Technology Overview
Researchers at the University of Nevada, Reno have developed an early wildfire smoke detection software that uses AI to analyze real-time video from remote camera networks. The system, named Nemo, uses an end-to-end Transformer architecture (based on DETR) to detect small, low-density smoke plumes—often visible only as faint wisps on the horizon—within minutes of fire ignition. By capturing long-range pixel dependencies across high-resolution imagery, Nemo delivers faster, more sensitive smoke detection compared to existing deep learning techniques. In validation tests using 95 real wildfire video sequences, the software achieved a 97.9% detection rates for fires in their earliest incipient stage, with an average detection latency of just 3.6 minutes.
Figure 1: Nemo Wildfire detection
Further Details:
A. Yazdi, H. Qin, C. Jordan, L. Yang, and F. Yan, “Nemo: An Open-Source Transformer-Supercharged Benchmark for Fine-Grained Wildfire Smoke Detection,” Remote Sensing, vol. 14, no. 16, 2022. DOI: https://doi.org/10.3390/rs14163979
Benefits
- Fast Detection: Detects wildfire smoke in as little as 3.6 minutes after fire ignition.
- Early-Stage Focus: Trained to detect faint smoke plumes rather than visible flames.
- High Accuracy: 97.9% success rate detecting fires during early incipient stage.
- Fine-Grained Analysis: Classifies smoke density to insights into severity.
Applications
The wildfire smoke detection tool, including the ML model, training data, and methodology, is publicly available under Apache License 2.0 on GitHub. However, detailed hyperparameter tuning for the final model is proprietary. An improved methodology is currently under development, and we welcome for collaboration.
