On 20 July, a satellite from the German company OroraTech detected a hot spot near Aberdeen, Calif., east of Kings Canyon National Park. It was the first detection of the Rock Fire, which prompted evacuations and road closures and grew to more than 12,000 acres in hot, windy conditions before it was contained by firefighters.The Rock Fire is one of a small but growing number of first detections attributed to OroraTech’s OTC-P1 constellation, part of a new generation of satellites developed specifically to detect and monitor wildfires. The advent of these satellites is coming as wildfires become more frequent and intense due to human-caused climate change. Early detection usually helps by letting fire agencies contain potentially destructive fires before they spread out of control. Long used to spot fires in very remote areas, satellites are poised to play a greater role in early detection, augmenting old-fashioned fire lookouts, ground-based cameras, and citizen calls to 911. OroraTech’s patented thermal-infrared camera is equipped with sensors that detect fire-specific temperature signatures.OroraTechDetecting wildfires from space is not a straightforward proposition, however. Just as human lookouts can fail to spot a new ignition or mistake dust for smoke, the detection algorithms used to identify fires in satellite data can miss small or smoldering fires and may be deceived by sun glints and industrial emissions. Improving the accuracy and timeliness of remote wildfire detections requires specialized sensors as well as robust algorithms, with researchers increasingly looking to machine learning to enhance the latter.Today’s fire satellites build on a nearly half-century legacy. In 1980, Jeff Dozier and Michael Matson were working together at the National Oceanic and Atmospheric Administration when they identified tiny bright spots—gas flares from Middle Eastern oil wells—in an image captured by a radiometer on the NOAA-6 satellite. The following year, Dozier published a mathematical method to identify high-temperature areas in satellite data, which became the basis for most classical detection algorithms.When vegetation burns, a portion of the energy released takes the form of infrared radiation. The burning area releases much more radiation, particularly in the mid-wave band, than does the surrounding area. In satellite imagery, the pixel that contains the burning area registers an increase in “brightness temperature,” a measure of the intensity of electromagnetic energy coming from a source. The first wildfire-detection algorithms used fixed thresholds for brightness temperature data to identify likely fires, although these quickly gave way to “contextual” algorithms that adjust their thresholds based on local conditions—compensating for regional and seasonal differences in radiation. Global databases of known hot spots, like gas flares and steel plants, help filter out false detections.NASA engineers used the experiences with the NOAA-6 satellite to design subsequent instruments used to detect fires from space, including NASA’s Moderate Resolution Imaging Spectroradiometer and Visible Infrared Imaging Radiometer Suite, which are deployed on low Earth orbit satellites and typically pass over a given location several times a day. They are complemented by NOAA’s Geostationary Operational Environmental Satellites (GOES), which provide continuous updates at a coarser resolution. Next-Generation Satellites Built for Early Wildfire DetectionPurpose-built fire satellites like the three launched by SpaceX in July for the California-based nonprofit Earth Fire Alliance (EFA) are further optimized for fire detection. EFA’s “FireSats” have a multiband suite of sensors including two sensitive to mid-wave infrared bands, one of which is attenuated for better detection of extremely hot fires that can saturate other sensors. The suite also includes sensors for visible, near-infrared, and short- and long-wave infrared radiation, allowing fire managers to accurately characterize the entire temperature profile of a fire throughout its life cycle. Michael Falkowski joined Earth Fire Alliance from NASA, where he served as program manager for the agency’s Wildland Fire Program and led its FireSense project.Earth Fire AllianceThe alliance ultimately aims to deploy more than 50 of its low Earth orbit satellites by the 2030s, which will allow it to image any location on the globe at 20-minute intervals. With an average image resolution of 80 meters per pixel compared to 500 meters for NASA’s Visible Infrared Imaging Radiometer Suite, EFA’s satellites will be able to detect much smaller fires than legacy satellites can—down to around 25 square meters, about the size of a standard shipping container.“From space, for decades, we’ve been really blind to where these small fires are, and we’re probably drastically underestimating the total amount of global burned area and the total amount of carbon emissions from fires because we don’t have a handle on where these small fires are,” says Michael Falkowski, EFA’s lead scientist. “One of the things that our system will enable is the ability to catch these small fires, view them more frequently, and not only improve fire operations but also improve global fire science.”EFA is currently working with a few early adopters who are helping refine its data products and delivery mechanisms; it plans to make those products more widely available to fire agencies and scientific researchers starting next year. Its first three operational satellites broadcast the images they acquire to dedicated ground stations, which then upload them into the cloud for processing and delivery to end users—a sequence of events that takes about 20 minutes from start to finish, according to Falkowski. EFA’s next satellites will be able to transmit images via Starlink to any ground node, cutting delivery times even further, he says.How Machine Learning Enhances Orbital Wildfire DetectionBecause EFA wants to distribute FireSat data to anyone with a legitimate need for it, images from its satellites can be used with both classical fire-detection algorithms and novel machine learning approaches, which have become increasingly popular in recent years for complex vision tasks. The alliance already has a partnership with Google Research to support development of new detection algorithms. These will use AI to compare operational FireSat data with historical images of the same location to detect small fires while minimizing false-positive rates. OroraTech says one of its wildfire satellites made the first detection of the Rock Fire in an area not visible to ground fire cameras.OroraTechOroraTech, which has a partnership with EFA to expand access to wildfire data for nongovernmental organizations, is already using AI models in addition to classical detection algorithms. Processing aboard its satellites allows the company to detect fires rapidly, downloading essential details about new detections to ground stations in advance of full images. “It’s important for us to be as fast as possible—we’re talking about minutes,” said Dima Rashkovetsky, OroraTech’s team lead for data engineering. “We know from talking to multiple customers that information after an hour is borderline useless for first responders.” Rashkovetsky says OroraTech developed its fire-detection models using supervised machine learning, which entailed manually labeling fires in the images used to train the model. “In the beginning, we prioritized precision, so we wanted to make sure that if we say something is a fire, it is really a fire, even at the expense of sometimes missing some of the smaller fires,” he said. The company iterated from there until it could reliably detect more of the smaller fires, while also compressing its machine learning model enough to run on the Nvidia Jetson Xavier NX GPU modules installed in its satellites.Compared to classical detection algorithms, Rashkovetsky says, AI makes it easier to incorporate contextual information like weather and site history to reduce the rate of false positives. Even so, all fire detections come with some degree of uncertainty, which fire agencies handle in different ways.“Different customers have different costs of missing a fire,” he said, explaining that some agencies would rather deal with false positives than run the risk of missing a fire, while others simply don’t have the resources to chase every alarm, including the occasional false one. OroraTech manages these divergent expectations by developing a confidence score for its detections, based on the confidence of its AI model as well as such factors as fire-weather indices, vegetation data, and the persistence of the detection. Users develop an “intuition” for these confidence scores and can filter their notifications accordingly, says Rashkovetsky.Most importantly, he adds, OroraTech tries to be as transparent as possible about how the confidence scores were derived, which he sees as essential for building trust in the technology.“Trust is a major issue in everything remote-sensing related, but specifically also with AI,” he said. “People are rightfully not willing to make a decision based on just a black box.”