Non-destructive sex determination of chicken hatching eggs using hyperspectral imaging and GA-LightGBM

Male chick culling in the layer industry creates major animal welfare and production-efficiency concerns. Hyperspectral imaging (HSI) offers a rapid, non-destructive alternative to invasive in-ovo sexing approaches, but its robustness for commercial use remains uncertain. In this single-breed, single-hatchery proof-of-concept study, daily transmission HSI images of Dawu Brown chicken eggs were acquired from incubation day 0 to day 14, with model development focused on day 13. We developed EggForecast, integrating Pearson correlation coefficient (PCC)-based spectral weighting, genetic algorithm (GA) wavelength optimization, and LightGBM classification. Six feature-processing or selection methods—RF, PCA, SPA, PLS, LDA, and GA—were compared. On an internal independent test set, GA-LightGBM achieved accuracy, precision, recall, F1-score, and Cohen’s Kappa values of 0.925, 0.938, 0.882, 0.909, and 0.845, respectively. GA selected 27 informative wavelengths while maintaining performance comparable to the full-spectrum model in this dataset. Repeated validation over 50 random partitions showed performance significantly above random classification and no significant degradation relative to the full-spectrum model. These findings support the feasibility of HSI-based non-invasive sex determination in the studied population, but validation across hatcheries, breeds, batches, and earlier incubation windows is required.