Semi-Supervised Semantic Segmentation for Artificial Intelligence Agriculture Field Boundary Delineation Under Sparse Cadastral Annotation: A Computer Boundary-F1 Evaluation

Field boundary delineation from satellite imagery is a prerequisite for plot-level agricultural monitoring, yet cadastral reference data are sparse across most smallholder systems, and the boundary-F1 metric that governs segmentation quality is rarely characterised under annotation scarcity. the present analysis treats semi-supervised semantic segmentation as a design problem whose operating characteristics can be derived rather than merely reported. It formalises boundary-F1 as a function of annotation density, pseudo-label error propagation and cluster-level heterogeneity, and it evaluates the resulting estimator through a replication-based design in which the before-after-control-impact estimator is applied to 2000 replications per design cell across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, under a declared variance structure with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.640.54 clusters are required to detect a 5 percentage-point boundary-F1 gain at 80% power, that the minimum detectable effect falls steeply between 8 and 24 clusters and flattens thereafter, and that the intraclass correlation of 0.64 dominates the power budget far more than annotation density. The contribution is a power-analytic account of when semi-supervised boundary delineation can be distinguished from its supervised baseline, together with the boundary conditions under which reported boundary-F1 differences are artefacts of design rather than segmentation quality.