Wild Fire Classification using Learning Robust Visual Features

Abstract The diversity in the texture, color, and shape of flames, and their similarity to most sunset skies and red forests, has always made reducing false alarms in forest fire detection a challenging task. Traditional image processing methods rely too heavily on manual feature annotation. Using conventional machine learning models can effectively extract flame characteristics and reduce the error rate of manual annotation, but their limited local perception often leads to a high rate of false alarms. This paper proposes the use of the self-supervised model with Learning Robust Visual Features (DINOV2) for feature extraction, and a classification head for categorizing forest fires. In tests across most fire datasets, DINOV2 achieved an accuracy rate of up to 99% while simultaneously reducing the rate of false alarms.