Benchmarking Graph-Neural Yield Forecasting for Computer Agriculture Under Intercropping Canopy Mixing: An Artificial Intelligence Spectral-Unmixing Bound for Smallholder Plot-Level Estimation

Intercropping mixes the spectral signatures of co-planted species inside a single smallholder plot, so that a graph-neural yield forecaster trained on pure-stand canopy reflectance receives an input whose graph nodes no longer correspond to the crop whose yield is being predicted. the present analysis treats that mismatch as a bounded estimation problem rather than a data-cleaning nuisance. It formalises plot-level reflectance as a convex mixture of species endmembers weighted by fractional canopy cover, defines the spectral-unmixing bound as the maximum yield-forecast error attributable to residual endmember ambiguity at a declared signal-to-noise ratio, and derives the conditions under which a graph-neural forecaster's error decomposes into an unmixing term and a graph-aggregation term. The derivation is illustrated with a design-based evaluation of a before-after-control-impact estimator under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), in which power was estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. The analysis is theoretical and methodological: no observational data are used or claimed. The bound implies that below roughly 54 clusters a 5-percentage-point yield effect cannot be separated from unmixing error at 80% power, which reframes intercropping benchmarking as a problem of spectral separability rather than of model capacity.