Using Machine Learning Techniques to Recover Prismatic Cirrus Ice Crystal Size from 2-Dimensional Light Scattering Patterns

In this paper, we present a prediction model developed to identify particles size of ice crystals in clouds. The proposed model combines a Feed Forward Multi-Layer Perceptron neural network withBayesian regularization backpropagation and other machine learning techniques for feature reduction with Principal Component Analysis androtation invariance with Fast Fourier Transform. The proposed solution is capable of predicting the particle sizes with normalized mean squared error around 0.007. However, the proposed network model is not able topredict the size of very small particles (between 3 and 10 µm size) with the same precision as for the larger particles. Therefore, in this work we also discuss some possible reasons for this problem and suggest future points that need to be analysed.