Few Labels, Shifting Sessions: Self-Supervised Features and Physical Priors for Plant Biomass from Multi-View RGB-D

Weighing a plant destroys it, so image-based biomass models are trained on few labels, and those labels arrive in capture sessions that differ in ways no plant-level feature sees. We study this regime on 38 potted Eucalyptus and Mango plants, each scanned from twelve RGB-D views and weighed, in four sessions that carry \( 67\% \) of the mass variance. We compare physical allometric priors, regressors on reconstruction descriptors, frozen supervised and self-supervised vision transformer tokens, and AB-SSViT, a head that anchors frozen patch tokens in 3D, pretrains on the training plants without labels, and learns a correction to a stem-diameter-and-height prior. Scored with each plant held out, frozen tokens reach \( R^2 \) \( 0.46 \) to \( 0.51 \), with the supervised ImageNet backbone ahead of DINOv2 and DINOv3; AB-SSViT reaches \( 0.48 \) with SAM masks and \( 0.66 \) with SAM 3 masks, ahead of the allometric prior \( (p=0.023) \) but not separable from a control that predicts each session's mean mass \( (R^2~0.58,~p=0.50) \) \( \). With a whole session held out every learned model falls below the mean, and the allometric prior degrades least. Weighing three plants of a new session and rescaling recovers \( R^2~0.73 \) for the prior and \( 0.56 \) for AB-SSViT. With few labels, the session, not the plant, is the unit of generalisation, and a handful of labels from the target session is worth more than a better representation.