IC4SD-Wood-Eucalyptus: A macroscopic transverse-section image dataset with metadata, split manifests, and leakage-audit reports for Eucalyptus wood identification

IC4SD-Wood-Eucalyptus is a macroscopic transverse-section image dataset for fine-grained Eucalyptus wood identification. The dataset contains 2,910 images from 86 physical wood specimens, acquired at 50× magnification from eight classes: seven Eucalyptus species (E. camaldulensis, E. cladocalyx, E. deglupta, E. diversicolor, E. grandis, E. microcorys, and E. saligna) and one Myrtaceae outgroup, Syzy- gium hemisphericum. The release includes raw images, class labels, original image dimensions, SHA-256 file hashes, physical-specimen group identifiers, perceptual-hash component identifiers, two split man- ifests, and leakage-audit outputs. The two distributed split manifests comprise a specimen-stratified reference split and a strict pHash-clean group-disjoint split; both are constructed from group identifiers anchored to physical-specimen subfolders and are audited for perceptual near-duplicate leakage. The strict split contains 2,025 training images, 437 validation images, and 448 test images, distributed across 56, 15, and 15 physical-specimen groups respectively, with all eight classes represented in every partition. The accompanying leakage-audit files document parsed-group overlap, exact file-hash overlap, filename overlap, and perceptual near-duplicate checks at two Hamming-distance thresholds. The dataset is intended to support reproducible benchmarking of macroscopic wood-image classifiers and to provide an example of leakage-aware data organization for biological image classification.