Early In Ovo Sex Identification of Chicken Embryos Using a Dual-Stage Difference-Enhanced Spectral Fusion Network Based on Visible-Near-Infrared Transmission Spectroscopy.

Early in ovo sex identification is important for improving hatchery resource utilization; however, sex-related optical differences are weak during early embryonic development and exhibit clear stage dependence. In this study, longitudinal visible-near-infrared transmission spectra were acquired from 1600 fertilized pink-shelled eggs of Bashang Long-tailed chickens across four independent incubation batches during incubation days 1-7. Before model development, 200 eggs were reserved as a fixed internal hold-out test set, while the remaining 1400 eggs were used for exploratory incubation-day and feature screening, five-fold cross-validation, and model comparison. Among all 21 dual-day combinations, D2 + D5 achieved the highest exploratory validation AUC of 0.951. The selected scheme required separate measurements on D2 and D5, with the final sex classification completed on D5. Further incorporation of the signed difference spectrum ΔS and relative change rate R increased the AUC to 0.965, demonstrating the incremental discriminative value of longitudinal developmental-change information. After the input scheme had been fixed, DSSF-Net achieved the numerically highest performance among the evaluated models under unified egg-level stratified five-fold cross-validation, with an accuracy of 0.915 ± 0.015, an AUC of 0.974 ± 0.012, and an F1-score of 0.915 ± 0.014. The final model achieved an accuracy of 0.915 and an empirical AUC of 0.966 (95% CI: 0.940-0.987) on the fixed internal hold-out test set. Complete four-round leave-one-batch-out validation yielded mean accuracy and AUC values of 0.846 ± 0.014 and 0.926 ± 0.012, respectively, indicating that the longitudinal spectral information from D2 + D5 maintained relatively stable discriminative performance across incubation batches within the investigated population. These results demonstrate the feasibility of longitudinal spectral modeling for early sex identification in fertilized pink-shelled eggs of Bashang Long-tailed chickens. Systematic offline validation was completed under laboratory conditions, including evaluation on the fixed internal hold-out test set and four-round leave-one-batch-out validation, supporting the effectiveness and batch-level stability of D2 + D5 longitudinal developmental spectra. Post hoc wavelength-level interpretation further identified 594-620 nm and 660-675 nm as two candidate spectral regions with relatively high predictive value. Collectively, these findings provide a reliable longitudinal spectral modeling framework for early non-destructive in ovo sex identification and establish an experimental and methodological foundation for wavelength selection, device development, and subsequent engineering validation of multispectral detection systems under hatchery conditions.