Umeå University's logo

Fire is considered one of the major threats to life, property, ecosystems, global warming, and the economy. Recent advancements in convolution neural networks have shown potential for vision-based fire detection; however, several challenges are associated with these techniques, such as limited model performance and high computational complexity. To address these issues, we present an efficient CNN-based model in which EfficientNetV2B0 is employed as a backbone feature extractor and is integrated with a channel and modified spatial attention mechanism to extract deeper spatial details, thereby weighting important features appropriately. The spatial attention is modified by introducing two depth-separable convolution layers to control computational complexity without affecting the performance. The proposed model is assessed on four benchmark datasets in the domain of remote sensing and CCTV-based systems for effective fire detection. Experimental analysis reveals that our model outperforms existing methods in terms of higher accuracy and inference speed, with lower model size and computational burden, indicating its suitability for deployment on resource-constrained devices in real time. To explain the predictions made by the proposed model, we use explainable artificial intelligence methods called Grad-CAM, guided backpropagation, and guided Grad-CAM to provide visualizations by localizing the most salient regions in the image, as emphasized by the attention mechanism.