An Improved MobileNetV3 for Classifying Potential Wildfire Types to Distinguish Smoke-Producing Events from Smoke-Free Instances

The recent, dramatic intensification of wildfires poses severe global threats, making rapid and accurate smoke detection essential for effective disaster response. To address this, our research presents an enhanced MobileNetV3 smoke classification model. Leveraging a comprehensive Kaggle dataset from Quang Vinh Vü (with varied smoke and non-smoke samples), we replaced the SE module and innovatively integrated cross-entropy and focal loss functions for this model. This approach significantly improves accuracy for challenging minority-class and complex smoke instances and ensures stable training. Critically, experimental results demonstrate the enhanced model's superior accuracy and computational efficiency, highlighting its suitability for real-time deployment in early wildfire intervention systems.