Corn rust detector

Upload a close-up of one corn leaf. It is checked for common rust, and you can see where the model looked.

No photo handy? Try a or a .

Method

Leaf-only crop. Each photo is cropped to the largest square that contains only leaf, so the background can't decide the answer. Our first models learned exactly that shortcut.

Input check. A pretrained EfficientNet-B0 compares the photo with four reference groups: corn rust or healthy, other corn diseases, other crops, and non-plant photos. Only photos closest to the first group go on to the classifier.

Classifiers. SmallCNN (0.39M parameters, trained from scratch), MobileNetV3-Small and EfficientNet-B0 (ImageNet, fine-tuned), trained on two corn photo sources. The ensemble averages SmallCNN and MobileNetV3. Held-out test accuracy: SmallCNN 99.7%, MobileNetV3 99.4%, EfficientNet-B0 98.5%.

Heatmap. Grad-CAM marks the regions that pushed the prediction.

Photos are analyzed in memory and never stored.

Limits