Fire Classification and Detection Using a CNN-YOLO Hybrid Model for Early Warning Systems

Abstract

Fire detection is essential for preventing and mitigating disaster impacts, requiring rapid and accurate identification to minimize damage and loss of life. Conventional fire detection systems based on sensors often lead to false alarms and struggle with delayed responses. This study proposes a hybrid CNN-YOLO approach, where a CNN is first used to classify images as fire or non-fire, followed by YOLOv8, which detects and localizes fire regions. The CNN model was trained on images collected from open databases, such as Kaggle, ensuring a diverse range of fire scenarios for classification. The CNN model achieved training, validation, and testing accuracies of 99.98%, 94.12%, and 93.01%, respectively, demonstrating its reliability in distinguishing fire from non-fire images. Additionally, the YOLOv8 model showed progressive accuracy improvements throughout 107 epochs, achieving mAP50 and mAP50-95 scores of 0.765 and 0.486, respectively. The hybrid model achieved an overall accuracy of 90%, with precision of 0.96, recall of 0.93, and an F1-score of 0.95, highlighting its practical effectiveness in detecting fire across diverse environmental contexts. This proposed hybrid approach demonstrates its potential for enhancing early fire detection systems, contributing to disaster management and mitigation efforts. Future research will focus on addressing misclassification issues, refining the model architecture, and expanding the dataset to improve robustness and generalization.