Smoke Detection

Machine Learning Smoke Detection

Early detection of wildfires is a critical aspect to FUEGO. One method of early detection in development is a machine learning smoke detection software. Using a network of more than 200 fire tower cameras, located throughout the western United States, we hope to autonomously detect and locate fires early. We have ~95% accuracy with our current iteration of the system. Further refinement of the current system will improve the accuracy of the system and reduce the time before first detection.

Our end goal is to integrate an automated smoke detection system into the larger FUEGO system. Smoke detection software will inform FUEGO of the location of a potential fire and the system will respond by focusing attention of other detectors onto the potential location. These other detectors could be PTZ cameras, aircraft, or satellite imagery. Once a fire is confirmed, the relevant authorities will be alerted. This will allow fire fighters to put out fires while they are still small and easier to control.

Current System

Sample image showing the detection software at work.

We currently pull live fire tower images from over 200 cameras and plan to expand the camera network further. This network of cameras is geographically diverse, including cameras in Oregon, Nevada, northern, and southern California. These cameras are run by several different organizations including (but not limited to) theĀ University of California San Diego, ALERTWildfire, and the United States Geological Survey.

We break each image into overlapping square segments and feed each segment into a our machine learning detection. The machine learning network, built using Tensorflow and our library of smoke and non-smoke images, then returns a smoke score for each segment. For average image is analyzed in around 2 seconds on our current hardware, although we will be moving to a more powerful computation system soon. Our goal is to analyze images at the rate they are made available.

Next Steps

Going forward we plan to improving its accuracy by iteratively retraining it. Using the current system, we will collect images of false positives/negatives that we can use to improve our image library.