From laboratory to field: Frozen foundation-model features toward robust plant disease recognition

Plant-disease classifiers trained on the laboratory PlantVillage benchmark exceed 99% in-domain accuracy yet collapse to ∼ 26–30% on the in-the-wild PlantDoc set, a ∼ 70 pp laboratory-to-field gap usually blamed on background bias, which background removal, augmentation and test-time adaptation close only marginally. We show the gap is largely a property of the representation : a frozen self-supervised DINOv2 backbone with a linear probe trained only on PlantVillage reaches 51.1 ± 0.7% on PlantDoc, + 2 3 . 4 pp over a fully trained EfficientNet-B0 and + 2 1 . 2 pp over MobileViT, using no field labels and no backbone training; the separation is statistically decisive (exact per-image McNemar 𝑝 < 1 0 − 7 on every seed; across seeds, Welch 𝑝 = 8 . 8 × 1 0 − 9 ). The gain is not a resolution artefact ( + 2 0 . 6 pp at matched 224 px). We explain it with an exact decomposition of full-label-space (full-27) accuracy into a taxonomic-retention term and a within-crop discrimination term: across architectures and on two field datasets, the foundation features win by keeping a leaf with its crop (retention 92% versus 21–60% for conventional backbones), not by discriminating diseases better. Adapting the backbone is harmful: full fine-tuning collapses cross-domain accuracy from 50.6% to 29.4%, and neither low-rank adaptation nor weight interpolation beats freezing. Finally, distilling the frozen teacher into a 4 M-parameter CNN yields an edge-sized model that runs in 33 ms per image on a laptop-class CPU. Across matched seeds, distillation improves PlantDoc transfer. On PP2021, TIFD improves full-27 accuracy over plain distillation, while no variant clearly improves over the highly variable undistilled baseline and disease discrimination does not improve. Absolute field accuracy ( ∼ 51%) nevertheless remains insufficient for deployment; the results should therefore be interpreted as progress toward robust field recognition rather than as a deployment-ready diagnostic.