<![CDATA[Fire detection using image-based computer vision is a promising alternative to conventional sensor systems, which are often limited by detection range and false alarms. This study presents a comparative evaluation of two lightweight deep learning architectures, EfficientNet-B0 and MobileNetV3 Small, for multi-class fire image classification across four categories: Urban Fire, Wild Fire, Non-Damage Building, and Non-Damage Wildlife. Both models were trained using transfer learning with ImageNet pre-trained weights on a curated dataset of 1,933 images with selective undersampling to address class imbalance. Grad-CAM (Gradient-weighted Class Activation Mapping) was applied to interpret model decisions and validate the learned visual features. EfficientNet-B0 achieved the highest overall accuracy of 93.8% with a macro F1-score of 0.925, while MobileNetV3 Small reached 91.0% validation accuracy with a macro F1-score of 0.900. Grad-CAM visualizations confirmed that EfficientNet-B0 develops more localized feature activations focused on flame-colored regions, while MobileNetV3 exhibited broader contextual activations. The results demonstrate that lightweight transfer learning models combined with Grad-CAM provide both high classification performance and meaningful interpretability for fire detection applications.]]>
Copyrights © 2026
