Can neural networks improve the accuracy of satellite data?

In this project we aim to tackle long-standing challenges in the retrieval of aerosol and cloud properties from satellite measurements using novel machine learning techniques.

Clouds and aerosols play a major role in our climate. Our limited understanding and ability to observe their interactions is still the largest source of uncertainty in predicting climate change.

The only way we can observe them on a global scale is via satellites. Satellite measurements are indirect, so we need specific algorithms to go from the reflected sunlight that the sensors measure to the physical properties we are actually interested in. The current so-called 'retrieval' algorithms make use of our knowledge about how aerosols and clouds reflect and scatter solar radiation. But they still struggle in some areas.

A first challenge is that behind one given retrieval estimate, there can actually be multiple different possible aerosol or cloud situations that all look the same to the satellite sensor. We are using neural networks that can learn to provide a complete probabilistic description of all the possible situations that might have generated a satellite observation. This will allow us to better quantify uncertainties, interpret the retrievals, and understand the retrieval limit for particularly challenging but important situations (e.g. low aerosol loadings that are conducive to cloud formation, but hard to detect).

Machine learning techniques have rarely been applied to this problem in the past, but some have the potential to fill these gaps and help improve our global picture of aerosols and clouds, and how they interplay with climate change.