Cloud and Aerosol Variability Across Different Synoptic Regimes Identified with a Deep Learning-based Unsupervised Clustering Model

To simultaneously constrain multiple large-scale conditions through clustering, we developed Deep Spatiotemporal Clustering (DSC) models with deep learning-based unsupervised clustering algorithms. Inspired by the U-net architecture, DSC utilizes an autoencoder integrating Convolution Neural Network (CNN) and Long Short-Term Memory (LSTM) layers to learn latent representations of the spatiotemporal data. DSC also includes a unique layer for cluster assignment on latent representations that uses the student’s t-distribution. Extensive experiments demonstrated the proposed approach outperforms both conventional and previous deep learning-based unsupervised clustering algorithms. Additionally, we created its various variants (CNN encoder, CNN autoencoder, CNN-LSTM encoder, CNN-LSTM autoencoder, etc.) to get insight into using both the CNN and LSTM layers in the autoencoder.

To investigate the influence of different synoptic conditions on aerosol and cloud variability over the eastern North Atlantic (ENA) region, we utilized reanalysis daily 2-D maps of multiple variables (500-hPa Geopotential height, Sea-level pressure, 10-m U, and 10-m V) from 2016 to 2021 as input data for the DSC model. This allowed us to classify the synoptic regimes that exhibited strong correlations with different cloud conditions. By characterizing and comparing aerosol and cloud variability using ENA ground-based observations from the U.S. DOE Atmospheric Radiation Measurement (ARM) ENA atmospheric observatory, aerosol reanalysis data, and satellite retrievals across different synoptic regimes, our analyses suggest that the differences in cloud macro and micro properties among these regimes are more significant than the changes observed in most aerosol properties.

Acknowledgement: This work is supported by the DOE Office of Science Early Career Research Program. This work was performed under the auspices of the U.S. DOE by LLNL under contract DE-AC52-07NA27344. LLNL-ABS-853417