AAAR 37th Annual Conference October 14 - October 18, 2019 Oregon Convention Center Portland, Oregon, USA
Abstract View
Aerosol Shape Classification by Deep Learning of Scattering Patterns
PATRICIO PIEDRA, Yong-Le Pan, Gorden Videen, U.S. Army Research Laboratory
Abstract Number: 189 Working Group: Remote and Regional Atmospheric Aerosol
Abstract Remote sensing of aerosol properties by inversion commonly requires directionally averaged aerosol light-scattering shape models such as spheres or spheroids. However, these shape-averaged models often yield discrepancies in retrievals at different wavelengths, leading to inaccurate or at-best ambiguous aerosol classification. Furthermore, shape-averaging does not allow discrimination of trace, non-averaged, scattering patterns. In this work, we have applied machine-learning algorithms to the calculated light-scattering patterns from particles of seven different, common, and naturally occurring shapes to test whether their shapes can be classified. Our scattering data set is produced from particles of volume-equivalent size parameter 5 and refractive index m = 1.5 + 0i, which have been isotropically and stochastically rotated. Furthermore, our dataset was either one-dimensional, depending on the polar angle, or two-dimensional, depending on both the polar and azimuthal angles. The neural network architectures required either a fully connected or a convolutional neural network. As expected, classification capabilities were much greater when the two-dimensional scattering data were used than when only one-dimensional data were considered. When the two-dimensional intensity patterns are considered, the prediction capabilities were approximately 70% for the regularly shaped particles and above 90% for the highly irregularly shaped particles. These capabilities increased slightly when linear polarization was used as input. These results suggest that high-accuracy (i.e., > 90 %) aerosol shape classification can potentially be achieved using a two-dimensional convolutional neural network.