Submitted to: Meeting Proceedings
Publication Type: Proceedings
Publication Acceptance Date: June 10, 2007
Publication Date: July 8, 2007
Citation: Smith, D.P., Lawrence, K.C., Heitschmidt, G.W., Park, B., Windham, W.R. 2007. Evaluation of Hyperspectral Imaging and Predictive Modeling to Determine Fertility and Development of Broiler Hatching Eggs. Poultry Science Meeting, p. 86 (suppl 1): 385. Interpretive Summary: Approximately 30 billion broiler hatching eggs are produced in the world each year. Up to 3 billion are not fertile, yet are incubated anyway. An accurate and automated system to remove these eggs prior to incubation would be an invaluable aid to poultry hatcheries. A camera system was configured to image fertile and infertile eggs prior to incubation and then each during the first 3 days of incubation. Data from the camera was then analyzed with a mathematical modeling process to separate fertile from infertile eggs. Although the initial testing and analysis was promising, eventually the model system did not produce results that were accurate enough to be considered for a commercial system. Further analysis and modeling of the collected data will continue to configure a more accurate predictive model.
Technical Abstract: A hyperspectral imaging system and a predictive modeling technique was evaluated for determining fertility and early embryo development of broiler chicken hatching eggs. Twenty-four broiler-strain eggs were collected (12 fertile, 12 infertile) for each of 8 replicate trials (n=192) and imaged on Days 0, 1, 2, and 3 of incubation for training and validation. Three replications of 30 eggs each (fertile and infertile eggs randomly mixed) were collected and imaged as above for verification (n=90). A tungsten-halogen lamp provided back illumination candling for eggs positioned vertically in relation to the imaging system, which consisted of a lens, spectrograph, and CCD camera mounted on a stand above the eggs. Spatial and spectral data from approximately 400 to 1000 nm were collected for each egg on each day of incubation, then refined to 550 to 899 nm. A Mahalanobis Distance (MD) supervised classifier was trained with spectral data from the first 5 replicate sets of eggs, then Principal Component Analysis (PCA) was performed. This model was applied to the next 3 sets for model validation and also the 3 sets of 30 eggs each (blind, randomly mixed fertile/infertile eggs) for verification. Fertility was confirmed on Day 5 of incubation by candling and breakout. The MD/PCA model predictions for the 3 validation sets of eggs were: 71% accuracy for Day 0; 63% for Day 1, 65% for Day 2, and 83% for Day 3. For the 3 sets of blind verification eggs, the MD/PCA model accurately predicted 46/90 on Day 1 and 45/90 on Day 3. The data indicate that the particular MD/PCA model used is not appropriate for predicting fertility and early development. Data will be further analyzed by other methods in an effort to reduce false positive and negative results.