Thermal-Only Fire Detection: Using FLAME 3 UAV Imagery for Real-Time Thermal Fire Detection

Fire detection in time and effective would save an ecosystem and infrastructure from impending devastation and human lives from destruction. Unmanned Aerial Vehicles (UAVs) with several thermal sensors provide the best opportunity for such monitoring, especially in low light or smoke-obscured circumstances. However, real-time processing of such data sets proves to be arduous work. This research investigates applying deep learning in developing a Thermal Transformer Network (TTN), based upon the Vision Transformer (ViT) architecture, for automated fire detection using thermal imagery from the FLAME 3 UAV dataset. In this article, we describe how to adapt the ViT for single-channel thermal data, including image preprocessing, patch embedding, and transformer encoder mechanisms. Model training and evaluation are done using a subset from the FLAME 3 dataset composed of thermal TIFF images during controlled burns. The performance of the model is measured by several conventional classification measures that include accuracy, loss, and confusion matrix analysis. So this research also testified about the scope of transformer-based architecture for great feature extraction and classification from thermal detection fire tasks, making further milestones for optimized UAV-based monitoring systems.