PNNL

The Science

Clouds affect the Earth’s weather and climate by influencing light, heat, and moisture in the atmosphere. They come in many varieties with unique impacts on the atmosphere. Researchers strive to understand these nuances to accurately model weather and climate. Artificial intelligence (AI) can help study clouds by analyzing satellite images. But typical image recognition AIs require vast amounts of human-generated labels to train. In this work, researchers demonstrate an AI that can be taught to recognize cloud types by looking at millions of satellite images of clouds without requiring human input.

The Impact

Clouds play a pivotal role in the Earth’s climate and weather. They form in a diversity of distinct patterns and types that can have profoundly different impacts on things such as the water cycle and how radiant energy moves through the atmosphere. Earth-observing satellites have been capturing images for decades and these datasets offer a unique opportunity to study clouds and their impact on the Earth system. The massive quantity of satellite images can be prohibitive to evaluate because there is far too much imagery for scientists to sort manually. Self-supervised AI can learn to interpret images without human input; in this work, researchers trained such an AI to differentiate between cloud types in satellite images. This AI will allow for a variety of scientific investigations into cloud properties and behavior. More broadly, self-supervised methods have not yet been widely adopted in the atmospheric sciences. This study demonstrates how those methods can be used to parse extraordinarily large datasets of satellite images.

Summary

AI developed in this research learns to generate a simple numerical representation of cloud images by performing what is essentially an image-matching task. It is trained on pairs of cloud images taken from nearby locations that are likely to contain the same cloud types. It is rewarded for learning similar representations of images with the same cloud type and differing representations for images with different cloud types. The model was evaluated on several datasets of cloud images that have been hand-labeled by past cloud classification studies and performed comparably to classification models that had been trained with human-generated labels. Researchers also demonstrated the model’s ability to generalize between satellite instruments and explored the internal representations of clouds learned by the model by forcing the model to hallucinate images based on its own parameters.

PNNL Contacts

Andrew Geiss, Pacific Northwest National Laboratory, andrew.geiss@pnnl.gov (Corresponding author)

Jerome Fast, Pacific Northwest National Laboratory, Jerome.Fast@pnnl.gov, (Project PI)

Funding

This research was a component of the Integrated Cloud, Land-surface, and Aerosol System Study (ICLASS) supported by the U.S. Department of Energy Office of Science Biological and Environmental Research as part of the Atmospheric Systems Research (ASR) Program. Pacific Northwest National Laboratory is operated by Battelle for the U.S. Department of Energy. Tianle Yuan and Hua Song were supported by NASA.

Related Links

“Using Contrastive Learning to Interpret Mesoscale Marine Cloud Imagery,” American Meteorological Society 2023 annual meeting presentation, Denver CO.