Résumé:
In the past years, 40 831 hectares of forests have been devastated by rampant wildfires, resulting in the tragic loss of numerous lives. To address this challenge, our research focuses on developing
an early wildfire detection system capable of identifying potential fire outbreaks before they escalate, as controlling them once they have spread becomes arduous. Our proposed approach
utilizes unmanned aerial vehicles (UAVs) to capture aerial data, which is then processed on board to automatically detect early signs of wildfires. This enables us to promptly alert relevant emergency services and facilitate a rapid response.
The core technique employed in our approach is transfer learning, specifically applied to the YOLOv3 model for object detection through several Batch sizes and epochs. We validate the effectiveness of our model using FLAME dataset. The performance metrics we have achieved demonstrate the success of our approach, with Precision, Recall, F1-Score and Accuracy rates reaching impressive levels of 100%, 96.66667%, 98.305085%, and 96.66667% respectively.
