Application of spectral techniques and machine learning for fertility, mortality, sex, and structural evaluation of eggs

Ahmed, MD Wadud

Permalink

https://hdl.handle.net/2142/129749 Copy

Description

Title

Application of spectral techniques and machine learning for fertility, mortality, sex, and structural evaluation of eggs

Author(s)

Ahmed, MD Wadud

Issue Date

2025-05-01

Doctoral Committee Chair(s)

Kamruzzaman, Mohammed

Committee Member(s)

• Emmert, Jason Lee
• Rausch, Kent D
• Malvandi, Amir

Department of Study

Engineering Administration

Discipline

Agricultural & Biological Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Hyperspectral imaging
• NIR spectroscopy
• Machine learning
• Fertility
• Embryo mortality
• Egg sex
• Shell strength
• Shell thickness
• Yolk ratio
• Animal welfare
• Egg industry.

Abstract

Eggs are considered one of the best dietary protein sources and have an important role in human growth and development. Fast and accurate quality assessment and grading of eggs is crucial to meet the growing demand for high-quality eggs driven by rapid population growth, economic sustainability, and increasing animal welfare concerns. Evaluation of some key parameters, such as fertility, mortality, sex, shell thickness, shell strength, and yolk ratio (i.e., yolk to egg mass ratio), is essential for optimizing hatchery productivity, ensuring egg quality, and maintaining ethical standards in the egg industry. Conventional methods for evaluating these parameters, such as candling, manual measurement, and mechanical testing, are inherently labor-intensive, time-consuming, and destructive, highlighting the need for non-destructive techniques in the egg industry. Spectral techniques such as near-infrared (NIR) spectroscopy and hyperspectral imaging (HSI) offer rapid, non-invasive, and accurate alternatives to conventional techniques, enabling the simultaneous assessment of multiple parameters without compromising the integrity of eggs. Spectral techniques collect spectra at numerous narrow and contiguous wavelengths. Consequently, chemometrics and machine learning (ML) techniques facilitate the analysis of these spectra by identifying complex patterns and relationships in the data, enabling accurate prediction, classification, and evaluation of egg parameters. The objective of this research was to apply spectral techniques (NIR spectroscopy and HSI) and ML for fertility, mortality, sex, shell thickness, shell strength, and yolk ratio analysis of eggs. This study can enhance automation, improve quality control, promote animal welfare, and optimize resource management for the sustainable development of the table and hatching egg industry. The poultry industry heavily relies on accurate detection of egg fertility to optimize hatchery operations. In this study, HSI spectra was used to develop pre-incubation chicken egg fertility classification models using XGBoost, CatBoost, RF, and SVM. Using full wavelengths, the CatBoost model with synthetic data showed the best classification performance, attaining 95.1% accuracy in independent validation. The CatBoost models with fewer important features showed good prediction performance, making them computationally efficient, robust, and interpretable. The Shapley additive explanation (SHAP) explainable AI technique was used to interpret the robust CatBoost model, revealing that wavelength regions associated with yolk color, pre-incubation cellular activities related to embryonic development, changes in hydration levels, and variations in protein and lipid contents between fertile and infertile eggs are crucial for pre-incubation chicken egg fertility classification. Non-destructive pre-incubation and early incubation chick embryo mortality prediction is crucial for optimizing hatchery operations and improving poultry production. This study explores potential of visible-near infrared (Vis-NIR) HSI combined with ML models to classify embryo mortality during pre-incubation and at 4 days of incubation periods. Partial least squares discriminant analysis (PLS-DA), RF, and CatBoost calibration models were developed, and performance of the calibration models was evaluated by independent validation and test sets. At full wavelength (501-921 nm), the PLS-DA model demonstrated the best performance for chick embryo mortality classification, achieving an accuracy of 91.3% in calibration, 88% in validation, and 86.7% in the test set for pre-incubation, while for day 4 incubation, it attained 97.3% accuracy in calibration, 96% in validation, and 97.3% in the test set, highlighting its robustness across different data sets. PLS-DA models with a reduced set of important spectral features demonstrated strong predictive performance, offering computational efficiency, robustness, and enhanced interpretability. SHAP model explantion revealed that wavelengths related to blood formation, embryo hydration status, and metabolic variations between live and dead embryos are critical for chick embryo mortality classification during early incubation. Non-destructive sex determination in eggs can enhance animal welfare, improve economic efficiency, reduce environmental impact, and foster technological innovation in sustainable hatchery operations. This study developed different classification models for pre-incubation sex prediction in chicken egg using HSI and ML. Partial least squares discriminant analysis (PLS-DA), XGBoost, RF, and CatBoost classification models were constructed using full-wavelength (452-899 nm) spectra and evaluated them through external validation. Multiple spectral pre-processing methods, feature selection approaches, and hyperparameter tuning were assessed to optimize and develop robust prediction models. Full featured CatBoost model exhibited superior performance among models tested, with 88.6% calibration accuracy and 75.5% validation accuracy. With 35 important features based optimized CatBoost model showed promising results with an accuracy of 82.9% and 75.5% in calibration and independent validation data. The eggshell protects internal contents, supplies calcium for embryos, and aids embryonic respiration. This study assessed NIR spectroscopy, ML, and explainable AI for real-time eggshell thickness prediction by developing partial least squares regression (PLSR), RF, K-nearest neighbors (KNN), and support vector regression (SVR) models using full wavelengths (1300-2525 nm) and selected important wavelengths. The PLSR model demonstrated superior and stable predictive performance, achieving a coefficient of determination (R2p) of 0.867, and root mean square error of prediction (RMSEP) of 0.015 mm. A new PLSR model using only five important variables showed promising results with an R2p of 0.910 and RMSEP of 0.012 mm. The SHAP explanation of the final PLSR model revealed wavelengths related to protein, moisture, and lipids are crucial for NIR spectroscopic prediction of eggshell thickness. Eggshell strength is crucial for ensuring high-quality eggs, reducing breakage during handling, and meeting consumer expectations for freshness and integrity. This study evaluated potential of NIR spectroscopy (1300-2525 nm), ML, and SHAP explainable AI for rapid, non-destructive method for detecting eggshell strength. Principal component analysis (PCA) and PLS-DA effectively classified eggs based on a threshold shell strength of 30 N. Regression models, including PLSR, RF, Light GBM, and KNN were used to predict eggshell strength and evaluated by independent validation test. Using only 14 selected important variables, the RF model achieved a very good prediction performance with coefficient of determination of prediction (R2p) of 0.83, minimum RMSEP of 1.49 N, and RPD of 2.44. The SHAP explanation of the RF model revealed wavelengths related to moisture (of shell and albumen) were crucial for NIR spectroscopic prediction of eggshell strength. In the final part, the potential of HSI combined with ML and explainable AI for yolk ratio prediction was investigated. Multiple full spectral (374-1015 nm) regression models were developed using the calibration set, validated with an independent dataset, and further tested using another independent dataset, ensuring robustness and generalizability across varying samples. The PLSR model showed superior predictive performance (R2: calibration=0.79, validation=0.73, test=0.68), with similarly robust results obtained using only 20 important wavelengths. SHAP explanation of the final PLSR model revealed wavelength region associated with water had the most influence on non-destructive yolk ratio prediction by HSI.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129749

Copyright and License Information

Copyright 2025 MD Wadud Ahmed

Owning Collections