A comprehensive benchmark of multi-generation YOLO architectures for forest species identification from macroscopic wood images.

Accurate identification of forest species from wood samples is essential for combating illegal logging and supporting sustainable wood trade. This study benchmarks 15 deep learning models from three YOLO generations (YOLOv8, YOLO11, and YOLO26), each in five sizes (nano through extra-large), to classify 73 forest species from macroscopic wood images. A dataset of 77,865 patches (448 × 448 pixels) was assembled from three heterogeneous image capture protocols with spatial resolution normalization. Four models achieved a test accuracy of 99.87% (MCC = 0.9987), with YOLO26-medium offering the best efficiency at 232.5 frames per second. All 15 models reached 100% Top-3 accuracy, and error analysis revealed that misclassifications were confined to individual patches. Image-level majority voting across patches yielded 100% accuracy for YOLO26-medium on test images. Multi-scale RISE saliency maps confirmed that the models attend to anatomically relevant features, such as pore arrangements and parenchyma patterns, while t-SNE analysis showed that YOLO26 produces the best-separated feature spaces. An out-of-distribution detection evaluation using three complementary methods (Maximum Softmax Probability, energy-based scoring, and Mahalanobis distance) confirmed strong model discriminability against both general out-of-domain content and unseen wood species acquired under the same protocols. These results establish a new benchmark for macroscopic wood species identification and suggest that medium-sized YOLO models are promising candidates for future deployment on resource-constrained devices for field inspection.