StreetTree: A Large-Scale Global Benchmark for Fine-Grained Tree Species Classification

Jiapeng Li; Yingjing Huang; Fan Zhang; Yu liu

cs.CVFeb 22, 2026

StreetTree: A Large-Scale Global Benchmark for Fine-Grained Tree Species Classification

Authors: Jiapeng Li, Yingjing Huang, Fan Zhang, Yu liu

Organizations: Peking University · University of Vienna

Abstract

The fine grained classification of street trees is a crucial task for urban planning, streetscape management, and the assessment of urban ecosystem services. However, progress in this field has been hindered by the lack of large scale, geographically diverse, and publicly available benchmark datasets specifically designed for street trees. To address this critical gap, we introduce StreetTree, the world's first large scale benchmark dataset dedicated to fine grained street tree classification. The dataset contains over 12 million images covering more than 8,300 common street tree species, collected from urban streetscapes across 133 countries spanning five continents, and supplemented with expert verified observational data. StreetTree poses challenges for pretrained vision models under complex urban environments including high inter species visual similarity, long tailed natural distributions, significant intra class variations caused by seasonal changes, and diverse imaging conditions such as lighting, occlusions from buildings, and varying camera angles. In addition, we provide a hierarchical taxonomy (order, family, genus, and species) to support research in hierarchical classification and representation learning. Through extensive experiments with various vision models, we establish solid baselines and reveal the limitations of existing methods in handling such real world complexities. We believe that StreetTree will serve as a key resource for driving new advancements at the intersection of computer vision and urban science.

Figures & tables

Figure 1: Overview of StreetTree dataset. The dataset comprises 12,235,152 street view images representing 3,365,485 individual urban trees across 133 countries and five continents, spanning 71 orders, 241 families, 1,747 genera, and 8,363 species. Each tree may correspond to one or multiple street view images. The map illustrates the global geographic distribution of tree samples, where darker shades indicate higher densities. Most individual trees are associated with long-term temporal records, capturing seasonal and interannual variations.

Dataset	Geographic Scope	Size	Taxonomic Granularity	Temporal Span	Individual Tree	Seasonal Coverage	Image Type	Region Type
GlobalGeoTree [ 28 ]	Global	6,263,345	Familiy / Genus / Species	2015-2024	Yes	No	Satellite Image	Forest
Auto Arborist [ 4 ]	USA	2,637,208	Genus	2018-2022	Yes	No	Satellite Image and Street View Image	Urban
U.S. 5 Million [ 27 ]	USA	5,660,237	Genus / Species	-	Yes	No	Not available	Forest / Urban
Urbanforest [ 7 ]	Australia	76,928	Family / Genus / Species	-	Yes	No	Not available	Urban
EU Forest [ 26 ]	Europe	588,983	Genus / Species	-	No	No	Satellite Image	Forest
PureForest [ 11 ]	France	135,569	Species	-	No	No	Satellite Image	Forest

Table 1: Overview of publicly available datasets for tree species classification.

Figure 2: Examples of the binary classification of street view images. Images with less than 20% “tree” class pixels are labeled as negative samples, whereas those exceeding 20% are labeled as positive samples. Segmentation masks of “tree” class are highlighted with green color.

Figure 3: Statistics of StreetTree dataset. (a)–(d) show distributions across the four taxonomic levels: Order, Family, Genus, and Species. The horizontal axis uses a logarithmic scale to represent the magnitude of occurrence counts; (e) presents the seasonal distribution; (f) illustrates the temporal span distribution of individual trees.

Model	Frequent				Common				Rare
Model	Order	Family	Genus	Species	Order	Family	Genus	Species	Order	Family	Genus	Species
0.1% of training set
Fine-tuned ViT	17.52 (52.31)	13.26 (34.56)	6.33 (20.16)	1.81 (5.76)	12.34 (38.11)	9.47 (24.22)	4.16 (12.35)	1.43 (3.28)	7.82 (25.33)	5.14 (15.29)	2.13 (7.41)	1.16 (1.82)
CLIP	20.35 (60.62)	16.58 (41.25)	5.21 (21.28)	1.54 (5.56)	14.28 (45.18)	11.31 (30.12)	3.14 (14.22)	1.27 (3.18)	9.17 (30.16)	7.18 (18.11)	1.56 (8.12)	1.08 (1.52)
Fine-tuned CLIP	20.06 (58.76)	15.16 (39.84)	6.08 (22.95)	1.76 (5.82)	13.82 (43.21)	10.43 (28.17)	4.07 (15.11)	1.41 (3.59)	8.87 (28.19)	6.42 (16.21)	2.18 (9.14)	1.19 (1.97)
Fine-tuned SigLIP	23.41 (63.26)	18.28 (44.31)	8.42 (25.18)	1.52 (7.82)	16.56 (48.11)	12.87 (33.22)	5.86 (18.81)	1.93 (5.16)	11.18 (33.65)	8.41 (20.17)	3.27 (11.18)	1.46 (2.81)

Table 2: Model evaluation on StreetTree-12M. Results are presented as mean accuracy (%) over 5 random runs.

Model	Top-1				Top-5
Model	Order	Family	Genus	Species	Order	Family	Genus	Species
0.1% of training set
Fine-tuned ViT	16.37 $\pm$ 0.58	12.39 $\pm$ 0.30	5.84 $\pm$ 0.88	1.73 $\pm$ 0.22	49.15 $\pm$ 1.24	32.27 $\pm$ 0.22	18.49 $\pm$ 0.64	5.23 $\pm$ 0.55
CLIP	19.01 $\pm$ 0.28	15.43 $\pm$ 1.12	4.76 $\pm$ 1.28	1.48 $\pm$ 0.56	57.15 $\pm$ 0.70	38.71 $\pm$ 1.03	19.72 $\pm$ 1.46	5.05 $\pm$ 0.89
Fine-tuned CLIP	18.69 $\pm$ 0.44	14.12 $\pm$ 1.37	5.63 $\pm$ 0.19	1.69 $\pm$ 0.42	55.27 $\pm$ 0.66	37.20 $\pm$ 2.34	21.24 $\pm$ 0.87	5.34 $\pm$ 0.54
Fine-tuned SigLIP	21.91 $\pm$ 1.55	17.09 $\pm$ 0.18	7.84 $\pm$ 1.42	1.58 $\pm$ 0.67	59.86 $\pm$ 0.58	41.74 $\pm$ 1.46	23.70 $\pm$ 1.12	7.23 $\pm$ 0.69

Table 3: Overall model evaluation. Results represent the mean accuracy (%)

\pm

standard deviation (%) across all frequency categories over 5 random runs. The experiments using 100% of training set do not include a standard deviation.

Figure 4: Long-tail distribution across genus and species level in StreetTree dataset.

Figure 5: Example of occlusions caused by street infrastructures.

Figure 6: Example of excessively strong or weak lighting conditions. (a) and (b) illustrate cases of excessively strong lighting, while (c) and (d) show cases of insufficient illumination.

Figure 7: Example of intra-species variability driven by seasonal changes. (a)-(d) and (e)-(h) illustrate the seasonal progression observed in two individual trees, illustrating their seasonal variability. (a) corresponds to early spring, where the tree has recently emerged from dormancy and the branches remain bare. (b) and (d) show the tree in mid to late spring, characterized by dense and deep green leaves. (c) appears to depict early winter or a late autumn state where most leaves have fallen, leaving only a few yellow leaves clinging to the branches. For another tree, a similar pattern is posed. (e) and (h) were captured in summer, showing full, vibrant green leaves. (f) represents winter, during which the branches are completely bare. (g) reflects an autumn condition, with abundant yellow leaves signaling seasonal transition.

Figure 8: Example of tree misalignment arising from imprecise camera angles. (a)-(d) illustrate how inaccurate camera angle cause the target tree to appear too low, too high, too left, or too right in the captured image respectively.

Figure 9: Example of visually ambiguous boundaries among highly similar species. The species name of each tree is shown below corresponding image.

Order	Family	Genus	Species	Sample count
Frequent species
Pinales	Pinaceae	Picea	Picea pungens	500
Rosales	Rosaceae	Malus	Malus sylvestris	500
Laurales	Lauraceae	Machilus	Machilus japonica	500
Fagales	Fagaceae	Quercus	Quercus ilicifolia	500
Myrtales	Myrtaceae	Melaleuca	Melaleuca linariifolia	500

Table 4: Taxonomic composition of StreetTree-18kEval dataset.

Explore similar work

Plant Taxonomy Meets Plant Counting: A Fine-Grained, Taxonomic Dataset for Counting Hundreds of Plant Species
Mar 22, 2026Jinyu Xu, Tianqi Hu, Xiaonan Hu +4Object Counting
Huazhong University of Science and Technology, China
SilvaScenes: Tree Detection and Species Classification from Under-Canopy Images in Natural Forests
Oct 10, 2025David-Alexandre Duclos, William Guimont-Martin, Gabriel Jeanson +7Instance Segmentation Tree Species Classification
Northern Robotics Laboratory, Universit´e Laval, Qu´ebec, QC, G1V 0A6, Canada · D´epartement des sciences du bois et de la forˆet, Universit´e Laval, Qu´ebec, QC, G1V 0A6, Canada · Institut des Sciences de la Forˆet temp´er´ee, Universit´e du Qu´ebec en Outaouais, Ripon, QC, J0V 1V0, Canada
LeafInst - Unified Instance Segmentation Network for Fine-Grained Forestry Leaf Phenotype Analysis: A New UAV based Benchmark
Mar 4, 2026Taige Luo, Junru Xie, Chenyang Fan +5Instance Segmentation Plant Phenotyping
College of Information Science and Technology & Artificial Intelligence, Nanjing Forestry University, Nanjing, 210037, Jiangsu, China · Department of Geography, College of Natural Resources and Environment, Virginia Tech, Blacksburg, VA, 24061, USA · School of Software, Hefei University of Technology, Hefei, 230601, Anhui, China +2

Mar 22, 2026J/K move · Enter open · S save

Plant Taxonomy Meets Plant Counting: A Fine-Grained, Taxonomic Dataset for Counting Hundreds of Plant Species

Jinyu Xu, Tianqi Hu, Xiaonan Hu, Letian Zhou+3

Huazhong University of Science and Technology, China

Object Counting

Open paper AskPDF arXiv

Visually cataloging and quantifying the natural world requires pushing the boundaries of both detailed visual classification and counting at scale. Despite significant progress, particularly in crowd and traffic analysis, the fine-grained, taxonomy-aware plant counting remains underexplored in vision. In contrast to crowds, plants exhibit nonrigid morphologies and physical appearance variations across growth stages and environments. To fill this gap, we present TPC-268, the first plant counting benchmark incorporating plant taxonomy. Our dataset couples instance-level point annotations with Linnaean labels (kingdom -> species) and organ categories, enabling hierarchical reasoning and species-aware evaluation. The dataset features 10,000 images with 678,050 point annotations, includes 268 countable plant categories over 242 plant species in Plantae and Fungi, and spans observation scales from canopy-level remote sensing imagery to tissue-level microscopy. We follow the problem setting of class-agnostic counting (CAC), provide taxonomy-consistent, scale-aware data splits, and benchmark state-of-the-art regression- and detection-based CAC approaches. By capturing the biodiversity, hierarchical structure, and multi-scale nature of botanical and mycological taxa, TPC-268 provides a biologically grounded testbed to advance fine-grained class-agnostic counting. Dataset and code are available at https://github.com/tiny-smart/TPC-268.

Abstract

Explore similar work

Plant Taxonomy Meets Plant Counting: A Fine-Grained, Taxonomic Dataset for Counting Hundreds of Plant Species

SilvaScenes: Tree Detection and Species Classification from Under-Canopy Images in Natural Forests

LeafInst - Unified Instance Segmentation Network for Fine-Grained Forestry Leaf Phenotype Analysis: A New UAV based Benchmark