Increasingly large amounts of satellite observations are produced by NASA and ESA Earth observation missions. Most of these observations measure a superposition of different sources and dynamics. In this project we aim, to separate respective satellite observations into contributing disjunct dynamics. For example, satellite observed ocean sea surface temperature (SST) is a very important variable to estimate the oceanic heat uptake and therefore the progress of climate change on Earth. However, the respective SST observations are superposed contributions from tides, convection, advection, turbulence, trends and measurement-noise (Fig. 1). Since not all of these processes contribute to the oceanic heat uptake in the same way (or even with the same sign), separating observed SST into its dynamic contributions would substantially improve the estimates of oceanic heat uptake and their respective long term forecasts as well as mitigation strategies. Existing methods for this task, e.g, data assimilation, strongly rely on and are biased towards the employed prior information, e.g., from numerical models. In data sciences,  these methods are called supervised.

The project’s main objective is to separate satellite-based Earth system observations into their dynamic components by unsupervised machine learning (ML). The advantages over existing methods are 1) easy re-application to all kinds of (also non-oceanic) satellite data without generating additional training data or encoding physical constraints;  2) omission of biases and errors usually introduced by the prior knowledge, e.g., from numerical models. 

In this project, the separation will entirely base on formal criteria and does not need prior domain specific information:

A separation is deemed successful, i.e, dynamically-disjunct, when by removing one of the separated parts, a ML forecast-skill that bases only on the remaining data set is decreased as much as possible. The ML forecast-skill will be evaluated by point-wise losses as RMSE and by value-distribution losses as Wasserstein distance. From the technical point, we develop an antagonistic back and forth training approach of two neural networks (NN), a Separation-NN  and a Forecast-NN that form a novel kind of dynamics-focused masked auto-encoder. 

This project s funded by Helmholtz Einstein International Berlin Research School in Data Science (HEIBRiDS) (https://www.heibrids.berlin/)