Spectral Domain Adaptation for Artificial Intelligence Agriculture Weed Detection Under Variable Illumination: A Computer Misclassification-Rate Evaluation

Weed detection models trained under one illumination regime and deployed under another suffer systematic spectral shift that inflates misclassification rates in ways aggregate accuracy conceals. The framework combines a physically motivated illumination-invariant spectral representation with an adversarial domain adaptation objective and a class-conditional calibration step. Because no field imagery was collected for the present analysis, the evaluation characterises the operating characteristics of the framework under a declared variance structure rather than reporting observational performance: between-cluster standard deviation was 6.0, residual standard deviation 4.5, and the intraclass correlation was 0.64, with power 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. Under this structure, 54 clusters were required to detect a 5 percentage point reduction in misclassification rate at 80 percent power. The dominant threat to weed-detection validity under variable illumination is not mean spectral shift but class-conditional shift concentrated in the minority weed class, and that adaptation gains measured on aggregate misclassification rate overstate the benefit available to the weed class by a wide margin. The contribution is a formal statement of the conditions under which spectral domain adaptation is identifiable, together with a design-level account of the cluster counts needed to evaluate it credibly.