Plantulus Doctor — plant disease classifier
EfficientNetV2-B0 (ImageNet pre-trained) fine-tuned to recognise 54 plant/disease classes across 19 plants from a photo of a single leaf. Live demo and JSON API: KumarjitC/indoor-plant-disease-gradio-app.
Files
| File | Use |
|---|---|
plantulus_doctor_fp16.tflite |
Float16-quantized TFLite model (~12 MB) for servers (LiteRT) and mobile |
plantulus_doctor.keras |
Full Keras 3 model |
classes.json |
Class labels in model output order |
class_info.json |
Per class: label, plant, disease, is_healthy, display name |
preprocessing.json |
Required input preprocessing |
Preprocessing
- Apply EXIF orientation and convert to RGB.
- Resize the shorter side to 256 px and centre-crop to 256×256.
- Centre-crop to 224×224.
- Feed float32 pixel values in [0, 255]. Do not divide by 255: the model normalizes internally.
Input shape [1, 224, 224, 3]; output: softmax probabilities over the 54 classes in classes.json order.
Training data
- PlantVillage (colour images only)
- New Plant Diseases (augmented PlantVillage)
- Indoor Plant Disease Detection
Images were deduplicated (MD5), and augmented copies were grouped by PlantVillage UUID and a rotation/flip-invariant perceptual hash so that copies of the same photo never appear in both train and test. Split ≈70/15/15, stratified by class.
Training: two stages (frozen backbone head training, then full fine-tuning with frozen BatchNorm and cosine learning-rate decay), random-crop/flip/rotation/zoom/colour augmentation, square-root class weights, label smoothing 0.1, mixed precision. Because of label smoothing, confident predictions typically score around 0.9 rather than 1.0.
Evaluation (held-out test set, 16,697 images)
| Subset | Images | Accuracy | Macro F1 |
|---|---|---|---|
| All | 16,697 | 98.04% | 0.984 |
| PlantVillage | 7,759 | 99.45% | 0.992 |
| New Plant Diseases | 5,919 | 99.27% | 0.936 |
| Indoor plants | 3,019 | 92.02% | 0.960 |
Top-3 accuracy: 99.75%.
Limitations
- Most training images are single leaves on plain backgrounds; expect lower accuracy on cluttered field or household photos (the indoor subset is the best indicator).
- The model always predicts one of its 54 classes, even for unrelated images or unknown diseases; use the confidence score and top-3 output.
- Not a substitute for expert diagnosis.
- Downloads last month
- 28