Authors: Mason Davis
Mentors: Mohammad Shekaramiz
Insitution: Utah Valley University
Forest fires are disasters that devastate our local communities here in Utah and communities abroad. Characterized as uncontrolled and unpredictable fires in areas with combustible vegetation, these phenomena cause ecological and economic harm. With global warming driving temperature increases and variability in weather patterns, these fires are becoming more severe and frequent. Calls from local and national leaders for solutions are ever-growing. An important factor in fighting these fires includes early detection and monitoring. With advances in artificial intelligence and computer vision, the accuracy and speed of detection can be greatly improved. In this research, two new deep learning approaches making use of transfer learning are developed and investigated for fire detection. To compare performance further, existing architectures are also deployed for analysis on the fire detection problem, including ResNet-50, Xception, MobileViT, and Support Vector Machine.
To train and evaluate the performance of the above models, the popular forest fire dataset known as DeepFire was utilized. This dataset consists of a symmetrical split of fire and no-fire images consisting of 1900 total images in varying forest environments. Each architecture was tuned through hyperparameter searches and trails to seek ideal combinations for optimal performance. A comparison was drawn with the most recent literature making use of this dataset. Here, our modified Xception architecture leveraging transfer learning topped all recent publications on the DeepFire dataset by achieving 99.211% accuracy.
With wildfires increasing in frequency and severity, the early detection of these disasters is paramount to controlling their spread. We have seen that deep learning can provide an increasingly accurate way to autonomously survey and detect these disasters. This is a promising step toward autonomous detection and early elimination of these disasters as they start. Future work will include the investigation of real-time processing techniques for fire detection, allowing for real-time data acquisition, inference, and transmission of geo-information to emergency and forest management teams.