Abstract
<title>Abstract</title> <p> Individual tree species mapping from LiDAR underpins forest inventory and biodiversity monitoring, yet remains difficult because supervised classifiers need expensive, geographically biased per-tree labels, and label-free geometric descriptors capture only predefined statistics. We propose a fully unsupervised, matching-free clustering framework on frozen off-the-shelf 2D vision foundation-model features: each tree is rendered into six pose-canonical depth views, encoded by a frozen image encoder, and clustered without species labels or text prompts. Across two independent LiDAR datasets, these features organize crowns by species more coherently than a hand-crafted geometric descriptor (matching-free Normalized Mutual Information (NMI) 0.28-0.35 versus 0.10) and than a frozen 3D point-cloud foundation model given the native geometry (OpenShape/PointBERT, 0.19-0.27), despite no 3D training. Because that 3D model is pre-trained on computer-aided-design (CAD) shapes, this is an availability comparison among frozen models, not an architectural verdict. The advantage holds for every pre-trained backbone, vanishes for a random-initialized control (isolating pre-training), and is stable across four clustering algorithms, fifteen seeds, and an embedding-dimension control. We further identify acquisition-source confounding as a first-class evaluation hazard: on the multi-sensor dataset most of the global gap recovers acquisition source, leaving a small, bootstrap-significant within-source residual of +0.044 NMI (foundation over geometry) on FOR-species20K, versus +0.204 on the single-sensor BioDiv-3DTrees-ULS. Although the uncontrolled gap is permutation-significant ( <italic>p</italic> < 1e-4), we anchor our evidence in this within-source bootstrap, treat the global gap as an upper bound, and scope our claims to feature coherence, not deployable classification. </p>