Challenges in zero-shot cross-domain transfer for plant disease classification

Population growth has increased the demand for food and water, leading to higher consumption and waste, which burdens society and the environment. Urgent action is needed to reduce crop damage and ensure food security. Deep Learning (DL) models require large, annotated datasets, which are difficult to obtain. We adopt a zero-shot cross-dataset transfer (ZS-CDT) approach for plant disease classification, merging publicly available datasets to improve detection accuracy. We took the leverage of a pipeline built on top of pre-trained backbone models and applied data augmentation, conducted several experiments, and found a performance gap between cross-datasets, highlighting the need for domain-generalised models.