Shaq2/chashibhai-rice-disease-cls

ChashiBhAI Rice Disease Classifier

Author	Shaq2 (Shakil Ahmed)
Crop	rice (ধান)
Task	Image classification (leaf disease)
Architecture	YOLO26-cls → TFLite FP16
Input	[1, 640, 640, 3] NHWC RGB /255.0
Output	[1, 8] softmax probabilities (nms: false)
Status	production-demo
App	ChashiBhAI (Expo / React Native, on-device diagnosis)
Code	GitHub
Collection	ChashiBhAI on-device classifiers

Disease ID in ChashiBhAI always runs on-device. Gemini / KrishokChat generate advisory text only and never see the photo.

Files

File	Role
model.tflite	On-device graph used by the Android app (~3.12 MB)
labels.json	Canonical class names and preprocess contract
best.pt	Ultralytics source weights (export parent)

Classes (8)

Label	English	Bangla
Rice__Bacterial_Leaf_Blight	Bacterial Leaf Blight	ব্যাকটেরিয়াজনিত পাতা পোড়া
Rice__Brown_Spot	Brown Spot	বাদামী দাগ
Rice__Healthy_Leaf	Healthy Leaf	সুস্থ পাতা
Rice__Leaf_Blast	Leaf Blast	ব্লাস্ট
Rice__Leaf_Scald	Leaf Scald	পাতা পোড়া
Rice__Narrow_Brown_Leaf_Spot	Narrow Brown Leaf Spot	সরু বাদামী দাগ
Rice__Rice_Hispa	Rice Hispa	হিসপা
Rice__Sheath_Blight	Sheath Blight	শেথ ব্লাইট

Preprocessing (variant C) — required

Do not letterbox. Letterbox disagreed with the .pt on non-square photos.

1. Resize the shortest side to imgsz = 640, keep aspect ratio.
2. Centre-crop to 640×640.
3. RGB, NHWC, float32 / 255.0.

labels.json is the source of truth (preprocess: centercrop).

Measured export checks

Check	Result
Preprocess verified vs .pt	yes (variant C, 100% top-1 vs .pt on the rice hold-out used for export)
FP16 vs FP32 top-1 agreement	1.0
FP16 vs FP32 max softmax diff	0.000372171
Independent machine test	pass (August 2026)
Bundled in APK	yes (rice)

Validation

Independent test on a separate machine (August 2026) confirmed this TFLite export loads and classifies correctly with the variant-C contract in labels.json. Rice, brassica, and corn all passed the same check.

Intended use

• On-device diagnosis in ChashiBhAI for Bangladeshi farmers (Bangla-first UI).
• Research reproduction of the mobile export.

Out of scope: detection / bounding boxes, crop auto-routing, chemical dosage (handled by a separate advisory stack with a refuse gate).

Limitations

Demo-grade on-device classifier. Not a plant-pathologist substitute. Retake if confidence is low or the leaf is not centred.

Load (Python)

import json
from pathlib import Path
import numpy as np
from PIL import Image
import tensorflow as tf

def preprocess_centercrop(path: str, imgsz: int = 640) -> np.ndarray:
    im = Image.open(path).convert("RGB")
    w, h = im.size
    scale = imgsz / min(w, h)
    nw, nh = int(round(w * scale)), int(round(h * scale))
    im = im.resize((nw, nh), Image.BILINEAR)
    left, top = (nw - imgsz) // 2, (nh - imgsz) // 2
    im = im.crop((left, top, left + imgsz, top + imgsz))
    return (np.asarray(im, dtype=np.float32) / 255.0)[None, ...]

labels = json.loads(Path("labels.json").read_text(encoding="utf-8"))
it = tf.lite.Interpreter(model_path="model.tflite")
it.allocate_tensors()
inp, out = it.get_input_details()[0], it.get_output_details()[0]
it.set_tensor(inp["index"], preprocess_centercrop("leaf.jpg", labels["imgsz"]))
it.invoke()
p = it.get_tensor(out["index"])[0]
i = int(p.argmax())
print(labels["names"][i], float(p[i]))

Related models (same author)

• Rice
• Brassica
• Corn
• Suite index: chashibhai-disease-classifiers

Credit (not this model)

KrishokChat Bengali advisory LLM / RAG is not this classifier. See RaiyanKhaan/KrishokChat-Advisory-System and arXiv:2606.29243 (Reza & Shahid).

Citation

@software{ahmed2026chashibhai_rice_cls,
  author = {Ahmed, Shakil},
  title  = {ChashiBhAI Rice Disease Classifier},
  year   = {2026},
  url    = {https://huggingface.co/Shaq2/chashibhai-rice-disease-cls}
}

License

MIT (see LICENSE). Ultralytics remains under its own license. This pack redistributes weights, not training images.