odiug77/wildfire-smoke-fire

Wildfire Smoke and Flame Detection – YOLOv26

Overview

This model is based on YOLOv26m and has been specifically trained to detect wildfire smoke and flames in outdoor environments.

It is designed for early wildfire detection using standard motorized RGB surveillance cameras positioned at strategic observation points such as hills, towers, or remote monitoring stations.

The system operates by continuously capturing images while the camera rotates across the monitored area. Each captured frame is processed by the model to automatically identify potential fire events.

🎯 Purpose

The primary goal of this model is early wildfire detection, enabling faster response times and reducing environmental and economic damage.

It is optimized for:

• Daytime wildfire detection
• Nighttime wildfire detection
• Smoke recognition
• Flame detection
• Standard RGB camera imagery (no thermal sensors required)

🛰️ Deployment Scenario

Typical workflow:

1. A motorized camera rotates and captures images at regular intervals.
2. Images are sent to the inference system.
3. The model analyzes each frame.
4. If smoke or flames are detected, an alert can be triggered.

🧠 Model Architecture

• Architecture: YOLOv26m (fused)
• Task: Object Detection
• Classes: smoke, fire
• Layers: 132
• Parameters: 20,350,994
• GFLOPs: 67.9
• Input: RGB images

📊 Model Performance

Validation set:

• Images: 100
• Instances: 301

Overall Metrics

Metric	Value
Precision (P)	0.632
Recall (R)	0.509
mAP@0.50	0.539
mAP@0.50:0.95	0.306

Per-Class Metrics

Class	Images	Instances	Precision	Recall	mAP@0.50	mAP@0.50:0.95
smoke	97	125	0.715	0.722	0.748	0.475
fire	34	176	0.550	0.295	0.330	0.138

Inference Speed (per image)

• Preprocess: 0.3 ms
• Inference: 12.7 ms
• Postprocess: 1.3 ms
• Total: ~14.3 ms per image

Approx. ~70 FPS theoretical throughput on Tesla T4 GPU.

This model does not replace professional fire detection systems or human supervision.

🚀 Inference Example (CPU)

Minimal example to run inference on CPU and print detection results.

from ultralytics import YOLO

# Load trained model
model = YOLO("wildfire-smoke-fire.pt")

# Run inference on CPU
results = model("image.jpg", device="cpu")

# Ultralytics returns a list (even for a single image)
r = results[0]

# Print detected boxes
print("Detected boxes:")
print(r.boxes)

# Print class names
print("Class names:")
print(r.names)

Notes

• Install Ultralytics:

pip install ultralytics