Graph-matching domain adaptation for cross-border cropland masks: an artificial intelligence evaluation of label-scarcity transfer failure in national agricultural statistics

Abstract

Cross-border transfer of cropland masks is routinely proposed as a remedy for label scarcity in national agricultural statistics, yet the conditions under which graph-matching domain adaptation fails have not been formalised. The analysis treats label scarcity as a structural property of the target domain rather than a sampling deficiency, and derives a bound on target risk that separates three additive components: graph-matching distortion, label-scarcity variance, and the divergence between source and target cropland definitions. The bound is used to identify a transfer-failure threshold below which adaptation is dominated by the unadapted source classifier. A formal proposition establishes that graph-matching distortion grows with the spectral gap between source and target adjacency structures, and a corollary shows that this distortion is not reduced by increasing the number of unlabelled target samples. The paper does not report empirical estimates; it provides the formal apparatus, the failure criterion, and the diagnostic quantities that a national statistical office would need to compute before adopting a transferred mask. The contribution is a specification of when cross-border graph-matching adaptation should be rejected, and a set of computable indicators that make that rejection auditable.