Classifying Global Low-Cloud Morphology with a Deep Learning Model: Results and Potential Use

Low cloud morphology has profound impact on low cloud evolution, cloud radiative effect on the ocean surface and climate, and how clouds may change under climate change. In this study, we present results of classifying low cloud morphology with deep learning models using both the supervised and unsupervised techniques. We use both autoencoders and generative adversarial nets to extract features of satellite images and classify them with extracted features. Both methods provide classifications that are physically sound based on expert inspection and geographic distribution of these morphological types. In the supervised approach, we iteratively improve our classification performance by increasing training data and tweaking model setup. We achieve state-of-the-art results on both training (99% accuracy) and validation (82% and improving at the time of submission). With these different models, we can start to analyze low cloud morphology distribution and their variability in a systematic way. We will present insights learned from these analyses as well as lessons learned from training our models.