کد مقاله | کد نشریه | سال انتشار | مقاله انگلیسی | نسخه تمام متن |
---|---|---|---|---|
5427030 | 1508614 | 2017 | 13 صفحه PDF | دانلود رایگان |
![عکس صفحه اول مقاله: Applying machine learning methods for characterization of hexagonal prisms from their 2D scattering patterns - an investigation using modelled scattering data Applying machine learning methods for characterization of hexagonal prisms from their 2D scattering patterns - an investigation using modelled scattering data](/preview/png/5427030.png)
- Zernike moments (ZM) are suitable for the analysis of 2DLSP with high symmetry.
- ZM and machine learning methods recover particles' properties from their 2DLSPs.
- Orientation, aspect ratio, size are recovered from 2DLSP for smooth hexagonal prisms.
Better understanding and characterization of cloud particles, whose properties and distributions affect climate and weather, are essential for the understanding of present climate and climate change. Since imaging cloud probes have limitations of optical resolution, especially for small particles (with diameter <â25â¯Âµm), instruments like the Small Ice Detector (SID) probes, which capture high-resolution spatial light scattering patterns from individual particles down to 1â¯Âµm in size, have been developed. In this work, we have proposed a method using Machine Learning techniques to estimate simulated particles' orientation-averaged projected sizes (PAD) and aspect ratio from their 2D scattering patterns. The two-dimensional light scattering patterns (2DLSP) of hexagonal prisms are computed using the Ray Tracing with Diffraction on Facets (RTDF) model. The 2DLSP cover the same angular range as the SID probes. We generated 2DLSP for 162 hexagonal prisms at 133 orientations for each. In a first step, the 2DLSP were transformed into rotation-invariant Zernike moments (ZMs), which are particularly suitable for analyses of pattern symmetry. Then we used ZMs, summed intensities, and root mean square contrast as inputs to the advanced Machine Learning methods. We created one random forests classifier for predicting prism orientation, 133 orientation-specific (OS) support vector classification models for predicting the prism aspect-ratios, 133 OS support vector regression models for estimating prism sizes, and another 133 OS Support Vector Regression (SVR) models for estimating the size PADs. We have achieved a high accuracy of 0.99 in predicting prism aspect ratios, and a low value of normalized mean square error of 0.004 for estimating the particle's size and size PADs.
Journal: Journal of Quantitative Spectroscopy and Radiative Transfer - Volume 201, November 2017, Pages 115-127