Abstract
Confirming seed variety identity is important in certification, breeding-material management and grain trade, but visually similar cultivars remain difficult to distinguish consistently. We present GAFE-Stack, a morphology-guided classifier that expands seven kernel measurements into twenty-one interpretable descriptors and combines five complementary learners. The method is evaluated on the small, balanced public UCI Seeds benchmark (N=210; 70 kernels per variety), whose measurements were extracted from soft X-ray images. Across ten repeats of stratified five-fold cross-validation, GAFE-Stack achieved 96.33 ± 2.29% accuracy and 96.67% under leave-one-out validation. Its observed mean differences from seven re-implemented references ranged from 0.48 to 4.29 percentage points; after accounting for dependence among repeated folds and applying Holm adjustment, none of the comparisons was significant at α=0.05. The strongest individual member, LightGBM, achieved a slightly higher mean accuracy (96.76%), whereas GAFE-Stack had lower fold-level dispersion (2.29% versus 2.60%) and fewer pooled Kama–Canadian confusions than the RBF-SVM baseline. The complete pipeline achieved 93.33% and 94.52% accuracy with 30 and 60 labelled training kernels, respectively, compared with 96.33% using the full training folds. From stored morphometric inputs, single-thread CPU training required 0.85 s and batch prediction processed approximately 52,600 kernels/s without a GPU. Applying the same dimension-typed construction rules to Raisin, Rice and Dry Bean datasets produced small positive mean changes of 0.09–0.37 points, with corrected intervals including zero. Lot-purity results are reported only as an exploratory resampling analysis of UCI observations. The present evidence therefore supports GAFE-Stack as an interpretable proof-of-concept approach to seed screening under benchmark conditions; validation on independently acquired kernels, measurement systems and physical seed lots remains future work.