Segmentation of polarimetric radar imagery using statistical texture

Weather radars are increasingly being used to study the interaction between wildfires and the atmosphere, owing to the enhanced spatio-temporal resolution of radar data compared to conventional measurements, such as satellite imagery and in situ sensing. An important requirement for the continued proliferation of radar data for this application is the automatic identification of fire-generated particle returns (pyrometeors) from a scene containing a diverse range of echo sources, including clear air, ground and sea clutter, and precipitation. The classification of such particles is a challenging problem for common image segmentation approaches (e.g. fuzzy logic or unsupervised machine learning) due to the strong overlap in radar variable distributions between each echo type. Here, we propose the following two-step method to address these challenges: (1) the introduction of secondary, texture-based fields, calculated using statistical properties of gray-level co-occurrence matrices (GLCMs), and (2) a Gaussian mixture model (GMM), used to classify echo sources by combining radar variables with texture-based fields from (1). Importantly, we retain all information from the original measurements by performing calculations in the radar's native spherical coordinate system and introduce a range-varying-window methodology for our GLCM calculations to avoid range-dependent biases. We show that our method can accurately classify pyrometeors' plumes, clear air, sea clutter, and precipitation using radar data from recent wildfire events in Australia and find that the contrast of the radar correlation coefficient is the most skilful variable for the classification. The technique we propose enables the automated detection of pyrometeors' plumes from operational weather radar networks, which may be used by fire agencies for emergency management purposes or by scientists for case study analyses or historical-event identification.