Stacking Ensemble Deep Learning for Automated Broiler Breeder Feather Scoring
Year: 2026
Type: article
Abstract: Assessing feather condition in broiler breeders is essential for monitoring poultry welfare and identifying feather damage and loss associated with management, health, natural mating, and behavioral challenges. Feather condition can be associated with fertility failures, as featherless areas can suffer scratches from roosters during mating, reducing reproductive success. Manual feather scoring (FS) is labor-intensive, subjective, and difficult to scale in broiler breeder production. This study proposed an explainable stacking ensemble framework for automated multiclass FS classification using RGB images. A total of 67,231 images were categorized into six FS classes (FS_0-FS_5) and used to train three complementary transfer-learning models: MobileNetV2, EfficientNetB0, and ResNet50. Their class probability outputs were integrated using a Logistic Regression (LR) meta-learner. The proposed stacking ensemble achieved superior performance, with an accuracy of 99.84%, precision of 99.85%, recall of 99.83%, F1-score of 99.84%, AUC of 99.99%, and a quadratic weighted Kappa value of 0.9997, corresponding to 99.97% ordinal agreement. The quadratic weighted kappa demonstrated near-perfect ordinal agreement, confirming minimal severe misclassification across feather-score categories. The experimental results demonstrate high accuracy and reliable inference on workstation hardware, indicating potential for real-time FS monitoring with future optimization for edge-device deployment. Five-fold cross-validation confirmed strong robustness and generalization, while Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations showed that the model focused on biologically meaningful feather regions. Compared with existing state-of-the-art (SOTA) methods, the proposed stacking ensemble achieved higher accuracy on a more challenging six-class task. These findings demonstrate the potential of stacking ensemble learning for objective, scalable, and interpretable poultry welfare monitoring.
Source: AgriEngineering
Institutions: North Carolina State University, Tribhuvan University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $1,930
OpenAlex ID: W7170618206