rice-disease-net

DINOv2-large fine-tuned to classify 15 paddy disease and stress conditions from field photographs. Trained on a deduplicated multi-source dataset of 9,376 images spanning Tamil Nadu, Bangladesh, and laboratory conditions.

Test accuracy: 92.96% · Weighted F1: 0.9288 · 15 classes · 304M parameters

Model Details

Property	Value
Backbone	facebook/dinov2-large (304M params)
Head	Linear(1024→512, GELU) → Dropout(0.3) → Linear(512→15)
Training	Linear probe (5 epochs) → Full fine-tune (24 epochs, early stopping)
Optimizer	AdamW, head LR 1e-4, backbone LR 1e-5
Schedule	Cosine decay with 5-epoch linear warmup
Hardware	2× RTX 3090, BF16 mixed precision, accelerate
Input size	224×224 RGB
Normalization	DINOv2 standard (mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225])

Input / Output

Input

A single RGB paddy leaf or plant photograph. The model works best on:

• Clear images of individual leaves or panicles
• Field or laboratory lighting (not heavily shadowed)
• Paddy/rice plants (Oryza sativa) only — not validated on other crops

from PIL import Image
from transformers import AutoImageProcessor

processor = AutoImageProcessor.from_pretrained("harinpurumandla/rice-disease-net")
image = Image.open("paddy_leaf.jpg").convert("RGB")

# Returns dict with "pixel_values" tensor of shape (1, 3, 224, 224)
inputs = processor(images=image, return_tensors="pt")

Output

Raw logits tensor of shape (batch_size, 15). Higher logit = higher confidence for that class. Apply softmax for probabilities.

import torch, json

config = json.load(open("config.json"))
idx_to_class = config["idx_to_class"]

with torch.no_grad():
    logits = model(inputs["pixel_values"])   # (1, 15)
    probs = torch.softmax(logits, dim=1)     # (1, 15)
    pred_idx = probs.argmax(dim=1).item()
    confidence = probs[0, pred_idx].item()

print(f"Predicted: {idx_to_class[str(pred_idx)]}  ({confidence:.1%} confidence)")
# Example: "Predicted: blast  (87.3% confidence)"

Full inference example

import json, torch
from PIL import Image
from transformers import AutoImageProcessor

# --- load once at startup ---
import sys
sys.path.insert(0, "path/to/rice-disease-net")
sys.path.insert(0, "path/to/rice-disease-net/train")
from train.model import PaddyClassifier

config = json.load(open("config.json"))
processor = AutoImageProcessor.from_pretrained("harinpurumandla/rice-disease-net")

model = PaddyClassifier(num_classes=15, hidden_dim=512, dropout=0.3)
from safetensors.torch import load_file
model.load_state_dict(load_file("model.safetensors"))
model.eval()

# --- per-image inference ---
def predict(image_path: str) -> dict:
    image = Image.open(image_path).convert("RGB")
    pixel_values = processor(images=image, return_tensors="pt")["pixel_values"]
    with torch.no_grad():
        probs = torch.softmax(model(pixel_values), dim=1)[0]
    idx = probs.argmax().item()
    return {
        "class": config["idx_to_class"][str(idx)],
        "confidence": round(probs[idx].item(), 4),
        "all_probs": {config["idx_to_class"][str(i)]: round(p.item(), 4)
                      for i, p in enumerate(probs)},
    }

result = predict("paddy_leaf.jpg")
print(result)
# {'class': 'blast', 'confidence': 0.8731, 'all_probs': {...}}

Classes

Index	Class	Disease / Condition	Causal Agent
0	bacterial_leaf_blight	Bacterial leaf blight	Xanthomonas oryzae pv. oryzae
1	bacterial_leaf_streak	Bacterial leaf streak	Xanthomonas oryzae pv. oryzicola
2	bacterial_panicle_blight	Bacterial panicle blight	Burkholderia glumae
3	blast	Blast (leaf + neck)	Magnaporthe oryzae
4	brown_spot	Brown spot	Bipolaris oryzae
5	downy_mildew	Downy mildew	Sclerophthora macrospora
6	hispa	Rice hispa	Dicladispa armigera
7	leaf_roller	Rice leaf roller	Cnaphalocrocis medinalis
8	leaf_scald	Leaf scald	Monographella albescens
9	sheath_blight	Sheath blight	Rhizoctonia solani
10	stem_rot	Stem rot	Sclerotium oryzae
11	tungro	Tungro	Rice tungro spherical + bacilliform virus
12	yellow_stem_borer	Yellow stem borer	Scirpophaga incertulas
13	normal	Healthy plant	—
14	potassium_deficiency	Potassium deficiency (abiotic)	Nutrient stress

Evaluation — In-Domain Test Set (n=938)

Summary Metrics

Metric	Value
Top-1 Accuracy	92.96%
Weighted F1	0.9288
Weighted Precision	0.9326
Weighted Recall	0.9296
Macro F1	0.9192
Macro Precision	0.9361
Macro Recall	0.9138
Best class F1	1.000 (yellow_stem_borer)
Worst class F1	0.692 (bacterial_leaf_streak, n=17)

Weighted metrics weight each class by its test-set support count. Macro metrics treat all 15 classes equally regardless of support.

Per-Class Results

Class	Precision	Recall	F1	Support
bacterial_leaf_streak	1.000	0.529	0.692	17
downy_mildew	0.727	0.828	0.774	29
blast	0.943	0.839	0.888	137
tungro	0.878	0.952	0.913	83
brown_spot	0.872	0.962	0.915	78
hispa	0.910	0.935	0.922	108
bacterial_panicle_blight	0.917	0.957	0.936	23
bacterial_leaf_blight	0.970	0.925	0.947	106
normal	0.924	0.965	0.944	113
leaf_scald	1.000	0.909	0.952	11
potassium_deficiency	0.952	0.976	0.964	41
leaf_roller	0.973	0.973	0.973	37
sheath_blight	1.000	0.958	0.979	24
stem_rot	0.976	1.000	0.988	41
yellow_stem_borer	1.000	1.000	1.000	90

Limitations and Known Issues

Weak classes:

• bacterial_leaf_streak (F1=0.692, support=17): Only 17 test samples; the model is precise when confident but misses ~47% of actual cases. More field data needed.
• downy_mildew (F1=0.774): Visually similar to early blast and nutrient deficiency; low precision suggests over-prediction.
• blast (F1=0.888, recall=0.839): Leaf blast and neck blast were merged. Some blast images are classified as brown_spot.

Geographic scope: Training data covers Tamil Nadu (Paddy Doctor) and Bangladesh (BRRI). Performance on Telangana, Andhra Pradesh, West Bengal, and Southeast Asian varieties has not been validated. Expect degraded accuracy on field conditions significantly different from the training distribution.

Not a diagnostic tool: Model output should be reviewed by an agronomist before treatment decisions. Abiotic stress (potassium_deficiency) shares visual symptoms with several diseases.

Training Data

Source	Classes Used	Images (kept after dedup)
Paddy Doctor (Kaggle 2022)	All 13 original classes	~6,600
BRRI Kaggle (valid set)	blast, bacterial_leaf_blight, brown_spot, hispa, leaf_scald, sheath_blight, tungro	~1,400
Mendeley Rice Disease V1 (2026)	bacterial_leaf_blight, leaf_roller, stem_rot, potassium_deficiency	~1,400

Total after SHA-256 exact dedup + pHash near-dedup (Hamming ≤ 10): 9,376 images. Split: 80% train / 10% val / 10% test (stratified, frozen test set).

Citations

Backbone:

@misc{oquab2023dinov2,
    title={DINOv2: Learning Robust Visual Features without Supervision},
    author={Maxime Oquab and others},
    year={2023},
    eprint={2304.07193},
    archivePrefix={arXiv}
}

Paddy Doctor dataset:

@misc{paddy-disease-classification,
    author={Paddy Doctor and Pandarasamy Arjunan (Samy) and Petchiammal},
    title={Paddy Doctor: Paddy Disease Classification},
    year={2022},
    howpublished={\url{https://kaggle.com/competitions/paddy-disease-classification}},
    note={Kaggle}
}

Mendeley dataset: Raki, Nishat Sultana; Bakki, Md. Abdul; Sheikh, Foysal; Pria, Mosa. Nadia Sultana; Parvin, Shahnaj; Matin, Mafiul Hasan (2026), "Rice Disease Image Dataset", Mendeley Data, V1, doi: 10.17632/jw7hp6r5gj.1

BRRI dataset: Attributed to Bangladesh Rice Research Institute (https://brri.gov.bd/). No formal citation provided by dataset authors.