Automated Early Detection of Wildfires Using GOES-16 Satellite Imagery

Wildfires burn millions of acres over the USA each year, generating billions of dollars of property losses, and taking many lives. Their destruction can be greatly reduced the earlier they are detected, as wildfires tend to spread faster the larger they grow. Wildfires are typically detected through 911 calls and observations from fire lookout towers. Unfortunately, these methods tend to be less effective at night and fires in underpopulated areas are often missed or detected quite late. Satellite data provides an alternative means of detecting wildfires. A few ongoing projects monitor burned area using infrared and visible satellite bands, but do not focus on early detection of fires. There are operational fire detection projects which issue warnings when fires are manually detected through satellite imagery. However, there is still no operationalized, automated algorithm optimized for early wildfire detection from satellite data.

Marshaling a unique satellite data platform and expertise in machine learning, Descartes Labs is developing a near-real time fire detection algorithm. The company's cloud-based platform ingests and processes images from the geostationary GOES-16 weather satellite within 5 minutes of the images being taken. Anomalies are detected by comparing the observed reflectance in the infrared bands, which are particularly sensitive to fire, to that of the predicted background state. The background state is constructed using data from surrounding pixels in both space and time, such that heterogeneity of landscape and the diurnal cycle are accounted for. Potential wildfire events are tracked to ensure uniqueness of fire starts. Validation of the algorithm over a scene in Colorado and northern New Mexico using spring/summer 2018 fire data will be presented. Challenges in working with historical wildfire data, sources of false positives, and possible further optimizations will be discussed.