Applications of transfer learning in sunflower disease detection: advances, challenges, and future directions

Sunflower (Helianthus annuus) is a crop of high economic and nutritional importance that continues to suffer significant yield losses due to foliar diseases. Traditional image-based and laboratory detection techniques remain limited by subjectivity, cost, and scalability. Transfer learning (TL) has recently emerged as an effective approach to overcoming these challenges involving the reuse of pretrained deep models for plant pathology tasks. Presented here is a systematic examination of recent TL-based studies on sunflower disease classification to identify prevailing trends, research gaps, and future opportunities.

A structured Scopus query was employed to retrieve peer-reviewed articles published between 2021 and 2025. Strict inclusion and exclusion criteria ensured technical relevance to TL-based sunflower disease detection. Subsequently, 30 studies meeting the criteria were critically reviewed and analyzed in terms of model architecture, dataset characteristics, preprocessing strategies, and reported evaluation metrics. The comparative assessment focused on convolutional neural networks (CNNs), transformer-based architectures, and hybrid models.

The analysis revealed a dominant reliance on pretrained CNNs such as ResNet, VGG, Inception, and EfficientNet. Several studies employed lightweight or federated learning variants to enhance deployment feasibility under field conditions. Among the commonly observed challenges were limited dataset diversity, class imbalance, and insufficient explainability. A key word cooccurrence analysis indicated an evolving research focus, transitioning from basic deep learning implementation to explainable and privacy-preserving frameworks optimized for edge devices.

The review revealed substantial progress in TL applications for the diagnosis of sunflower disease but underscored the need for larger, standardized datasets and cross-regional validation. Future studies should prioritize interpretable, adaptive architectures that can function in real-world agricultural environments. The insights drawn from this synthesis extend beyond sunflower pathology, offering a foundation for scalable, domain-transferable TL solutions in broader plant disease detection contexts.