual tree segmentation competition for the 9th China LiDAR Conference.
For detailed competition specifications, please refer to the official conference website.
Since the competition has concluded, we will no longer provide the competition-version data.
Statistics of WHU-STree dataset.
The folder structure of our dataset:
WHU-STree-NJ/SY
├── 00/ # Road 00
│ ├── hdi/ # Processed POS Data
│ │ ├── 1/ # POS data for Trajectory 1
│ │ │ └── traj.csv
│ │ └── 2/ # POS data for Trajectory 2
│ │ └── traj.csv
│ ├── image/ # Panoramic Image Data
│ │ ├── 1/ # Images for Trajectory 1
│ │ │ ├── 00000000-01-20201218104358926.jpg
│ │ │ ├── 00000000-01-20201218104401126.jpg
│ │ │ └── ...
│ │ └── 2/ # Images for Trajectory 2
│ │ ├── 00000000-01-20201218104402944.jpg
│ │ ├── 00000000-01-20201218104404539.jpg
│ │ └── ...
│ └── PCD/ # Point Cloud Data
│ ├── 1.ply # Point cloud for Trajectory 1
│ └── 2.ply # Point cloud for Trajectory 2
├── 01/ # Road 01
│ └── ... # Identical substructure
├── ... # Other roads
├── reference_data/ # Reference data for test dataset
│ ├── 00_1.npy
│ └── ...
└── test_split.txt # Road and point-cloud sequence IDs of the test set
Specifically,
Each traj.csv records the exterior orientation parameters for every panoramic image at its exposure moment along the same trajectory.
Per-record structure:
[image_filename, X, Y, Z, roll, pitch, heading] (One record per image)
Functionality:
These parameters enable coarse registration between point clouds and images.
Each .ply point cloud file contains per-point attributes:
[x, y, z, intensity, tree, label]
Among them:
tree: Instance labellabel: Tree species label
Details of tree species annotations are shown in the following table. It should be noted that, during the benchmark experiment, we have grouped species less than 150 instances into "others" as shown in the table. However, we consider the original annotations of these "others" species to remain valuable, and therefore their labeling information has been preserved in the "label" field, specifically for those with label IDs greater than 17.
Researchers can use the provided ground truth data to evaluate algorithm performance. The files are in int16 .npy format and are named as [road_id]_[ply_id].npy.
We have established benchmarks for two core tasks: tree species classification and individual tree segmentation. Additionally, we have introduced a species-specific benchmark for individual tree segmentation. Detailed information can be found in our paper.
If you find our work useful in your research, please consider citing:
@article{ding2026whu,
title={WHU-STree: A multi-modal benchmark dataset for street tree inventory},
author={Ding, Ruifei and Chen, Zhe and Fan, Wen and Long, Chen and Xiao, Huijuan and Zeng, Yelu and Dong, Zhen and Yang, Bisheng},
journal={ISPRS Journal of Photogrammetry and Remote Sensing},
volume={233},
pages={519--542},
year={2026},
publisher={Elsevier}
}