A Multimodal Deep Learning Approach to Forest Fire Early Detection Systems

Abstract

Wildfires in which smoke and flames are detected late cause widespread ecological damage, high carbon emissions, and significant economic losses worldwide. In recent years, driven by the effects of climate change, wildfires have become increasingly frequent natural disasters. Therefore, early wildfire detection is of critical importance for environmental sustainability and disaster management. In this study, a prototype multimodal deep learning-based early wildfire detection system that integrates image processing and environmental sensor data is developed. The proposed architecture is built upon the YOLOv8 object detection model and enhanced with SPD-Conv and GSConv modules. A total of seven different model configurations M1–M7 were evaluated. The real-world behavior of the system was assessed in a video-based simulation environment, where fire and smoke detections were analyzed together with simulated sensor data using a decision-level fusion approach. Experimental results demonstrate that the use of a cleaned dataset and a fine-tuning strategy significantly improves model performance. The final model (M6) achieved a precision of 0.84640, an mAP50 of 0.72887, and an mAP50-95 of 0.32988, making it the best-performing model in terms of precision and mAP50. The findings indicate that the proposed multimodal data fusion approach, which combines image-based detection with sensor simulation, has the potential to improve the reliability of decision-making in early wildfire detection systems.