— license: mit

Fire Detection Model

By Roy An Ocean Senior, University of Washington

Scinetific context

Wild fire are unplanned, uncontrolled and unpredictabe fire that does tremendous dama to wildlife, human communities.Furthermore, a 2016 study found climate change enhanced the drying of organic matter dobuling the number of large fires between 1984 and 2015(Abatzoglou and Williams,2016). This arises the need of an alert system as distingusihing a fire early is the most effective and practial way to control these massive disasters.

Dataset utilization

Number of Images: 9359 combination of aerial and ground images curated from a diverse array of online platforms such as goverment databases All images has been resized by 640 X 640.

Train : 8115 Val: 829 test :415

Classes:

Fire(including smoke images)
Non Fire

Model selection

I used Ultralytics Yolov11-cls because it was the most up to date Yolo at the time. I used classification as for fire detection, you only need to know yes, if there is an fire so fast processing time can help as well.


model = YOLO('yolon.pt')

#path to Yaml
dataset_config = '/content/Dataset/data.yaml'

#Train the model
results = model.train(
    data=dataset_config,
    epochs=80,
    batch=-1,
    imgsz=640,
    plots=True,
    patience=50,

)

Model Assessment

Confusion Matcix and Normalized confusion matrix. The model had an 96% accuracy and 98% accuracy for fire and non-fire detection

train/loss and val/loss graph over metrics. You can see that loss is gradually decreasing over time indicating that the model is learning effectively. However, Val/loss is slightly higher then train/loss which means that the model might be over fitting meaning it is learning the train data test too much catching too much detail on the image making in not good for doing generalization.

Model usecase

This can obviously be deployed in an area to where wild fire statistically occurs a lot. This model can be the initial alert system where we could set a threshold and if the model thinks a fire is detected, it could send the image or footage of that fire to humans. It can act like an double verification system for further applications. It didn't do well on distingushing on Fogs vs Smoke so adding another class to improve its accuracy in foggy areas might be needed

Disclaimer

Credit for all the images go to https://www.kaggle.com/datasets/sayedgamal99/smoke-fire-detection-yolo/data