Impact of hyper-parameter tuning on CNN accuracy in agricultural image classification

This study explores the impact of hyper-parameter optimization on the performance of convolutional neural networks (CNNs) for olive cultivar classification using transfer learning. Pre-trained ImageNet models such as VGG16, InceptionV3, and ResNet50 were adapted to a proprietary dataset, with VGG16 selected for detailed evaluation. Key hyper-parameters, including layer count, neurons per layer, dropout rate, learning rate, and batch size, were tuned using random search. The best configuration achieved a validation accuracy of 87.5%, significantly outperforming the control model. Sensitivity analyses with Morris and Sobol methods identified the number of layers as the most influential factor, followed by dropout and learning rates through interaction effects. These findings demonstrate the importance of tailoring CNN architecture and regularization settings to the problem domain. These results underscore the importance of tuning architectural depth and regularization mechanisms for performance optimization. As a practical guideline, models with fewer layers and intermediate dropout levels demonstrated higher robustness and generalization, offering an effective strategy for adapting CNNs to agricultural classification tasks. • Systematic analysis of the impact of hyper-parameter tuning on CNN performance. • Sensitivity analysis reveals the number of layers as the most influential hyper-parameter. • Innovative application of transfer learning for olive cultivar classification. • Publicly available dataset ensures reproducibility and transparency (DOI link). • Comprehensive framework for optimizing CNNs in agricultural tasks with practical applications.