New AI framework uses synthetic disease images to boost potato crop health detection

A team of researchers has developed PotatoGANs, a machine learning approach designed to overcome a persistent challenge in agricultural automation: deep learning models that work well in labs often fail when deployed in real farm conditions.

The challenge of overfitting in crop disease detection#

Training neural networks to identify crop diseases typically requires large volumes of labeled image data. When models are exposed to limited datasets during training, they can become overspecialized to those particular examples, performing poorly when encountering images taken under different lighting, weather, or field conditions. Traditional image augmentation techniques—such as rotating, flipping, or shifting images—have proven insufficient to generate the diversity needed for robust generalization, especially for sensitive applications like potato farming where disease outbreaks directly threaten yields and farm profitability.

How PotatoGANs generates synthetic training data#

Rather than manually collecting more disease images, the researchers employed Generative Adversarial Networks (GANs) to create synthetic but realistic potato disease photographs. The system takes images of healthy potatoes and generates corresponding images showing two common potato diseases: black scurf and common scab. Two different GAN architectures were tested—CycleGAN and Pix2Pix—with CycleGAN producing higher-quality synthetic images as measured by the Inception Score metric (1.2001 for black scurf, 1.0900 for common scab).

The synthetic images expand training datasets while introducing visual variation that helps models learn disease signatures rather than memorizing specific image patterns. This approach reduces the cost and labor of collecting real disease samples while improving the diversity of training material.

Improving model interpretability#

The researchers further enhanced their system by layering Explainable AI techniques on top of the disease classification models. They combined three visualization methods—GradCAM, GradCAM++, and ScoreCAM—with three standard CNN architectures (DenseNet169, ResNet152 V2, and InceptionResNet V2). These explainability tools help reveal which regions of an image the model focuses on when making predictions, making the system more transparent and trustworthy for agricultural practitioners who need to understand why a model flagged a particular plant.