Convolution Neural Network to Classify Ice Crystal Habit

Knowledge about the light scattering and absorption properties of cloud particles is crucial to know the effect clouds have on climate. Accurate information about the size, shape and phase of cloud particles are needed to represent cloud processes in climate models. Current programs use parameters (i.e. maximum dimension, projected area, perimeter) from pictures obtained by aircraft mounted cloud particle imagers to classify the ice crystals by shape, but these programs are not nearly as precise or accurate enough as manual classification, which is tedious, time consuming, and subjective. A more reliable automatic method to identify crystal shape would be a powerful tool for scientists to further study radiative transfer through clouds. In this project, a machine learning program is created using a convolution neural network to more accurately and quickly identify ice crystal shapes in clouds. Currently it classifies 8 crystal habits (quasi sphere, bullet rosette, column, plate, aggregates of bullet rosette, column, plate, and frozen droplets), with 78% accuracy. The program is still being improved to show skill and practicality over “Ice-cloud particle habit classification using principal components” Lindqvist et al. 2012 which uses principle component analysis coupled with Bayesian and kNN machine learning techniques to achieve 81% accuracy with 8 ice crystal habits. Current advantages over Lindqvist et al. 2012 are a probability distribution assigned to each crystal instead of definite classifications, the program can be easily trained to add more ice crystal habits by other people, and the program will be available publicly.