Real-time wildfire instance segmentation from self-supervised training labels

Abstract

Tracking wildfires in real time is beneficial for fire management response times, studying the environmental impacts of wildfires, and improving single- or multi-sensor autonomy, among many other benefits. Supervised machine learning algorithms enable deep learning models to learn and automatically detect patterns in images efficiently and adaptively. Advances in hardware and associated on-device capabilities are at a stage that makes onboard detection for satellite- and airborne instruments possible. However, supervised methods typically require large hand-labeled datasets. To address this, we train the onboard-capable version of the YOLOv11 instance segmentation architecture with airborne infrared imagery of wildfires using training labels obtained from a separate self-supervised learning framework. Using this newly trained, onboard-capable model, we successfully detect fire sources with a structural similarity index of 0.911 and an intersection over union of 0.796 relative to labels generated by the self-supervised model. The results give 0.850 precision and 0.928 recall in the test dataset, demonstrating the feasibility of self-supervised machine learning for creating training labels. The performance of the self-supervised-to-supervised transfer is evaluated on an emulator of an onboard processor, which found the inference speed to range from 18.6 ms to 50.7 ms with a median of 30.9 ms per 160 by 160 pixel input image tile.