Charting the Realms of Mesoscale Cloud Organisation using Unsupervised Learning

TL;DR

This paper introduces an unsupervised neural network approach to map and analyze the continuum of mesoscale shallow cloud organization, revealing its environmental drivers and impact on Earth's albedo, with implications for climate modeling.

Contribution

It presents the first continuous mapping of cloud organization states using neural networks, linking environmental factors and cloud albedo variations.

Findings

Identified key environmental variables influencing cloud organization.

Quantified cloud organization impact on albedo variations.

Demonstrated temporal evolution capture of cloud states.

Abstract

Quantifying the driving mechanisms and effect on Earth's energy budget, of mesoscale shallow cloud organisation, remains difficult. Partly because quantifying the atmosphere's organisational state through objective means remains challenging. We present the first map of the full continuum of convective organisation states by extracting the manifold within an unsupervised neural networks's internal representation. On the manifold distinct organisational regimes, defined in prior work, sit as waymarkers in this continuum. Composition of reanalysis and observations onto the manifold, shows wind-speed and water vapour concentration as key environmental characteristics varying with organisation. We show, for the first time, that mesoscale shallow cloud organisation produces $\pm 1.4\%$ variations in albedo in addition to variations from cloud-fraction changes alone. We further demonstrate how the manifold's continuum representation captures the temporal evolution of organisation. By enabling study of states and transitions in organisation (in simulations and observations) the presented technique paves the way for better representation of shallow clouds in simulations of Earth's future climate.