Interpretable deep learning for studying the Earth system

Abstract

The Earth system is a highly complex dynamical system. While considerable process understanding has been achieved in past research, many processes and relations in the Earth system remain poorly understood due to this complexity. A better understanding of these processes and relations can improve weather and climate predictions and eventually help make decisions that protect life and property. In this thesis, I evolve the recently proposed approach of using interpretable deep learning to gain new scientific insights into the Earth system. In the approach, a deep learning model is trained to predict one Earth system variable (referred to as target variable) given some others as input. After training the model, the relations between input and target variables that the model learned are analyzed to gain new scientific insights. The major challenge to the approach is that the model may learn spurious correlations rather than actual causal relations. This is a challenge, not only because the scientist cannot gain new scientific insights from a model that learned spurious correlations, but also because detecting whether a given model learned spurious or causal relations is difficult in complex systems. <br />Here, I propose a variant approach to identify spurious correlations that any given statistical model learned. Furthermore, I develop a methodology of causal deep learning models, which combines the approach of using interpretable deep learning to gain new scientific insights with findings from causality research to actually obtain a causal deep learning model, i.e. a model that learns the causal relations between input and target variables. Applied to several examples from hydrometeorology, the variant approach is superior to other commonly applied approaches for identifying spurious correlations that statistical models learn. Moreover, results obtained with causal deep learning models differ entirely from results obtained with a simple linear correlation analysis, which stresses the importance of considering non-linear effects and the difference between correlation and causation. <br />Finally, I apply both methodologies to gain new insights into soil-moisture-precipitation coupling, i.e. the question how soil moisture affects precipitation. Improving our understanding of soil-moisture-precipitation coupling can help to better understand and mitigate extreme events like droughts and floods, and the effects of land management and climate change. The developed methodology of causal deep learning models overcomes several common limitations of previous studies on soil-moisture-precipitation coupling and reveals that an increase in local soil moisture leads to a subsequent increase in precipitation locally, and a simultaneous decrease in precipitation in a surrounding area. The non-local coupling strength exceeds the local coupling strength. These findings contribute to our understanding of soil-moisture-precipitation coupling and stress the importance of non-local effects, which have commonly been neglected in previous studies.