Answers are grounded in this paper and cite their sources.
Tomato cultivars differ in soluble solids content (SSC) and fruit optical properties, making chemical reference replication a critical design factor for hyperspectral regression. We analysed 480 fruit-level visible–near-infrared spectra (125 bands, 401–997 nm) from eight cultivars (60 spectra each) paired with archived cultivar-mean SSC references, comparing cultivar-stratified random testing with eight leave-one-cultivar-out (LOGO) tests across classical regressors and the original DG-SSDA v1 and v2 implementations. Under random splitting, three-seed DG-SSDA v2 reached R² = 0.7837, while the known-cultivar mean baseline reproduced the archived targets exactly (R² = 1.0000), revealing that random splits reward cultivar recognition rather than SSC calibration when references are replicated at cultivar level. LOGO evaluation supplied the rigorous test of generalisation: v2 achieved pooled spectrum-level R² = −1.74 and cultivar-level R² = −1.56 (RMSE 0.66 °Brix vs. 0.47 °Brix for the training-mean baseline), delimiting the generalisation boundary of current spectral regression under cultivar-mean supervision. These results establish a reproducible distinction between spectral replication and chemical reference replication and provide concrete design guidance for cultivar-diverse SSC studies.
No takes yet. Share an insight, caveat, or question.
Tian et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: