Self-supervised Vision Transformer Pretraining for Artificial Intelligence Weed Detection Under Herbicide-Resistant Biotype Drift: A Computer Robustness Evaluation

Weed detection models trained on one biotype assemblage often fail when herbicide-resistant biotypes shift leaf morphology, canopy architecture and colour distribution. the present analysis evaluates whether self-supervised pretraining of a Vision Transformer (ViT) backbone confers robustness to such drift relative to supervised ImageNet initialisation. Because no field imagery was available for this run, the study is framed as a computer robustness evaluation: a controlled, replication-based benchmark in which biotype drift is operationalised as a distribution shift between pretraining and evaluation domains, and detection performance is characterised across a declared grid of drift severities and label budgets. The evaluation uses a before-after-control-impact design with 8, 16, 24 and 32 clusters, effect sizes of 3, 5 and 8 percentage points, and 2000 replications per design cell. Under the declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), 52 clusters are required to detect a 5-percentage-point effect at 80% power.