Jamee Crop Disease Classifier

Image classifier that recognises 15 crop / disease classes (Pepper bell, Potato, Tomato) from a photo of a single leaf. It is the ML component of Jamee, an AI crop-disease detection system. It runs on CPU with ONNX Runtime; no GPU or PyTorch is needed.

Metrics (held-out test set, evaluated once)

20624 images after cleaning, split 14434 / 3095 / 3095 (train / validation / test; 70/15/15, stratified, duplicate-aware, made before augmentation). Best validation macro-F1: 99.67% (epoch 25).

Per-class test results

Use it from Python

Install the runtime (tested on Python 3.12; onnxruntime 1.30 does not install on Python 3.10):

pip install huggingface_hub onnxruntime==1.30.0 numpy==2.4.4 pillow==12.3.0

import sys
from huggingface_hub import snapshot_download

MODEL_DIR = snapshot_download("sarang-sketch/sarox-plangton", local_dir="jamee_model")  # ~17 MB; unchanged files are not downloaded again
sys.path.insert(0, MODEL_DIR)
from inference import load_model, predict_image

model = load_model(MODEL_DIR)                    # load once, reuse for every image
result = predict_image("leaf.jpg", model=model)  # path, bytes or PIL image
print(result["status"], result.get("predicted_class"), result["confidence"])

For production, pin an exact model version with snapshot_download(..., revision="<commit id>"). If this repository is made private, set the HF_TOKEN environment variable on the machine that downloads it.

Successful prediction:

{"status": "success", "predicted_class": "Tomato_Early_blight", "display_name": "Tomato — Early blight",
 "crop": "Tomato", "condition": "Early blight", "is_healthy": false, "confidence": 0.995, "threshold": 0.99,
 "top_predictions": [{"class": "Tomato_Early_blight", "display_name": "Tomato — Early blight", "confidence": 0.995},
                     {"class": "Tomato_Late_blight", "display_name": "Tomato — Late blight", "confidence": 0.003},
                     {"class": "Tomato_Septoria_leaf_spot", "display_name": "Tomato — Septoria leaf spot", "confidence": 0.001}]}

Low confidence (top_predictions is still included for logging — do not present it as a diagnosis):

{"status": "uncertain", "predicted_class": null, "confidence": 0.42, "threshold": 0.99,
 "message": "Jamee could not confidently identify this image.", "top_predictions": ["..."]}

(The two JSON examples show the response format; their values are illustrative.)

Optional: FastAPI wrapper for the Jamee backend

# Run the download step above first, so MODEL_DIR is on sys.path.
from fastapi import FastAPI, File, HTTPException, UploadFile
from inference import InvalidImageError, load_model, predict_image

app = FastAPI()
model = load_model(MODEL_DIR)

@app.post("/predict")
async def predict(file: UploadFile = File(...)):
    try:
        return predict_image(await file.read(), model=model)
    except InvalidImageError as exc:
        raise HTTPException(status_code=400, detail=str(exc))

pip install fastapi uvicorn python-multipart, then uvicorn api:app --host 127.0.0.1 --port 8000. Add authentication and an upload-size limit before exposing it to the internet.

Try the demo locally

In the downloaded folder:

pip install gradio==6.29.1 -r requirements.txt
python app.py

The Gradio page is served at http://127.0.0.1:7860, reachable only from your own computer.

Configuration

• config.json — confidence_threshold, temperature, preprocessing, class list (original + display names).
• Environment variable JAMEE_CONFIDENCE_THRESHOLD overrides the threshold without editing files.

Limitations

• The model only recognises the diseases represented in its training dataset (15 classes: Pepper bell, Potato, Tomato). Any other crop or disease is forced into one of these classes or reported as uncertain.
• A high confidence score does not guarantee real-world correctness.
• Images taken in uncontrolled field conditions (cluttered background, several leaves, blur, unusual light) may produce lower performance: PlantVillage photos are lab images of single leaves on plain backgrounds.
• The confidence threshold is a basic rejection mechanism, not out-of-distribution detection. It and the temperature were fitted on PlantVillage validation images, so they may be optimistic for field photos.

Responsible use

The model returns a class label and probabilities only. It does not recommend pesticides, doses or treatments. Agricultural guidance must come from Jamee's separate knowledge layer and should be reviewed by experts.

Security

inference.py rejects files it cannot decode (InvalidImageError) and images larger than 40 megapixels. Any service that exposes the model on the internet still needs its own authentication, upload-size limit and rate limiting.

Files