Arko007/nfnet-f1-plant-disease

NFNet-F1 – Multi-Crop Plant Disease Classification

This repository contains a high-accuracy CNN model for multi-crop plant (crop) disease classification using leaf images.
The model is trained from scratch using NFNet-F1 and evaluated under a strict, non-leaky validation protocol.

---

🧠 Task Description

Task: Image-based plant / crop disease classification
Input: RGB leaf images
Output: One of 88 crop–disease classes, e.g.:

• Tomato__late_blight
• Rice__hispa
• Wheat__yellow_rust
• Apple__black_rot

This formulation is commonly referred to as crop disease detection in agricultural computer vision literature.

---

📊 Dataset

• Source: Merged multi-crop plant disease dataset
• Total images: ~79,000
• Number of classes: 88
• Crops included: Tomato, Rice, Wheat, Apple, Grape, Tea, Corn, Soybean, Potato, and more
• Split:
  • Train: ~71k images
  • Validation: ~8k images

The dataset contains real-world variation (lighting, background, leaf morphology) and class imbalance, making it significantly more challenging than lab-style benchmarks (e.g. PlantVillage).

---

🏗️ Model Architecture

• Backbone: NFNet-F1
• Parameters: ~130M
• Training: From scratch (no ImageNet pretraining)
• Precision: bfloat16
• Optimizer: AdamW (fused)
• Stabilization: Adaptive Gradient Clipping (AGC)
• Scheduler: Linear warmup + cosine decay
• Input resolution: 512 × 512

---

✅ Performance (Validation)

Best checkpoint: Epoch 25

Metric	Value
Validation Accuracy	95.40%
Weighted F1	95.38%
Macro F1	~88.6%

These results are achieved on a strict validation split with no data leakage and macro-F1 reporting, providing a realistic estimate of real-world performance.

---

📁 Repository Files

• model.safetensors
  Secure, production-ready model weights (preferred format)

• classification_report.json
  Full per-class precision, recall, and F1-score

• metrics.json
  Summary validation metrics for the best checkpoint

• config.json
  Model configuration, class names, and preprocessing details

---

🔐 Why Accuracy Is Below “99%+” Benchmarks

Some papers report 99%+ accuracy on PlantVillage, which uses controlled lab images with uniform backgrounds and fewer classes.

This model is evaluated on a much harder setting:

• Multi-crop (88 classes)
• Real-world images
• Strong class imbalance
• Macro-F1 reporting

Under these conditions, ~95% accuracy represents state-of-the-art performance.

---

🚀 Intended Use

• Crop disease diagnosis support systems
• Agricultural decision-support tools
• Research and benchmarking
• Backend inference APIs

⚠️ This model is not a medical or agronomic diagnostic tool and should be used as an assistive system only.

---

🔧 How to Use the Model (Inference)

import torch
import timm
from PIL import Image
from torchvision import transforms
from safetensors.torch import load_file
import json

# -------------------------
# Load config
# -------------------------
with open("config.json", "r") as f:
    config = json.load(f)

class_names = config["class_names"]
img_size = config["input_size"]

# -------------------------
# Build model
# -------------------------
model = timm.create_model(
    config["architecture"],
    pretrained=False,
    num_classes=len(class_names)
)

# Load safetensors weights
state_dict = load_file("model.safetensors")
model.load_state_dict(state_dict)
model.eval()

# -------------------------
# Preprocessing
# -------------------------
mean = config["normalization"]["mean"]
std = config["normalization"]["std"]

transform = transforms.Compose([
    transforms.Resize((img_size, img_size)),
    transforms.ToTensor(),
    transforms.Normalize(mean=mean, std=std),
])

# -------------------------
# Load image
# -------------------------
image = Image.open("leaf.jpg").convert("RGB")
x = transform(image).unsqueeze(0)

# -------------------------
# Inference
# -------------------------
with torch.no_grad():
    logits = model(x)
    pred_idx = logits.argmax(dim=1).item()

prediction = class_names[pred_idx]
print("Predicted class:", prediction)

---

🧾 License

This project follows the dataset’s original license: CC BY-NC-SA 4.0
Commercial usage may require additional permissions.

---

✍️ Author

Arko007