A Robust Ensemble Deep Learning Model for Lumpy Skin Disease Identification

Early diagnosis of Lumpy Skin Disease (LSD) in cattle is crucial for maintaining meat quality and ensuring livestock productivity. This study introduces an innovative Ensemble Deep Learning (EDL) framework with Contrast Limited Adaptive Histogram Equalization (CLAHE) preprocessing, L2 regularization-based feature optimization, and adaptive ensemble weighting to efficiently and accurately detect LSD. Comparative experiments were conducted with feature selection methods, such as Information Gain (IG), Chi-square, and L1 regularization, as well as ensemble methods, such as Simple Averaging, Majority Voting, and Stacking. The L2-norm-based proposed ensemble attained the highest Dice score of 97.50%, the highest Jaccard index of 98.20%, and the best AUC of 0.980 at α = 0.01. It also attained an accuracy of 99.70%, precision of 98.73%, recall of 96.20%, and a 97.80% F1-score, with a moderate Inference Time (IT) of 0.17 s per image. In comparison to marginally slower but simpler ensemble methods, it also exhibited large improvements in robustness, generalization, and feature stability. Overall, when combined, the ensemble optimization and feature regularization enable the Regularized-ensemble Convolutional Neural Network (RoFR-eCNN) to deliver precise, accurate LSD detection, providing a scalable and efficient framework for automatic and early livestock disease diagnosis.