Constraining aerosol-cloud interactions through multi-sensor fusion of satellite-based Earth observations using machine learning


Aerosol-cloud interactions remain the largest uncertainty in anthropogenic perturbations to the Earth’s radiation balance underlying climate change. In this project we will develop new machine learning based approaches to combine the strengths and information content of multiple heterogeneous satellite observations to provide novel observational constraints on aerosol-cloud interactions.


The Earth is being continuously observed by a wide array of satellite sensors, some of which are flown in synchronised constellations observing the same scenes only seconds apart. Yet, retrievals of atmospheric properties are almost exclusively done independently for each instrument. This is because traditional methods make it difficult to combine heterogeneous satellite observations (e.g. along track aerosol-backscatter curtains from lidars with wide-swath aerosol optical depth retrievals from spectral imagers or along-track cloud radar reflectivity curtains with wide-swath spectral imager cloud property retrievals) in an optimal way. Therefore, significant opportunities to strengthen our observational constraints on aerosol-cloud interactions, in particular in terms of the vertical distribution of clouds and aerosol properties, remain currently unexplored.


This project will develop novel machine learning methods for multi-sensor fusion of heterogeneous satellite observations based on modern deep convolutional neural network architectures, such as Conditional Generative Adversarial Networks (CGANs), to provide constraints on the vertical distribution of clouds and aerosols, which will also provide more robust constraints on the retrieved physical cloud and aerosol properties.


The derived products will be applied as novel observational constraint on the highly uncertain representation of aerosol-cloud interactions in cloud-resolving and climate models, reducing the overall uncertainty in anthropogenic perturbations to the climate system.


Philip Stier (supervisor), Yarin Gal, Department of Computer Science (co-supervisor)

Planned secondments


Enrolment in Doctoral degree

PhD in Physics at University of Oxford


oxlogo sq noborder