Wildfires are increasing in intensity and frequency as climate change alters air temperature, humidity, and vegetation. Fires like the ones that tore across Australia and the western United States will become much more common.

The early detection of smoke from wildfires is critical to saving lives and infrastructure. Once ignited, a fire can spread at speeds of up to around 14 mph or 2.3 miles every 10 minutes. Wildfires are first detected via fire watchtowers, camera networks, or satellite images of areas of interest. Because the data streams must be monitored by a human being, the task is challenging and prone to error. Deep learning, a subset of machine learning that focuses on neural networks with a high number of layers, is exceptional at rapidly synthesizing gigabytes of data and identifying complicated patterns.

Working with an image dataset from the HPWREN9 tower network in southern California, Argonne and its collaborators are applying deep learning algorithms for smoke detection; the algorithms can constantly and automatically ​“monitor” the image and video streams from fire watchtower networks and satellites and alert officials of smoke in the images.

The team is developing an algorithm that learns temporal, as well as spatial information: combining spatial data models with a long short-term memory (LSTM) network — a kind of recurrent neural network (RNN) used for learning patterns over time. Such hybrid spatial-temporal models have been used in a variety of other applications that involve video or sequential image data, such as person re-identification, object tracking, etc.

The research is part of the SAGE project to design and build a national-scale, reusable cyberinfrastructure to enable AI at the edge. Funded by the National Science Foundation, the research team includes Northwestern University, the University of Chicago, George Mason University, the University of California San Diego, Northern Illinois University, the University of Utah, and the Lincoln Park Zoo.