Examining the Research Gap in Transfer Learning for Agricultural Image Classification in West African Contexts

Transfer learning enables effective image classification with limited labelled training data, and evaluating its performance for agricultural applications in West African agro-ecological contexts addresses an important gap between global deep learning advances and local food system data science needs. This study examined the research gap in transfer learning for agricultural image classification in West African and specifically Nigerian contexts through systematic review and original experimental study. A systematic scoping review identified 143 publications from 2019 to 2024 on transfer learning for agriculture, of which only 7 included West African crops and zero focused on Nigerian staple crop disease or quality classification. Original experiments compared five pre-trained CNN architectures (ResNet-50, VGG-16, EfficientNet-B4, MobileNetV3, and DenseNet-121) fine-tuned for cassava leaf disease classification on a 4,200-image dataset compiled from South East Nigerian field photography. EfficientNet-B4 achieved 91.3 percent classification accuracy across six disease classes, significantly outperforming the closest competitor ResNet-50 at 88.7 percent. Model performance degraded substantially (to 74.2 percent) when tested on images captured under North West Nigerian lighting conditions, suggesting domain shift challenges. Data augmentation with brightness and contrast jitter reduced this degradation to 81.6 percent. The study fills a critical gap in transfer learning evaluation for Nigerian crop disease contexts and recommends establishing a Nigerian Agricultural Image Repository to support model development, and deploying EfficientNet-B4-based apps through IITA extension channels.

Keywords: transfer learning, agricultural image classification, cassava disease, Nigeria, CNN