Semi-Supervised Phenological Stage Alignment for Artificial Intelligence Agriculture Yield Inference Under Uneven Ground Observations: A Computer Temporal-Label-Noise Evaluation

Phenological stage labels collected from ground observations are unevenly distributed across space and time, and the resulting temporal label noise propagates into any downstream yield inference that treats those labels as ground truth. the present analysis formalises the problem as a semi-supervised alignment task in which a small set of stage-labelled observations must be reconciled with a larger pool of unlabelled or weakly labelled crop sequences, and it evaluates the operating characteristics of a before-after-control-impact (BACI) estimator under a declared variance structure. No observational data are used or claimed: the analysis is a replication-based design evaluation in which the between-cluster standard deviation was set to 6.0, the residual standard deviation to 4.5, and the intraclass correlation to 0.64, with 2000 replications per design cell across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis found that 54 clusters were required to detect a 5 percentage point effect at 80 per cent power under this variance structure, and that power curves, minimum detectable effect curves, coefficient intervals, residual diagnostics and kernel density estimates of estimated power all behaved in the manner the design theory predicts. The contribution is a formal statement of the temporal-label-noise problem, a semi-supervised alignment operator with stated consistency conditions, and a design-level characterisation of how much ground observation is needed before BACI inference on phenological alignment becomes reliable. The implication is that temporal label noise is not a nuisance to be cleaned post hoc but a design parameter that governs the cluster budget of any yield inference pipeline.