Scalable Automated Video Labeling for Early Wildfire Smoke Detection with Fast-Then-Precise Two-Stage Inference

Srikantnag Angondalli Nagaraja , Imre Bartos, Chang Zhao

Early wildfire response depends on detecting the first faint appearance of smoke while maintaining low false-alarm rates across diverse cameras, lighting conditions, and environments. A central barrier to progress is the lack of scalable, reliable supervision for subtle early-stage smoke plumes, which makes models brittle under real-world domain shift. We address this challenge by introducing a scalable automated video labeling pipeline based on SAM2 mask propagation, including a reverse-frame processing strategy that runs the model on temporally reversed video sequences. This enables faster and more consistent annotation of early smoke emergence by propagating masks backward from clearly visible smoke to its faint initial appearance. Segmentation masks are converted into tight bounding boxes with targeted human validation to remove cloud artifacts, producing a large and diverse training set spanning fixed-view and zoom-capable wildfire camera networks. Building on this dataset, we design a fast-then-precise two-stage smoke detection system that mirrors operational alerting logic. A high-recall early-warning stage based on RT-DETR prioritizes rapid detection, while a high-precision confirmation stage using YOLOv11 stabilizes alerts and suppresses false positives. The system is evaluated under two complementary settings. On a strict temporally held-out benchmark consisting of all available ignition sequences in the HPWREN Fire Ignition images Library (FIgLib) from 2023 to 2025 (excluded from training), the early-warning stage detects smoke in 7.0 (+-) 6.3 minutes on average. Separately, on the overall test dataset, the early-warning stage achieves high recall (0.94), while the confirmation stage reaches high precision (0.95), with no false positives observed across the full 2023–2025 FIgLib evaluation sequences. These results demonstrate that scalable video labeling enhanced by reverse temporal propagation and complementary two-stage inference enable reliable early wildfire smoke detection under realistic operating conditions.