Benchmarking YOLO models for forest species identification from macroscopic wood images

A comparative study of 15 models from the YOLOv8, YOLO11, and YOLO26 families for automated classification of 73 wood species, with t-SNE separability analysis and out-of-distribution detection.

99.87%
Accuracy
YOLO26m (best model)
73
Species
in the unified dataset
15
Models
trained & compared
3
Datasets
heterogeneous conditions
The problem

Why automate identification?

Illegal logging

Enforcement relies on rapidly identifying seized wood species — a process that currently demands scarce specialists.

Shortage of anatomists

Few professionals master wood anatomy. Samples are sent via courier for remote analysis, causing delays and inaccuracies.

Field-ready solution

A model deployed on a smartphone allows environmental agents to identify species on-site, generating instant reports.

Approach

Datasets and models

Combined datasets

DS1
Species41Resolution3264×2448 pxDeviceSONY DSC-T20
DS2
Species40Resolution2080×1540 pxDeviceZeiss Discovery V12
DS3
Species44Resolution640×480 pxDeviceHAIZ USB microscope

15 models compared

YOLOv8
NSMLX
YOLO11
NSMLX
YOLO26
NSMLX

All trained with mosaic augmentation · 448×448 px · 70/15/15 split

Results

Model performance

YOLO26m achieved the highest overall accuracy on the unified 73-species dataset. Detailed separability and misclassification analysis is available in the t-SNE Viewer.

Best model

YOLO26m

YOLO26 family · medium variant

99.87%
Accuracy
73
Species
Explore model separability in the t-SNE Viewer
Publication

Paper

in preparation

Benchmarking YOLO architectures for forest species identification from macroscopic wood images

Jandrei Sartori Spancerski, Pedro Luiz de Paula Filho, Mauricio Kugler, Fabio Kurt Schneider
UTFPR · ESIGELEC / Univ Rouen
DOI (coming soon) PDF (coming soon)
Back to projects