AICCA: AI-Driven Cloud Classification Atlas

Takuya Kurihana, Elisabeth Moyer, Ian T. Foster

Introduction

Clouds represent one of the most significant uncertainties in climate change projections. As the "wild card" of the Earth's energy budget, their behavior—how they reflect solar radiation and trap heat—is critical to understanding future warming. While decades of multispectral satellite data provide a wealth of information, much of this 801 TB archive remains underutilized due to the limitations of traditional classification methods. The AI-driven Cloud Classification Atlas (AICCA) addresses this gap by leveraging unsupervised deep learning to categorize 22 years of satellite imagery into a manageable, physically meaningful dataset.

The Challenge: Beyond Pixel-Level Analysis

Traditional cloud classification systems, such as the International Satellite Cloud Climatology Project (ISCCP), typically rely on pixel-by-pixel analysis of cloud top pressure and optical thickness. While useful, this approach often simplifies the complex nature of clouds into roughly nine broad categories. Crucially, it misses "texture"—the spatial organization of cloud formations.

Mesoscale organization, such as the distinction between open-cell and closed-cell stratocumulus clouds, is vital for determining a cloud's radiative effect and precipitation potential. Because manual labeling of millions of images to capture these textures is impossible, there is a pressing need for automated systems capable of discovering physically relevant cloud regimes without human bias.

Methodology: Rotation-Invariant Cloud Clustering (RICC)

To capture these spatial nuances, researchers developed a workflow called Rotation-Invariant Cloud Clustering (RICC). The process involves several sophisticated stages:

  1. Rotation-Invariant Convolutional Autoencoder: Standard neural networks might interpret the same cloud pattern differently if it is rotated. RICC utilizes a custom loss function that forces the model to learn "latent representations"—compact numerical summaries—of cloud patches that remain identical regardless of orientation.
  2. Hierarchical Agglomerative Clustering (HAC): These latent summaries are grouped into clusters based on similarity.
  3. The Stability Protocol: To determine the optimal number of clusters ($k$), the authors developed a four-part protocol. This ensured the chosen number of classes maximized information richness while remaining stable across different seasons and data subsets. The protocol identified 42 distinct cloud classes as the "sweet spot."
  4. Computational Scaling: Using the Parsl parallel programming library, the team scaled the inference process to analyze nearly 200 million image patches (128 x 128 pixels, or roughly 100 km x 100 km) spanning the years 2000 to 2021.

Data and Compression

The primary input for AICCA consists of 22 years of multispectral data from the MODIS instruments on NASA’s Aqua and Terra satellites. By converting 801 TB of raw imagery into 54.2 GB of labels and physical properties, the project achieved a data compression factor of 15,000. This reduction "democratizes" climate research, allowing scientists to perform global-scale analyses on standard hardware rather than requiring supercomputing resources.

Key Insights and Results

The AICCA dataset provides a much more granular view of the atmosphere than previous models:

  • Enhanced Granularity: While ISCCP identifies a single "Stratocumulus" category, AICCA identifies 30 different low-cloud clusters. This captures subtle regional variations, such as those found off the coasts of North and South America, which have distinct radiative signatures.
  • Physical Coherence: The unsupervised classes align closely with physical expectations. For example, classes identified at high altitudes (low pressure) consistently show high ice content, while low-altitude classes are predominantly liquid.
  • Spatio-temporal Consistency: The clusters exhibit clear geographic "homes" and migrate seasonally with the sun. For instance, specific classes are uniquely associated with the West Pacific warm pool, demonstrating that the AI has captured real-world meteorological phenomena.

Limitations and Future Directions

Currently, AICCA is restricted to ocean regions to avoid the spectral noise introduced by varying land surfaces. Additionally, the 100 km patch size focuses on mesoscale organization, potentially overlooking very small-scale features or extremely large weather systems.

Future work aims to expand the model to include land surfaces and utilize the 22-year record to detect long-term trends in cloud organization. By identifying how these 42 cloud classes shift in response to global warming, AICCA provides a powerful new tool for refining climate models and reducing uncertainty in Earth's future climate trajectory.

Summary updated

Abstract

Clouds play an important role in the Earth’s energy budget, and their behavior is one of the largest uncertainties in future climate projections. Satellite observations should help in understanding cloud responses, but decades and petabytes of multispectral cloud imagery have to date received only limited use. This study describes a new analysis approach that reduces the dimensionality of satellite cloud observations by grouping them via a novel automated, unsupervised cloud classification technique based on a convolutional autoencoder, an artificial intelligence (AI) method good at identifying patterns in spatial data. Our technique combines a rotation-invariant autoencoder and hierarchical agglomerative clustering to generate cloud clusters that capture meaningful distinctions among cloud textures, using only raw multispectral imagery as input. Cloud classes are therefore defined based on spectral properties and spatial textures without reliance on location, time/season, derived physical properties, or pre-designated class definitions. We use this approach to generate a unique new cloud dataset, the AI-driven cloud classification atlas (AICCA), which clusters 22 years of ocean images from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA’s Aqua and Terra instruments—198 million patches, each roughly 100 km × 100 km (128 × 128 pixels)—into 42 AI-generated cloud classes, a number determined via a newly-developed stability protocol that we use to maximize richness of information while ensuring stable groupings of patches. AICCA thereby translates 801 TB of satellite images into 54.2 GB of class labels and cloud top and optical properties, a reduction by a factor of 15,000. The 42 AICCA classes produce meaningful spatio-temporal and physical distinctions and capture a greater variety of cloud types than do the nine International Satellite Cloud Climatology Project (ISCCP) categories—for example, multiple textures in the stratocumulus decks along the West coasts of North and South America. We conclude that our methodology has explanatory power, capturing regionally unique cloud classes and providing rich but tractable information for global analysis. AICCA delivers the information from multi-spectral images in a compact form, enables data-driven diagnosis of patterns of cloud organization, provides insight into cloud evolution on timescales of hours to decades, and helps democratize climate research by facilitating access to core data.
BibTeX
@article{kurihana2022aicca,
  author = {Takuya Kurihana and Elisabeth Moyer and Ian T. Foster},
  title = {{AICCA: AI-Driven Cloud Classification Atlas}},
  journal = {Remote Sensing},
  year = {2022},
  doi = {10.3390/rs14225690},
}
Figure