minhhungg/rice-disease-dataset

This is the dataset for my project of Plant Diagnosis Suite at here. You can check out to see a Disease Detector trained on this dataset

Vietnamese Rice Disease & Crop Recommendation Dataset

An agricultural AI dataset collected in Vietnam, containing 37,978 rice plant images across 21 classes (diseases, pests, nutrient deficiencies, healthy) — for image classification.

---

Dataset Summary

Component	Type	Samples	Classes	Task
Rice Disease Images	Image (JPG)	37,978	21	Image Classification

Splits Available

Split	Description	Size
healthy	Original healthy rice images	3,764 images
pests	Original pest/insect images (9 classes)	~14,284 images
diseases	Original rice disease images (8 classes)	~17,618 images
nutrition	Original nutrient deficiency images (3 classes)	2,312 images
train	Stratified training split (70%)	26,584 images
validation	Stratified validation split (15%)	5,697 images
test	Stratified test split (15%)	5,697 images

The train / validation / test splits draw from all 21 classes combined and are stratified — use these for model training. The four category splits (healthy, pests, diseases, nutrition) reflect the original folder structure of the raw collection.

---

Dataset Structure

Image Component — Fields

Field	Type	Description
image	Image	The rice plant photograph
label	string	English class label (see label list below)

Loading the Dataset

from datasets import load_dataset

# Load a specific split
ds = load_dataset("minhhungg/rice-disease-dataset", split="train")

# Load all splits
ds = load_dataset("minhhungg/rice-disease-dataset")

# Access train/val/test for model training
train = ds["train"]
val   = ds["validation"]
test  = ds["test"]

# Quick check
print(train[0])
# {'image': <PIL.JpegImagePlugin...>, 'label': 'Healthy'}

---

Label Definitions

Image Classes (21 total)

Healthy (1 class)

Label	Vietnamese Name	Count
Healthy	Cây lúa khỏe mạnh	3,764

Pests / Insects (9 classes)

Label	Vietnamese Name	Approx. Count
Tungro Virus	Tungro virus	3,480
Hispa	Sâu gai	2,922
Rice Gall Midge	Sâu năn (Muỗi hành)	1,582
Chilo Stem Borer	Sâu đục thân (Sọc nâu)	1,490
Rice Leaf Folder	Sâu cuốn lá nhỏ	1,210
Thrips	Bọ trĩ	1,160
Rice Skipper	Sâu cuốn lá lớn	950
Yellow Stem Borer	Sâu đục thân (vàng)	910
Brown Plant Hopper	Rầy nâu	580

Diseases (8 classes)

Label	Vietnamese Name	Approx. Count
Leaf Scald	Bệnh cháy lá	3,340
Sheath Blight	Bệnh đốm vằn / khô vằn	3,156
Brown Spot	Bệnh đốm nâu	3,140
Bacterial Leaf Blight	Bệnh bạc lá	2,950
Narrow Brown Spot	Bệnh gạch nâu	2,832
Blast	Bệnh đạo ôn lá và cổ bông	2,000
Bakanae Disease	Bệnh lúa von (lúa đực)	100
False Smut	Bệnh than vàng	100

Nutrient Deficiencies (3 classes)

Label	Vietnamese Name	Count
Nitrogen Deficiency	Thiếu đạm (N)	880
Potassium Deficiency	Thiếu kali (K)	766
Phosphorus Deficiency	Thiếu lân (P)	666

---

Class Imbalance

The dataset has significant class imbalance across the image component:

Most common : Healthy           — 3,764 images
Least common: Bakanae Disease   —   100 images
Imbalance ratio: 37.6 ×

When training on the combined train split, we recommend using class weights or focal loss to account for this. The stratified split ensures proportional class representation across train/val/test.

---

Image Properties

Based on EDA of 1,050 sampled images (50 per class):

Property	Value
Mean height	1,223 px
Mean width	1,053 px
Most common aspect ratio	1:1 (square)
Height range	217 – 4,301 px
Width range	201 – 4,364 px
Mean file size	282 KB
Mean brightness	149.5 / 255
Color channels	RGB (3)
File format	JPEG (.jpg / .JPG / .jpeg)

Images vary widely in resolution. All models trained on this dataset resized inputs to 224 × 224 using transforms.Resize((224, 224)).

---

Data Collection

• Source: Field photographs collected across multiple sources on the Internet
• Geographic scope: Vietnam (Mekong Delta and surrounding agricultural regions)

Geographic bias: Images were collected exclusively in Vietnam. Generalization to rice-growing regions with substantially different climate, rice varieties, or lighting conditions (e.g., South Asia, East Africa) has not been validated.

---

Dataset Creation

Preprocessing

Images were not resized or augmented in the raw splits — they are served at original resolution. The train split CSV was generated with a stratified 70/15/15 split using sklearn.model_selection.train_test_split(stratify=label) with random_state=42.

Recommended Preprocessing for Training

from torchvision import transforms

train_transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.RandomHorizontalFlip(p=0.5),
    transforms.RandomVerticalFlip(p=0.3),
    transforms.RandomRotation(30),
    transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),
    transforms.RandomAffine(degrees=0, translate=(0.1, 0.1)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225]),  # ImageNet stats
])

val_transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225]),
])

---

Usage Examples

Basic Classification with HuggingFace Trainer

from datasets import load_dataset
from transformers import AutoFeatureExtractor, AutoModelForImageClassification, TrainingArguments, Trainer
import numpy as np

ds = load_dataset("minhhungg/rice-disease-dataset")

# Get label list from train split
labels = sorted(set(ds["train"]["label"]))
label2id = {l: i for i, l in enumerate(labels)}
id2label = {i: l for l, i in label2id.items()}

model_name = "google/efficientnet-b0"
extractor  = AutoFeatureExtractor.from_pretrained(model_name)

def preprocess(batch):
    inputs = extractor(images=batch["image"], return_tensors="pt")
    inputs["labels"] = [label2id[l] for l in batch["label"]]
    return inputs

ds_processed = ds.map(preprocess, batched=True, remove_columns=["image", "label"])

model = AutoModelForImageClassification.from_pretrained(
    model_name,
    num_labels=len(labels),
    id2label=id2label,
    label2id=label2id,
    ignore_mismatched_sizes=True,
)

args = TrainingArguments(
    output_dir="rice-classifier",
    per_device_train_batch_size=32,
    num_train_epochs=20,
    evaluation_strategy="epoch",
    save_strategy="best",
    metric_for_best_model="accuracy",
)

def compute_metrics(eval_pred):
    logits, labels = eval_pred
    preds = np.argmax(logits, axis=1)
    return {"accuracy": (preds == labels).mean()}

trainer = Trainer(
    model=model,
    args=args,
    train_dataset=ds_processed["train"],
    eval_dataset=ds_processed["validation"],
    compute_metrics=compute_metrics,
)
trainer.train()

PyTorch DataLoader (direct)

from datasets import load_dataset
from torch.utils.data import DataLoader
from torchvision import transforms
from PIL import Image

ds = load_dataset("minhhungg/rice-disease-dataset", split="train")
labels = sorted(set(ds["label"]))
label2id = {l: i for i, l in enumerate(labels)}

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

def collate_fn(batch):
    images = [transform(item["image"].convert("RGB")) for item in batch]
    labels_t = [label2id[item["label"]] for item in batch]
    import torch
    return torch.stack(images), torch.tensor(labels_t)

loader = DataLoader(ds, batch_size=32, shuffle=True, collate_fn=collate_fn)
images, labels_t = next(iter(loader))
print(images.shape)  # torch.Size([32, 3, 224, 224])

Filtering to a Single Category

# Work only with disease images
diseases = load_dataset("minhhungg/rice-disease-dataset", split="diseases")

# Or filter the combined train split
from datasets import load_dataset
train = load_dataset("minhhungg/rice-disease-dataset", split="train")
diseases_only = train.filter(lambda x: x["label"] in [
    "Leaf Scald", "Sheath Blight", "Brown Spot",
    "Bacterial Leaf Blight", "Narrow Brown Spot",
    "Blast", "Bakanae Disease", "False Smut"
])

---

Limitations & Biases

• Geographic: Collected exclusively in Vietnam. Pest and disease appearance may vary under different climate zones, rice varieties, or soil types.
• Class imbalance: 37× imbalance between the largest and smallest classes. Models trained without compensation may underperform on Bakanae Disease and False Smut (100 images each).
• Image quality variability: Images range from 201px to 4,364px wide, captured under varying field lighting conditions (overcast, direct sunlight, shade). Resolution normalization is required before training.
• Single crop: Only covers rice (Oryza sativa). Not applicable to other crops without retraining.
• Annotation granularity: Nutrient deficiency labels (N, P, K) are assigned at the nutrient level, not by deficiency severity stage.