Abstract
Abstract
Non-destructive fruit weight estimation can provide low-cost visual support for automated grading and target-weight packaging, but small-sample models are prone to systematic errors when illumination, scale, background, and imaging batch change. This study proposes a four-view strawberry weight estimation framework for acquisition condition shift. The framework first uses SAM2 to extract the fruit region and construct morphological, color, and cross-view digital phenotypes, while a frozen DINOv2-small encoder is used to obtain visual semantic representations. Subsequently, the Biology- and Reliability-Guided Mixture of Experts (BRG-MoE) is used to improve in-distribution performance on the development set, whereas the Guarded Anchor-Residual TabPFN (GAR-TabPFN) uses an engineered-phenotype SVR as a stable anchor and suppresses residual extrapolation without training support through 5-nearest-neighbor novelty-aware gating. In 10 repeated nested five-fold validations on 174 strawberry samples, BRG-MoE achieved an MAE of 1.123±0.040 g and an R
2
of 0.855±0.007. On the independent fixed acquisition batch, the MAE of BRG-MoE increased to 2.758 g and R
2
decreased to -0.808; GAR-TabPFN classified all 38 samples as outside the support range and fell back to the phenotype anchor, reducing MAE to 0.969 g and increasing R$^2$ to 0.685. The results show that accuracy advantages within the development distribution cannot replace acquisition shift testing, and that auditable novelty-aware gating can provide an explicit failure-degradation path for small-sample visual measurement.