Transfer Learning Enhanced Deep Learning Model for Wildfire Flame and Smoke Detection

Autonomous UAVs equipped with deep learning and RGB cameras are becoming pivotal for wildfire monitoring. This study highlights the importance of Transfer Learning (TL) to enhance the YOLOv5 convolutional neural network (CNN) model's performance for detecting wildfire smoke and flames, particularly when trained on limited datasets. We introduce the Aerial Fire and Smoke Essential (AFSE) dataset as the target dataset, utilizing the Flame and Smoke Detection Dataset (FASDD) and the general Microsoft Common Objects in Context (COCO) dataset as source datasets. TL is meticulously applied in two phases: feature extraction and fine-tuning, leading to significant improvements in detection precision (reaching up to 83.0% for the mean Average Precision (mAP@0.5)), reduced training time, and increased model generalizability on the AFSE dataset. The results confirm TL's pivotal role in optimizing CNN models for complex tasks like wildfire detection, demonstrating its potential to significantly enhance model accuracy and efficiency.