Spectral-Spatial Distillation for Artificial Intelligence Agriculture Crop-Type Mapping Under Cloud-Contaminated Satellite Time Series: A Computer Label-Efficiency Guarantee Evaluation

Cloud contamination in optical satellite time series degrades crop-type mapping in smallholder agricultural systems, and the cost of dense reference labels makes exhaustive supervised training infeasible. The framework is formalised as a label-efficiency guarantee: under stated assumptions on cloud-cover probability, spectral separability and spatial autocorrelation, the student's generalisation error is bounded by the teacher's error plus a distillation gap that shrinks with the fraction of cloud-free observations per pixel. A before-after-control-impact design is used to characterise the operating characteristics of the evaluation that would test this guarantee, with a between-cluster standard deviation of 6.0, a residual standard deviation of 4.5 and an intraclass correlation of 0.64; power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design requires 54 clusters to detect a 5-percentage-point improvement at 80% power. The contribution is a formal statement of when spectral-spatial distillation can substitute for dense labels, together with the design parameters that a field evaluation would need to detect the resulting gains.