Tracking Pyrometeors With Meteorological Radar Using Unsupervised Machine Learning

• Published in: Geophysical research letters 2020-04, Vol.47 (8), p.n/a
• Main Authors: McCarthy, N. F., Guyot, A., Protat, A., Dowdy, A. J., McGowan, H.
• Format: Article
• Language: English
• Subjects: Burning; Classification; Combustion; Debris; Detritus; Fires; Gaussian Mixture Model; Hydrometeor; Learning algorithms; Machine Learning; Meteorological radar; Plumes; Probabilistic models; Pyrometeor; Radar; Radar data; Radar signatures; Radar tracking; Smoke; Smoke plumes; Speciation; Unsupervised learning; Wildfire; Wildfires; Wind
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2019GL084305

Pyrometeors are the large (>2 mm) debris lofted above wildfires that are composed of the by‐products of combustion of the fuels. One speciation of pyrometeor is firebrands, which are burning debris that lead to ignitions ahead of the surface fire and can be the dominant mechanism of fire spread and structure loss. Pyrometeors are observed by meteorological radar. To date, there have been no investigations into identification of pyrometeor speciation with radar. Here we present an unsupervised machine learning method (Gaussian mixture model) to classify pyrometeor modes using X‐band radar data. The coherent features of the mode of pyrometeor identified most likely to transport firebrands were tracked in time and space. The radar classification and tracking method shows that wildfires do produce signatures in radar returns that could be used for spot fire risk prediction. In wildfires, different types of debris (known as pyrometeors) are lofted in the smoke plumes and transported downwind. Some types of pyrometeors may, when in the air, still be burning and capable of starting new wildfires. Here we investigate the potential for meteorological radar to classify different types of pyrometeors and to track them to determine their potential for starting new fires downwind of the main fire front. Key Points An unsupervised pyrometeor (lofted wildfire debris) classification algorithm is applied to mobile X‐band dual‐polarization radar data Insights on the internal differences of pyrometeors from a wildfire are obtained from the radar data Potential is demonstrated to detect and track pyrometeors with radar including detection of possible firebrands that may ignite spot fires