Plant Disease Segmentation

U-Net Segmentation

Identify visible disease regions on a plant leaf photo using a U-Net trained from scratch, pixel by pixel, not just a single sick/healthy label.

Trained from scratch

No pretrained backbone. 4-stage encoder, bottleneck, 4-stage decoder with skip connections. Every layer built directly in PyTorch on the PlantSeg dataset.

Model stats

75% whole-image classification accuracy on held-out validation

11,000+ images from PlantSeg (Kaggle / Zenodo)

15 classes across 15 plant species, chosen by sample size and visual diversity

Input fixed at 256×256, images resized before inference

What makes it different

Pixel-level mask output, not a yes/no label

Combined cross-entropy + Dice loss; every class improved after the switch

Mask bug caught by raw pixel inspection, not symptoms. Values were 4-114, not 0/255

Exported to ONNX and verified against PyTorch output before deployment

Known limitation

Apple scab lesions cover a median of ~6% of the image vs 40%+ for the strongest class. Small, sparse targets are harder for a fixed-architecture segmentation model. Explainable, not a bug.

Upload a clear, well-lit photo of a single affected leaf for best results.

Upload Leaf Photo

Click to upload, or drag a photo here

JPG or PNG, ideally a single leaf, filling most of the frame

Change

Analysis takes 8-12 seconds. The model runs segmentation pixel by pixel across the full image.

No Image Selected

Please upload a leaf photo before running the analysis.

Analyzing your leaf

Preparing image...

This usually takes 8-12 seconds