A deep learning based approach for semantic segmentation of small fires from UAV imagery

In this paper, we propose novel techniques for fire segmentation from unmanned aerial vehicle (UAV) images. (1) We propose the ObjectDetection+CIELAB thresholding technique, which leverages a pre-trained object detector such as YOLO (``you only look once’‘) to generate fire bounding boxes. We then apply thresholding in the CIELAB colour space within these regions to detect fire pixels. This approach significantly improves speed by streamlining the segmentation task into a more focused detection and classification task. (2) We further introduce the SEG-4CHANNEL technique, which generates a fire pixel mask using the ObjectDetection+CIELAB thresholding method. This mask is integrated as a fourth channel into various segmentation networks, allowing models to concentrate on fire detection while minimising background interference. (3) Finally, we explore the AttentionSeg technique, which incorporates an attention module within a segmentation framework (e.g., SegFormer-B5) that utilises all four channels. It combines the advantages of CIELAB colour-space model, convolution neural network (CNN) and transformer. We explore a large design space of various networks and backbones on the FLAME (Fire Luminosity Airborne-based Machine learning Evaluation) segmentation dataset. Our best model, AttentionSeg-B5, segments fire with an intersection-over-union (IoU) score of 84.15% and 91.39% F1-score. The code has been released at https://github.com/CandleLabAI/FireSegmentation.