A research team has built a pipeline that teaches neural networks to spot potato diseases without needing large libraries of real infected crops.
The system, called PotatoGANs, feeds healthy potato images into two generative adversarial network models — CycleGAN and Pix2Pix — which produce synthetic images of diseased plants. Those fakes then pad the training dataset, giving classifiers more variety to learn from. The researchers tested three CNN architectures (DenseNet169, ResNet152 V2, and InceptionResNet V2) on the resulting data and measured image quality using Inception scores. CycleGAN outperformed Pix2Pix, scoring 1.2001 for black scurf and 1.0900 for common scab. Three gradient-based explainability tools — GradCAM, GradCAM++, and ScoreCAM — were layered on top to show which image regions drove each classification decision.
The approach targets a real bottleneck in agricultural AI: disease image datasets are small, expensive to label, and skewed toward conditions photographed in controlled settings. Models trained on them tend to overfit and fall apart in the field. Synthetic augmentation is not new, but applying it specifically to close the healthy-to-diseased image gap is a cleaner framing than generic rotation-and-flip pipelines.
Potato blight famously reshaped modern history; whether a GAN pipeline reshapes modern farming is a different question entirely — field deployment data would be more persuasive than Inception scores alone.