Efficient Outdoor Fire Detection Using MobileNetV2 for Skewed Dataset

Outdoor fire detection is key to the control of fire hazards and the protection of the ecosystems. In this study, we propose an efficient hybrid framework for outdoor fire detection, combining a CNN and MobileNetV2 which can tackle problems raised by skewed datasets and real-world complexity. To improve the detection of minority classes with computational efficiency, advanced data augmentation techniques, cost-sensitive learning, and fine-tuned MobileNetV2 architecture are integrated into the proposed model. We perform extensive experimentation, showing that the model achieves perfect recall on fire images, and an Fl-score of 98.8 % overall, demonstrating the robustness of the model. The framework has the guarantee of zero false negatives for fire detection and is therefore highly suited for safety-critical applications. The proposed model based on lightweight design and high accuracy is suitable for real-time deployment on resource-constrained environments like remote monitoring systems and edge devices. Future work will rely on multimodal data, increase explainability, and further minimize false positives to make the system more practical in different environments.