WHU-STree: Urban Tree Dataset Benchmark
- WHU-STree is a multi-modal, cross-city urban street tree dataset coupling synchronized LiDAR and panoramic imagery to enable automated inventory.
- It supports diverse tasks including 3D instance segmentation, species classification, and morphological parameter estimation across varied urban climates.
- Its cross-city design, advanced preprocessing, and rich annotations enhance research on domain generalization and urban asset management.
Searching arXiv for WHU-STree and closely related street-tree dataset work to ground the article. WHU-STree is a multi-modal, cross-city benchmark dataset for automated street tree inventory from Mobile Mapping Systems (MMS). It couples dense, geo-referenced point clouds with high-resolution panoramic imagery over two climatically distinct Chinese cities and provides 21,007 annotated tree instances, 50 species, and two morphological parameters, namely height and DBH. The benchmark is designed to support more than ten tasks spanning perception, inventory, and analysis, including individual tree instance segmentation, species classification, morphological parameter estimation, multi-modal fusion, and cross-domain generalization (Ding et al., 16 Sep 2025).
1. Definition, scope, and naming
WHU-STree was introduced to address the limitations of existing MMS-acquired tree datasets, which were described as limited by small-scale scene, limited annotation, or single modality. In contrast, WHU-STree is explicitly framed as a cross-city, richly annotated, and multi-modal urban street tree dataset collected for detailed and dynamically updated street tree inventory in space-constrained urban environments (Ding et al., 16 Sep 2025).
The dataset is motivated by the practical role of street trees in urban livability. The underlying problem statement emphasizes that street trees provide cooling and shading, air purification, noise buffering, safer and more pleasant streets, and public health and equity gains. This motivates the need for a detailed, accurate, and frequently updated inventory of tree locations, species, and structure, which conventional field surveys cannot scale to because they are too slow and labor-intensive. MMS is positioned as the alternative acquisition paradigm because vehicles carrying synchronized LiDAR and cameras with precise GNSS/IMU positioning can provide close-range, high-throughput coverage of streetscapes while capturing the geometry, texture, and context needed for comprehensive tree inventories (Ding et al., 16 Sep 2025).
The name should be distinguished from unrelated uses of “STree” in other research areas. In the literature provided here, “STree” also denotes a speculative tree decoding method for state-space models (Wu et al., 20 May 2025) and an SVM-based oblique decision tree algorithm (Montañana et al., 2024). By contrast, WHU-STree refers specifically to the street tree inventory benchmark dataset (Ding et al., 16 Sep 2025).
2. Data acquisition, sensing configuration, and preprocessing
WHU-STree was collected in Nanjing and Shenyang, approximately 970 km apart and in distinct climate zones. The two subsets are named WHU-STree-NJ and WHU-STree-SY. This cross-city design intentionally induces variation in species composition, tree morphology, urban form, and acquisition conditions so that algorithms can be evaluated under nontrivial domain shift rather than within a single homogeneous acquisition regime (Ding et al., 16 Sep 2025).
Two MMS platforms with comparable sensor suites were used. Hiscan-Z was used in Nanjing and Alpha3D in Shenyang. Both systems are synchronized by a central controller and acquire dense 3D point clouds, 360° images, and high-precision positioning/orientation data. The panoramic images are time-stamped, and extrinsics at exposure time are provided so that 3D labels can be projected onto 2D imagery.
| Subset | Coverage | Platform |
|---|---|---|
| WHU-STree-NJ | ~59.182 km over 31 roads | Hiscan-Z |
| WHU-STree-SY | ~7.062 km over 8 roads | Alpha3D |
The Nanjing platform specification is LiDAR with max range 119 m, laser emission 1010 k pts/s, 200 Hz scanning, 9 mm ranging accuracy, and a 360° panoramic camera at px. The Shenyang platform specification is LiDAR with max range 420 m, laser emission 1000 k pts/s, 250 Hz scanning, 5 mm ranging accuracy, and a 360° panoramic camera with 8192 px width. The combined data volume is 108.2 GB, partitioned as 92.9 GB for Nanjing and 15.3 GB for Shenyang. Average point densities are reported as approximately 616 points/ in Nanjing and approximately 1377 points/ in Shenyang (Ding et al., 16 Sep 2025).
The preprocessing pipeline is specified in four stages: voxel downsampling at 0.05 m, adaptive segmentation at road intersections, statistical outlier removal, and ground removal using cloth simulation filtering. For downstream learning, point clouds are further partitioned into cylindrical tiles with radius 16 m. For the multi-modal LCPS baseline, panoramas are unwrapped into 120° FOV perspective views to reduce distortion, and co-located LiDAR is cropped to a m window to ensure coverage (Ding et al., 16 Sep 2025).
This design suggests that WHU-STree is not merely a perception benchmark in the narrow sense of segmentation accuracy; it is structured as an operational inventory dataset in which geo-referencing, point-to-pixel projection, and multi-view alignment are first-class components of the problem definition.
3. Annotation schema and dataset composition
WHU-STree contains 21,007 annotated tree instances across two cities, with 50 species identified in WHU-STree-NJ. The data product integrates synchronized LiDAR and panoramic images and provides two morphological parameters per tree. The dataset also includes over 12,000 panoramic images aligned with the 3D data, and road-segment partitions serve as scenes or tiles for training and evaluation (Ding et al., 16 Sep 2025).
The annotation protocol is instance-centric. Trained annotators delineated 3D instances, and overlapping crowns were separated using morphological cues to preserve reasonable 3D structure. Species identification in Nanjing integrated field surveys with evidence from LiDAR geometry and panoramic texture and color. Species abbreviations such as Platanus × acerifolia “PA,” Cinnamomum camphora “CC,” Zelkova serrata “ZS,” and Prunus serrulata “PS” are provided for consistency. The 2D–3D alignment is generated by projecting 3D instance annotations into 2D using camera extrinsics at exposure time, so no separate manual 2D annotation is required (Ding et al., 16 Sep 2025).
The two reported morphological parameters are defined explicitly. Height is computed as
where is the maximum of the tree points and is the local ground elevation under the tree. DBH is estimated at 1.2–1.4 m above ground. Trunk points are first extracted with a trunk classifier, specifically MinkNet trained on FOR-instance, then an ellipse is fitted to a horizontal slice, and
where and 0 are the fitted major and minor axes. Field DBH measurements are used where LiDAR points are insufficient. Tree position is defined as the trunk bottom center from the annotated point cloud (Ding et al., 16 Sep 2025).
The benchmark also documents substantial cross-city differences. Nanjing trees are generally larger, with height concentrated in 10–20 m and many exceeding 30 m, and DBH concentrated in 0.2–0.5 m. Shenyang trees cluster around height 5–15 m and DBH below 0.3 m. Species abundance is highly imbalanced spatially and across classes, and many species are rare with fewer than 150 instances. For baseline classification, classes with fewer than 150 instances are grouped into “Others,” explicitly acknowledging a long-tail regime (Ding et al., 16 Sep 2025).
No specific annotation tooling details, quality-control procedures, or inter-annotator agreement statistics are reported. That absence is itself relevant because it constrains how uncertainty in species labels, instance boundaries, and derived morphology should be interpreted in downstream benchmarking.
4. Supported tasks and evaluation protocol
WHU-STree is designed to support more than ten tasks. These include individual tree instance segmentation in 3D, image, and multi-modal settings; species-specific instance segmentation; species classification from 3D, 2D, or fused inputs; semantic segmentation of trees versus background; object detection; morphological parameter estimation including prediction of 3D parameters from 2D images alone; multi-modal fusion; cross-domain generalization and domain adaptation; spatial pattern and context learning; inventory mapping and update; multi-task learning; and foundations for Multi-modal LLM-based analytics for street tree asset management (Ding et al., 16 Sep 2025).
The split strategy is explicitly bifurcated into intra-city and cross-city protocols. The intra-city split is class-balanced: road segments are uniformly sampled so that train and test subsets maintain similar species ratios within Nanjing. The cross-city split trains on WHU-STree-NJ and tests on WHU-STree-SY, thereby measuring domain generalization under different climates, species composition, density regimes, and urban forms (Ding et al., 16 Sep 2025).
The benchmark formalizes several metrics. For species classification, it reports mean IoU and Overall Accuracy. For instance segmentation, it reports Detection, Omission, Commission, and F1-score at IoU threshold 0.5. For species-specific instance segmentation, it reports mean Precision, mean Recall, mean Coverage, mean Weighted Coverage, and F1-score. The paper also lists standard formulas for accuracy, cross-entropy loss, per-class IoU, mIoU, OA, precision, recall, F1, and the coverage variants, and notes that mAP and Dice are standard but were not the primary focus in the reporting (Ding et al., 16 Sep 2025).
A plausible implication is that WHU-STree is structured as a benchmark suite rather than a single-task corpus. The combination of aligned 3D geometry, panoramic imagery, species labels, and morphology permits evaluation under multiple supervisory granularities, ranging from instance-level reasoning to inventory-oriented analytics.
5. Baselines and quantitative findings
For species classification on WHU-STree-NJ, using classes with at least 150 instances plus “Others,” the reported baselines are MinkNet, PointMLP, PTv2, and TSCMDL. Each 3D instance is downsampled to 8192 points for 3D models, while TSCMDL associates tree instances to projected 2D bounding boxes and resizes image patches to 1. The reported results are: MinkNet, mIoU 49.94% and OA 81.67%; PointMLP, mIoU 58.00% and OA 84.51%; PTv2, mIoU 64.09% and OA 87.99%; and TSCMDL, mIoU 60.49% and OA 86.20%. PTv2 yields the best overall classification result, while TSCMDL improves over its 3D PointMLP counterpart by 2 mIoU, which the benchmark interprets as confirmation that texture and color are valuable for fine-grained species discrimination (Ding et al., 16 Sep 2025).
For individual tree segmentation, the evaluated architectures are SegmentAnyTree, SoftGroup, SPFormer, and LCPS. On WHU-STree-NJ, the species-agnostic results are: SegmentAnyTree with Detection 87.6%, Omission 12.4%, Commission 14.1%, and F1 86.8%; SoftGroup with 59.0%, 41.0%, 33.0%, and 62.7%; SPFormer with 93.2%, 6.8%, 26.6%, and 82.2%; and LCPS with 75.6%, 24.4%, 17.8%, and 78.8%. On WHU-STree-SY under cross-city testing, the results are: SegmentAnyTree with 86.0%, 14.0%, 12.6%, and 86.7%; SoftGroup with 79.9%, 20.1%, 20.1%, and 79.9%; SPFormer with 87.4%, 12.6%, 17.5%, and 84.9%; and LCPS with 83.0%, 17.0%, 7.8%, and 87.4%. The benchmark identifies SegmentAnyTree as best F1 in Nanjing and LCPS as best F1 in Shenyang (Ding et al., 16 Sep 2025).
For species-specific instance segmentation on Nanjing, the results are: SegmentAnyTree with mPrec 56.9%, mRec 46.5%, mCov 43.2%, mWCov 46.8%, and F1 51.2%; SoftGroup with 36.0%, 31.1%, 31.4%, 31.6%, and 33.4%; SPFormer with 52.8%, 63.9%, 59.4%, 61.8%, and F1 57.8%; and LCPS with 41.9%, 38.0%, 35.3%, 35.0%, and F1 39.9%. This confirms that integrating species classification within instance segmentation is materially harder than classifying pre-cropped individual trees (Ding et al., 16 Sep 2025).
Ablation results isolate two central effects. First, multi-modal fusion improves LCPS relative to its 3D-only version: on Nanjing species-agnostic segmentation, F1 rises from 74.3% to 78.8% and Commission drops by 7.5%; on Shenyang, F1 rises from 83.2% to 87.4%. In species-specific segmentation, the multi-modal version improves mPrec by 6.9%, mRec by 12.0%, mCov by 10.8%, mWCov by 8.5%, and F1 by 9.9%. Second, multi-task collaboration improves SPFormer: training it to predict species rather than only tree versus non-tree raises species-agnostic F1 on Nanjing from 69.2% to 82.2% and raises cross-city F1 from 77.5% to 84.9%, although Detection slightly drops in the cross-city setting (Ding et al., 16 Sep 2025).
6. Cross-domain behavior, spatial priors, limitations, and applications
One of the most distinctive empirical observations is that the cross-domain setting does not exhibit the typical out-of-distribution degradation. When models are trained on WHU-STree-NJ and tested on WHU-STree-SY, performance remains robust or even improves. The paper attributes this to four factors: the large and diverse Nanjing training data, voxel downsampling that reduces density gaps, the Shenyang scenes having less crown overlap and fewer shrub confusions, and the benefit of RGB fusion in countering point-cloud distribution shifts. The best cross-city F1, 87.4%, is achieved by the fusion model LCPS (Ding et al., 16 Sep 2025).
The benchmark explicitly argues that spatial pattern learning is an open direction rather than a resolved component. Street trees are described as being laid out under human design, with linear alignment along roads, quasi-regular spacing per local standards, and zone-specific species preferences. The paper states that learning this structure could improve 3D segmentation through priors on alignment to road centerlines, planting intervals, and topological relations to curbs, poles, and buildings, while also improving species classification through contextual priors such as homogeneous species per road or zoning preferences. This suggests that WHU-STree can support models that move beyond local appearance and local geometry toward structured urban-context reasoning (Ding et al., 16 Sep 2025).
Several limitations are stated directly. Geographic coverage is limited to two cities, species labels are currently comprehensive for Nanjing while Shenyang species labeling is planned, seasonal variability is not explicitly controlled or documented, occlusions and shrub confusion remain challenging, sensor asynchrony and multi-view temporal offsets can induce small but consequential projection and alignment errors for fusion, and long-tail species distribution together with inter-species similarity limit ceiling performance without specialized imbalance handling or open-vocabulary methods. Annotation tooling details, inter-annotator agreement, and license terms are not reported in the paper (Ding et al., 16 Sep 2025).
The practical application scope is broader than benchmark reporting. WHU-STree is intended for building and updating city-scale street tree inventories, supporting urban planning and ecological assessment, prioritizing maintenance and risk management, and training perception modules for municipal digital twins. The paper also identifies Multi-modal LLMs as a future direction for integrating perception, expert knowledge, and policy rules into end-to-end “perception–analysis–decision” asset management. This suggests that WHU-STree is positioned not only as a perception benchmark but also as a substrate for integrated urban forestry analytics (Ding et al., 16 Sep 2025).