Understanding Low Cloud Mesoscale Morphology with an Information Maximizing Generative Adversarial Network

This is a Preprint and has not been peer reviewed.

This Preprint has no visible version.

Download Preprint
Downloads

Download Preprint

Authors

Tianle Yuan 

Abstract

Generative adversarial networks (GANs) are a class of machine learning algorithms with two neural networks, one generator and one discriminator, playing adversarial games with each other. Information maximizing GANs (InfoGANs) is a particular GAN type that tries to maximize mutual information between a subset of latent variables and generated samples, thereby establishing a mapping between the latent variables and generated images. Here we demonstrate the feasibility of classifying low clouds mesoscale morphology with no human supervision by training an InfoGAN. We take a set of latent variables as mesoscale cloud morphology categories and successfully train a generator to map each category variable to realistic images that belong to a corresponding mesoscale morphology. The trained generator generates visually realistic cloud scenes. Furthermore, the model learns ten physically meaningful categories each corresponding to a particular morphology. We also show that by perturbing other latent variables while keeping the cloud category variable the same, the model can generate images that have the same morphology but with substantial variations. The trained discriminator can be used to classify real cloud scenes with limited training samples in the future.

DOI

https://doi.org/10.31223/osf.io/gvebt

Subjects

Artificial Intelligence and Robotics, Atmospheric Sciences, Computer Sciences, Oceanography and Atmospheric Sciences and Meteorology, Physical Sciences and Mathematics

Keywords

machine learning, GAN, cloud morphology, low clouds

Dates

Published: 2019-06-05 17:55

Older Versions
License

GNU Lesser General Public License (LGPL) 2.1

Add a Comment

You must log in to post a comment.


Comments

There are no comments or no comments have been made public for this article.