Benchmarking Evaluation of Semantic Segmentation for Computer Agriculture Boundary Mapping Under Sub-Meter Imagery: An Artificial Intelligence Area-Error Bound for Land-Tenure Adjudication

Sub-metre imagery now resolves field boundaries at the scale of cadastral adjudication, yet the segmentation models that extract those boundaries are benchmarked with pixel-overlap metrics that do not translate into the quantity a tenure tribunal actually needs: the area error of an adjudicated parcel. The bound decomposes adjudicated area error into a boundary-localisation term, a topological term penalising spurious and missed parcels, and a sampling term that propagates residual variance through the before-after-control-impact estimator used to attribute error to model revision. Under a variance structure with between-cluster SD 6.0, residual SD 4.5 and intraclass correlation 0.64, power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. Fifty-four clusters were required for 5 percentage points at 80% power. The analysis is a design and power study, not an observational evaluation: no imagery, parcels or adjudications were observed, and the reported quantities are the operating characteristics of the proposed benchmark under the stated assumptions. The contribution is a bound that converts segmentation benchmark output into an adjudication-relevant error statement, together with the design parameters under which that statement is estimable.