WoodBR-Cross: A public wood-image dataset and individual-tree-disjoint benchmark for species identification

Daniel Casanova; Alexandre Bahia Gontijo; Arnaldo Candido Junior; Silvana Nisgoski; Pedro Luiz de Paula Filho

cs.CVOct 8, 2026

WoodBR-Cross: A public wood-image dataset and individual-tree-disjoint benchmark for species identification

Authors: Daniel Casanova, Alexandre Bahia Gontijo, Arnaldo Candido Junior, Silvana Nisgoski, Pedro Luiz de Paula Filho

Organizations: Department of Computer Science, Federal University of Technology–Paraná (UTFPR), Medianeira, Paraná, Brazil · Forest Products Laboratory (Laboratório de Produtos Florestais), Brazilian Forest Service (Serviço Florestal Brasileiro), Brasília, Brazil · São Paulo State University (UNESP), São José do Rio Preto, São Paulo, Brazil · Department of Forest Engineering and Technology, Federal University of Paraná (UFPR), Paraná, Brazil

Abstract

Wood-image benchmarks can yield optimistic estimates when images from the same tree occur in both training and test sets. WoodBR-Cross contains 4,875 transverse-surface images in 38 operational taxonomic classes, linked to 303 distinct trees identified by numeric catalogue witness codes (codigos-testemunho) in the Brazilian Forest Service collection. A species-level subset comprises 3,406 images, 20 classes and 214 trees. Three tasks compare duplicate-aware image partitions for all classes (A), image partitions for the species subset (B1), and individual-tree-disjoint partitions for that subset (B2). In B2, all images from each tree are kept together, so no tree is shared across training, validation or test within a fold. ResNet-50, ConvNeXt-Tiny and ViT-B/16 were evaluated with three outer folds, internal validation and five training seeds. Handcrafted and frozen-feature baselines complement border-only and randomized-label controls. Validation-selected ConvNeXt-Tiny achieved B2 tree-level macro-F1 of 0.8621 +/- 0.0317; frozen DINOv2 features yielded 0.9000 +/- 0.0722. Values are means and standard deviations across three seed-averaged fold means. Results estimate performance on previously unseen trees from this collection.

Figures & tables

Task	Classes	Images	Trees	Partition unit	Macro-F1 unit
A	38	4,875	303	Image, with duplicate groups intact	Image
B1	20	3,406	214	Image, with duplicate groups intact	Image
B2	20	3,406	214	Tree ID	Tree

Table 1: Dataset subsets and evaluation units. Each tree is identified by one LPF catalogue witness code ( código testemunho ). B1 and B2 use the same 20 species-level classes.

Figure 1: One example per species-level class. Labels report image support, distinct-tree support and the example image identifier. Examples were selected by recorded quality and then the smallest image identifier. Files contain no EXIF/GPS information and were not retouched. No physical scale is inferred.

Task	Fold	Training	Validation	Test
A	1	2,600	649	1,626
A	2	2,600	649	1,626
A	3	2,601	651	1,623
B1	1	1,817	454	1,135
B1	2	1,817	454	1,135
B1	3	1,816	454	1,136

Table 2: Partition sizes. B2 counts distinct trees identified by catalogue witness code; tree IDs are disjoint across training, validation and test within each fold. Training and validation are combined for final fitting after model selection.

Task	Backbone	Macro-F1	Accuracy	Balanced accuracy
A	ResNet-50	0.9563 $\pm\,0.0016$	0.9561 $\pm\,0.0015$	0.9609 $\pm\,0.0013$
A	ConvNeXt-Tiny	0.9893 $\pm\,0.0023$	0.9914 $\pm\,0.0027$	0.9877 $\pm\,0.0023$
A	ViT-B/16*	0.9893 $\pm\,0.0019$	0.9897 $\pm\,0.0033$	0.9894 $\pm\,0.0005$
B1	ResNet-50	0.9730 $\pm\,0.0017$	0.9756 $\pm\,0.0014$	0.9753 $\pm\,0.0005$
B1	ConvNeXt-Tiny	0.9736 $\pm\,0.0226$	0.9800 $\pm\,0.0164$	0.9752 $\pm\,0.0212$
B1	ViT-B/16*	0.9837 $\pm\,0.0123$	0.9861 $\pm\,0.0087$	0.9842 $\pm\,0.0125$

Table 3: Outer-test means

\pm

SD across three seed-averaged fold means. In panel (a), B2 macro-F1 is tree-level, computed after averaging image probabilities within each catalogue witness code; all other metrics in panels (a) and (b) are image-level. Within each task and metric, bold marks all best displayed means (including ties) and underlining marks the next-highest distinct mean. Lower ECE and Brier values are better. Asterisks mark validation-selected backbones.

Figure 2: Primary macro-F1 across folds and training seeds. Dots are seed scores, squares are fold means, and diamonds with whiskers are the mean and SD across three folds. A/B1 evaluate images; B2 evaluates distinct trees (numeric catalogue witness codes in the figure). Asterisks mark validation-selected backbones.

Backbone	B1 image	B2 image	B2 tree-level
ResNet-50	0.9730 $\pm\,0.0017$	0.7749 $\pm\,0.0185$	0.8778 $\pm\,0.0321$
ConvNeXt-Tiny	0.9736 $\pm\,0.0226$	0.7979 $\pm\,0.0260$	0.8621 $\pm\,0.0317$
ViT-B/16	0.9837 $\pm\,0.0123$	0.7583 $\pm\,0.0293$	0.8231 $\pm\,0.0279$

Table 4: Macro-F1 for B1 and B2 by evaluation unit. Values are means

\pm

SD across three seed-averaged fold means. Bold marks the best displayed mean; underlining marks the next-highest distinct mean in each column.

Baselines
Method	A images	B1 images	B2 trees
HSV + LBP + GLCM	0.9221 $\pm\,0.0126$	0.9527 $\pm\,0.0046$	0.7024 $\pm\,0.0952$
Frozen DINOv2	0.9475 $\pm\,0.0154$	0.9719 $\pm\,0.0070$	0.9000 $\pm\,0.0722$
Majority class	$0.0035{\color[rgb]{0.55,0.55,0.55}\footnotesize\pm\,0.0000}$	$0.0092{\color[rgb]{0.55,0.55,0.55}\footnotesize\pm\,0.0000}$	$0.0069{\color[rgb]{0.55,0.55,0.55}\footnotesize\pm\,0.0014}$
Controls
Border-only inputs	$0.9804{\color[rgb]{0.55,0.55,0.55}\footnotesize\pm\,0.0093}$	$0.9726{\color[rgb]{0.55,0.55,0.55}\footnotesize\pm\,0.0181}$	$0.8428{\color[rgb]{0.55,0.55,0.55}\footnotesize\pm\,0.0798}$

Table 5: Baseline and control macro-F1. Values are means

\pm

SD of three fold scores. Among baselines, bold and underlining mark the best and second-best mean in each task; controls are unranked.

Figure 3: Image-level OOF confusion for ConvNeXt-Tiny in B2: mean of five seed matrices normalized by true class. Column numbers match the numbered row labels. Annotations are rounded percentages; cells below 1% are unannotated. Images come from held-out trees; predictions are not aggregated by tree in this matrix. Class support is provided in the supplementary class-count tables.

Figure 4: The 60 B2 learning-curve runs. Each $k$ uses five training subsets in each of three folds, with all images of each sampled tree retained. Black markers and whiskers show the mean and descriptive SD of 15 runs per $k$ , not confidence intervals. Architecture is selected separately within each fold.

Tree Species Classification

Tree species classification identifies tree taxa from images, spectral measurements, 3D structure, and satellite or aerial observations.

See all 18 papers

BarkNet-Lite: A Lightweight Texture and Colour Network with the BarkBD Benchmark for Bark-Based Tree Species Recognition in Bangladesh
Sep 7, 2026Aroshi Ali, Saad Ahmed, Md. Khalid SyfullahEfficient Neural Network Inference Tree Species Classification
Department of Computer Science and Engineering, Khulna University of Engineering & Technology, Khulna, 9203, Bangladesh · Department of Computer Science and Engineering, Bangladesh Army University of Science & Technology, Saidpur, Bangladesh
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
Three LiDARs, Twenty-Five Datasets: Cross-Source Generalization in Tree-Species Recognition
Oct 15, 2026Jae Joong Lee, Bedrich BenesTree Species Classification Cross-Dataset Generalization
Purdue University, Department of Computer Science 305 N University St, West Lafayette, IN 47907, USA

Sep 7, 2026J/K move · Enter open · S save

BarkNet-Lite: A Lightweight Texture and Colour Network with the BarkBD Benchmark for Bark-Based Tree Species Recognition in Bangladesh

Aroshi Ali, Saad Ahmed, Md. Khalid Syfullah

Department of Computer Science and Engineering, Khulna University of Engineering & Technology, Khulna, 9203, Bangladesh · Department of Computer Science and Engineering, Bangladesh Army University of Science & Technology, Saidpur, Bangladesh

Efficient Neural Network Inference Tree Species Classification Lightweight CNNs Image Classification

Open paper AskPDF arXiv

Tree species recognition supports forest inventory and biodiversity monitoring but still depends on scarce taxonomic expertise. Bark is visible year-round at ground level, yet bark recognition has concentrated on temperate floras and on large ImageNet-pre-trained backbones. We address both gaps. First, we release BarkBD, a bark dataset for Bangladesh: 14,258 uncropped smartphone photographs of 20 native species across four districts and three weather conditions, with a fixed stratified split. Second, we propose BarkNet-Lite, a 2.96M-parameter network trained from random initialisation, pairing a multi-scale texture pathway with a parallel colour-aware pathway. Over five seeds it reaches 96.64+-0.66%accuracyunderstrict single-image inference, within 2.3 points of nine ImageNet-pre-trained backbones fine-tuned under an identical protocol and within one seed-level standard deviation of the smallest ofthem, andtransfers to public benchmarks (95.86% on BarkVN-50, 92.85% on BarkNet 1.0). Grad-CAM, validated by faithfulness and weight-randomisation checks, confirms its decisions rest on bark structure rather than background. The exported single-precision model classifies one photograph in 15.34ms on a commodity smartphone.

Abstract

Tree Species Classification

Explore similar work

BarkNet-Lite: A Lightweight Texture and Colour Network with the BarkBD Benchmark for Bark-Based Tree Species Recognition in Bangladesh

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

Three LiDARs, Twenty-Five Datasets: Cross-Source Generalization in Tree-Species Recognition