Robust Fire Classification Across Multiple Datasets using Self-Attention Based Convolutional Neural Network

Climate change, urbanization, and industrialization have all contributed to an increase in fire-related incidents, which has raised the need for efficient fire detection systems. Traditional sensor-based methods have limitations in complex scenarios, whereas computer vision and deep learning offer feasible alternatives. The study introduces a fire classification system utilizing Convolutional Neural Network (CNN) enhanced with self-attention techniques to boost classification accuracy in this field. The system was trained on different fire image datasets, addressing critical challenges such as data imbalance and real-time performance on limited resources. Comprehensive studies demonstrate that the proposed model substantially surpasses established transfer learning models, including VGG16, DenseNet, and InceptionV3. Our proposed model with augmentation and self-attention mechanism attained an accuracy of 95% on the FireNet dataset, 95% on the Fire_Dataset, 92.43% on the Firesmoke dataset, and 98.91% on the Forest_Fire_BigData dataset. The integration of augmentation techniques with self-attention processes resulted in the suggested model demonstrating improved precision, recall, and F1-scores across many datasets, making it appropriate for applications such as intelligent surveillance and forest monitoring.