Improving Wildfire Detection Accuracy Using MobileNetV3-YOLOv8n

Wildfires cause significant damage to ecosystems and human society, making timely fire detection crucial for minimizing losses. With the advancement of deep learning and Internet of Things (IoT) technologies, intelligent wildfire detection systems show promising application prospects. However, deploying high-precision detection models on resource-constrained IoT devices and drones still faces numerous challenges. To address this issue, this paper integrates the MobileNetV3 lightweight neural network with the YOLOv8 object detection model and proposes two improved models: MobileNetV3 Small-YOLOv8n and MobileNetV3 Large- YOLOv8n. Systematic evaluation on our custom wildfire image dataset demonstrates that the MobileNetV3 Large-YOLOv8n model achieves a flame detection Box(P) of 61.1%, a 4.2 percentage point improvement over the original YOLOv8n. To thoroughly investigate model performance, we designed a series of ablation experiments, comparing different YOLOv8n variants and exploring methods to enhance wildfire feature detection capabilities through the introduction of various feature pyramid structures. This research provides an efficient solution for wildfire detection systems in resource-constrained environments, achieving not only significant improvements in detection accuracy but also maintaining sufficient real-time performance, making important contributions to the advancement of intelligent disaster prevention, mitigation, and emergency response systems.