FDL 2020: How does atmospheric dust affect marine clouds?
Marine Boundary Clouds, pollution and climate
Marine boundary layer (MBL) clouds span thousands of kilometres over the world’s oceans in the tropics. Through the reflectance of incoming solar radiation they exert a large cooling effect upon the climate. This cooling effect is dependent not only on the thickness of the clouds, but also upon their meso-scale structure. Even small changes in cloud structure may result in large, non-linear changes in the Earth’s energy budget.
Aerosols (dusty pollution particles) are thought to be responsible for the most rapid and extreme changes to marine boundary clouds - possibly even more rapid than the effects of climate heating. Aerosol-cloud interactions (ACI) not only increase the brightness of clouds directly, but may also influence cloud structure, which in turn further changes the response of the cloud to aerosols. This results in a complex system of interactions that is both important to understanding future climate change, but also challenging to understand.
The FDL Europe 2020 Clouds and Aerosols team applied self-supervised learning to classify cloud structure as a function of time and space. They analysed the results using a powerful causal inference model - a machine learning technique that can tease out cause and effect from complex data.
MBL clouds and interactions
MBL clouds exhibit a large range of structures which typically span hundreds of kilometres. Some examples shown in the image above include closed-cell (a), open-cell (b) and broken stratocumulus (c), and cumulus clouds (d), which have taller individual cloud columns. These clouds decrease in reflectivity (a) - (d) and hence have different radiative impacts upon the climate, but also may respond to ACI differently.
Investigating the impact of ACI cloud structure is difficult due to the complexity of interactions between and within clouds, and feedback processes with the environment - as shown above. There is no direct causal link to ACI, so the team opted to understand how aerosols affect cloud droplet density and processes such as precipitation.
Tracking clouds in complex fused data
The team combined a time-series of infrared images from the SEVRI instrument on the MSG satellite with the ERA-5 meteorological data from the European Centre for Medium-Range Weather Forecasts. This fused dataset offers the possibility of making predictions based off of individual tracked clouds, rather than bulk statistics.
To classify the MBL cloud structures, the team implemented a novel mixture-of-experts model utilising LTSMs to predict future observations from a time-series of past observations. A mixture assignment network was trained to select the best expert prediction for each pixel of the input image, resulting in a semantic segmentation for cloud type.
Finding meaning
Using droplet number density as a proxy for aerosol concentration, the team built a Recurrent Neural Network whose architecture encoded the complex interactions from the previous diagram. The network was trained on the fused dataset of over 20,000 time-series of cloud structures (see the FDL tech memo for more details).
After marginalising out unobserved parameters (e.g., the effect of in-cloud circulation), the model can predict and estimate the causal effect of droplet number density on cloud type - even at night, when satellite measurements are no longer available.
Results of causal analysis
An example prediction of cloud structure occurrence using the causal model are shown below. The model ingests a time-series of real cloud observations and predicts how the observed cloud types would change for a given droplet number density intervention, for a single trajectory. The different colour lines represent the different cloud type classifications including clear sky, broken cumulus, stratus, and various types of stratocumulus cloud.
For the no-intervention case, the team observed a mixture of closed-cell stratocumulus and cumulus clouds prior to clear sky. If droplet density is reduced, they sew a large increase in the proportion of broken cumulus. Contrary to this, for increased droplet density they predicted an increase in the proportion of closed-cell stratocumulus and a delay in the length of time taken for clear sky conditions.
The team’s predictions suggest that an increase in aerosol concentration leads to an increase in the occurrence of closed-cell stratocumulus, and an increase of cloud lifetime.
Summary and future work
The combination of novel cloud classification techniques and causal inference of time series of classified satellite observations shows great potential for investigating the complex aerosol impacts on cloud structure and lifetime. However, as these techniques are new and experimental, further evaluation is required to ensure that the results provide solid evidence for real cloud processes.
Further work will involve predicting the radiative forcing impacts of cloud structure changes in order to better understand the climate impacts of aerosol effects on cloud structure and lifetime.
Resources:
FDL Tech Memo [LINK]
Acknowledgments
This work is the result of the 2020 ESA Frontier Development Lab (FDL).
