Chapter 7: Perspective on Deep Learning for Earth Sciences
Machine learning and deep learning (DL) in particular have made a huge impact on many fields of science. In the last decade, advanced deep learning methods have been developed and applied to Earth data science problems extensively. Applications on classification and parameter retrieval are making a difference: methods are very accurate, can handle large amounts of data and deal with spatial and temporal data structures efficiently. Nevertheless, several important challenges still need to be addressed. Current standard deep architectures struggle to learn useful Earth feature representations in an unsupervised way, and struggle with long-range dependencies so distant driving processes (in space and time) are not captured, and they cannot cope with non-Euclidean spaces efficiently. DL models are still obscure and resistant to interpretability too and, as other data-driven techniques, they do not necessarily learn physically meaningful and, more importantly, causal relations. Advances are needed to cope with arbitrary signal structures and data relations, physical plausibility and interpretability. This chapter reviews the current approaches and discusses ways forward to develop new DL methods for the Earth sciences in all these directions.