Abstract
One of the most expensive and fatal natural disasters in the world is forest fires. For this reason, early discovery of forest fires helps minimize mortality and harm to ecosystems and forest life. The present research enriches the body of knowledge by evaluating the effectiveness of an efficient wildfire and smoke detection solution implementing ensembles of multiple convolutional neural network architectures tackling two different computer vision tasks in a stage format. The proposed architecture combines the YOLO architecture with two weights with a voting ensemble CNN architecture. The pipeline works in two stages. If the CNN detects the existence of abnormality in the frame, then the YOLO architecture localizes the smoke or fire. The addressed tasks are classification and detection in the presented method. The obtained model's weights achieve very decent results during training and testing. The classification model achieves a 0.95 F1-score, 0.99 accuracy, and 0.98e sensitivity. The model uses a transfer learning strategy for the classification task. The evaluation of the detector model reveals strong results by achieving a 0.85 mean average precision with 0.5 threshold (mAP@0.5) score for the smoke detection model and 0.76 mAP for the combined model. The smoke detection model also achieves a 0.93 F1-score. Overall, the presented deep learning pipeline shows some important experimental results with potential implementation capabilities despite some issues encountered during training, such as the lack of good-quality real-world unmanned aerial vehicle (UAV)-captured fire and smoke images.