Influence of Optimizers on the Performance of Deep CNN for Cattle Lumpy Skin Disease Diagnosis

Abstract

Lumpy skin disease (LSD) poses a significant threat to cattle industry, leading to reduced productivity and economic losses. Deep learning can be leveraged for effective differentiation of LSD-affected from healthy cattle. However, the effectiveness of the deep learning based solution depends on the models ability to learn discriminative features that accurately distinguish between the two classes. In addition limited attention has been given to understand, optimizer influence on feature learning, convergence behavior, and classification accuracy across different network architectures. To address this gap, this study presents a comprehensive analysis of the impact of five distinct optimizers that include Adam, Adagrad, Adamax, Nadam, and SGD on seven benchmark deep learning architectures, namely DenseNet201, ResNet152V2, ResNet101V2, InceptionV3, DenseNet121, Inception ResNetV2, and Xception for LSD diagnosis. The proposed methodology uses an initial 80:10:10 train–validation–test split-based analysis, followed by five-fold cross-validation with class-weight balancing to improve the robustness of performance assessment. Comparative analyses were performed using accuracy, precision, recall, F1-score, confusion matrices, loss landscape visualization, and statistical validation through one-way ANOVA with Tukey’s HSD post hoc test. ResNet101V2-Nadam, InceptionResNetV2-Adam and DenseNet201-Nadam consistently achieved the highest classification performance across the cross-validation folds, exhibiting better discrimination capability and stable optimization behavior. One-way ANOVA confirmed that optimizer selection significantly influenced classification performance (p = 0.003663), with Adam and Nadam exhibiting better classification accuracy with reduced variability across the cross-validation folds compared with other optimizers. The DenseNet201-Adam and DenseNet121-Nadam achieved classification accuracy of 99.69% and 99.38% respectively on the independent test set, along with attaining a recall of 100% and 99.07% for LSD. The radar plots illustrating the precision, recall, and F1-score for all model–optimizer combinations on the independent test dataset demonstrate that DenseNet121–Nadam achieved the highest precision (1.00), whereas DenseNet201–Adam achieved the highest recall (1.00); both combinations attained an F1-score of 0.99, demonstrating highly accurate classification performance. Overall, the results demonstrate that optimizing the training process through appropriate optimizer selection can unlock the full potential of existing deep learning architectures by improving precision, recall, F1-score, convergence behaviour, overall classification accuracy and generalization on unseen data while avoiding additional architectural complexity.