YOLOv11s for Early Wildfire Detection

This repository contains a compact object detection model based on YOLOv11s, optimized by Pyronear for early-stage wildfire detection. The model is trained to identify smoke plumes with high precision and minimal false positives using aerial and ground-level imagery collected across diverse environments.

🔍 Model Overview

Architecture: YOLOv11s (small variant)
Task: Single-class object detection (wildfire smoke)
Backends: TorchScript, ONNX (CPU, MPS), NCNN (CPU, MPS)
Resolution: 1024 × 1024
Training epochs: 45
Optimizer: AdamW
Augmentations: RandAugment, mixup, mosaic, erasing, color jitter

📦 Available Formats

Format	Backend	File
NCNN	CPU	`ncnn_cpu_yolo11s-colorful-chameleon_v3.0.0_7bd9f32.tar.gz`
NCNN	MPS	`ncnn_mps_yolo11s-colorful-chameleon_v3.0.0_7bd9f32.tar.gz`
ONNX	CPU	`onnx_cpu_yolo11s-colorful-chameleon_v3.0.0_7bd9f32.tar.gz`
ONNX	MPS	`onnx_mps_yolo11s-colorful-chameleon_v3.0.0_7bd9f32.tar.gz`
PyTorch	Any	`best.pt` (SHA256: `e3667439...`)

The _7bd9f32 suffix refers to the source Git commit used to build this release.

📁 Dataset

Source: Pyronear wildfire imagery repository
Size: ~5.9M images
Structure: YOLO-format bounding boxes with a single fire class
Preprocessing:
- 5% random sampling
- Cleaned, stratified by region
- Full-resolution pipeline with 1024px square crops

🏋️‍♂️ Training Details

Config: best.yaml

Command:

python train.py --data data.yaml --model yolo11s.pt --epochs 45 --imgsz 1024 --batch 16 --optimizer AdamW --pretrained

DVC hash:
- Model: 6ea728f6abf7c8de6eee661e52840893.dir
- Data: 08dc4b6823c84cd24507003aaa4d5cfb.dir

Perfect, thanks for sharing the full Classifier class implementation.

Here is the updated Inference section for your README.md, reflecting the correct usage of Classifier from pyroengine.vision, and linking back to the source code on GitHub:

🔧 Inference

You can run inference using the format that best suits your environment:

✅ PyTorch (.pt) with Ultralytics

from ultralytics import YOLO

model = YOLO("best.pt")  # or path to your model
results = model.predict(source="image.jpg", imgsz=1024, conf=0.25)
results.show()  # Display results with smoke bounding boxes

This method uses the official Ultralytics YOLO library.

⚙️ ONNX and NCNN via Pyro‑Engine

For ONNX or NCNN inference, use the Classifier class from pyroengine.vision, developed by Pyronear.

from PIL import Image
from pyroengine.vision import Classifier

# Load image
img = Image.open("your_image.jpg")

# Load classifier (auto-downloads model from Hugging Face if needed)
clf = Classifier(format="onnx", conf=0.15, model_folder="data")  # or format="ncnn"

# Run inference
preds = clf(img)

# preds is a NumPy array of shape (N, 5): [x1, y1, x2, y2, confidence]
print(preds)

The Classifier automatically handles preprocessing, backend selection (ONNX or NCNN), and postprocessing including confidence thresholding, normalization, and occlusion masking.

📎 Source code on GitHub

📜 License

This model is released under the Apache 2.0 License. Feel free to use, adapt, and redistribute with attribution to Pyronear.

🤝 Citation

If you use this model in your research or product, please cite Pyronear and/or link to https://pyronear.org.