Determination of brown egg fertility by multi-directional visible/near-infrared spectroscopy and machine learning

Poultry products provide affordable proteins to meet daily nutrient requirements for human. In aligned with the growing human population, the poultry industry has been intensified and mechanized over the past decades. Hatchery is a key component in the modern integrated poultry industry by providing chicks for commercial production. Detection of egg fertility during the first few days of incubation allow hatcheries to exclude infertile or problematic eggs in a timely manner, reducing incubator space, utility costs, and contamination from exploder eggs. The objective of this research was to investigate brown egg fertility detection via multi-directional point visible/near-infrared (VIS/NIR) spectroscopy and machine learning. A total of 130 brown eggs were measured with the point VIS/NIR spectroscopy from days 0 to 5 of incubation. Spectral images of the eggs were collected from three directions (X, Y, and Z axes) and in the spectral range of 190-1100 nm. Four preprocessing methods (e.g., Savitzky-Golay, Gaussian, Multiplicative Scatter Correction, and Standard Normal Variet) were separately or jointly applied to filter un-wanted data and noises in spectrums. Then feature selection was performed using Principal Component Analysis and T-test. Finally, the Support Vector Machine classifier was used to determine brown egg fertility based on the filtered data and extracted features. The results indicate that the best precisions (%) of determining egg fertility on days 0, 1, 2, 3, 4 and 5 of incubation were 90.3, 90.9, 91.3, 91.9, 92.3, and 94.3, respectively. The proposed framework is a useful tool to identify egg fertility in early stages of incubation and support efficient and sustainable poultry production.