Potato Leaf AI Study Reveals Visual Similarity Is Not Leaf Identity

A team of Peruvian computer scientists says a common assumption in plant-pathology AI may be misleading for potato leaf disease image testing: that visually similar photos must depict the same leaf. Published in Frontiers in Computer Science, the study led by Hugo Vega-Huerta of Universidad Nacional Mayor de San Marcos ran 240 training runs across four neural network architectures, two datasets, and 15 random seeds. Using PlantVillage lab images (2,152 total) mapped to 538 physical leaf identities, the researchers created holdout splits where a leaf never appears in both training and testing. They then compared DINOv2-based visual clusters to the known identities. Results were poor—pair precision was 3.89%—showing that clustering alone cannot prove duplicates. The field dataset lacked identity metadata, so duplicates were grouped via DINOv2 clusters without direct identity checks.