From 2D labels to 3D structure: Scalable label transfer and benchmarking of 3D vegetation models in rangeland ecosystems

Laura N. Sotomayor , Arko Lucieers, Darren Turner, Johan Barthelemy, Teja Kattenborn

Three-dimensional (3D) characterisation of vegetation structure at the level of individual growth forms is critical for understanding ecosystem function and resilience, yet remains challenging in rangelands because vegetation is sparse, low-stature, and structurally heterogeneous. Recent 3D deep-learning models perform strongly in forests, but their transfer beyond closed-canopy benchmarks is limited by scarce labelled 3D data and forest-centred architectural assumptions.
We present a modular 2D-to-3D label-transfer workflow that projects high-resolution UAS multispectral labels into co-registered LiDAR point clouds, generating height-plausible and object consistent volumetric supervision without manual point-cloud annotation. Using this supervision, we evaluated zero-shot transfer, few-shot fine-tuning, and training from scratch.
Forest-trained models declined from 85.2% mIoU in forest to 41.1% and 16.1% in medium- and low-density rangelands, while instance F1 dropped
sharply. Curriculum fine-tuning produced the strongest rangeland performance, reaching 49.6% and 63.1% mIoU. Effective 3D rangeland mapping therefore requires scalable target-domain supervision and adaptation beyond forest-centred structural assumptions.