IEEE Geoscience and Remote Sensing Magazine - June 2021 - 101

widely used approach in Earth and climate sciences to
quantitatively identify relationships between time series. It
tests whether including past states of a variable, X, improves
the prediction of an output variable, Y, more than considering
other covariates. GC is a linear test, but nonlinear (kernel)
versions have also been proposed [85]. In [86], a generalized
kernel GC is presented, able to discover footprints
of El Niño-Southern Oscillation (ENSO) on soil moisture
(SM) and vegetation optical depth records [see Figure 9(a)].
However, GC approaches have problems in nonstationary,
nonlinear, and deterministic relationships, especially in
dynamic systems with weak-to-moderate coupling.
The second family considers nonlinear state-space methods,
such as convergent cross mapping (CCM) [87]. CCM
attempts to address GC problems by reconstructing the
variable's state spaces, (Mx, My), using time embeddings
and concludes, based on XY " , whether points on Mx can
be predicted more accurately using nearest neighbors in My
as more points are used for prediction. However, CCM is
very sensitive to noise and time-series length. Recent works
have included bootstrap resampling to alleviate such problems
and shown good results in identifying causal links
in long global records of carbon and water fluxes [88] [see
Figure 9(b)].
The third family, collectively known as causal network
learning algorithms, relies heavily on conditional-independence
tests. Its methods iteratively remove the links
between pairs of variables, (X, Y), if they are found to be
independently conditioned on any subset of the other variables.
The PC algorithm (named after its inventors Peter
and Clark) allows us to identify parents and can be flexibly
implemented with different kinds of conditional-independence
tests, which can handle nonlinear dependencies and
variables that are discrete or continuous and are univariate
or multivariate. Finally, structural causal models (SCMs)
are used when time is not involved or the sampling frequency
is too low. SCMs search for the causal direction within
Markov-equivalent classes by exploiting the asymmetries
between cause and effect. Additive noise models rely on the
principle of independence between the cause and the generating
mechanism and have recently shown good results
in remote sensing and geosciences in cases where time is
not involved and only two variables are observed [21].
PERSPECTIVES
Although the field of machine and deep learning has traditionally
progressed very rapidly, we observe that this is
not the case in tackling the challenge of learning causal
relationships from Earth observation data. The role that
deep learning will play on causal discovery is, at best, uncertain
because deep learning models focus mostly on fitting
and are largely overparameterized, which is (apparently)
against causal, sparse, reasoning. Only very recently have
we witnessed efforts toward either incorporating or understanding
deep models causally: the authors in [89] implemented
a metalearning objective that maximizes the speed
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
of domain transfer, which, under certain assumptions, can
be seen as a way to localize changes in causal mechanisms.
In [90], the authors learned individual-level causal effects
from observational data that can efficiently handle confounding
(hidden) factors. Both methods are, in principle,
well suited to the problems in remote sensing and geoscience
data sets, which exhibit spatiotemporal relationships
to be exploited but have not (thus far) been considered.
Yet we will have to face a more important challenge:
cognitive barriers. Domain knowledge is elusive and difficult
to encode, interaction between computer scientists
and physicists is still a barrier, and education in synergistic
concepts still needs to become a reality in coming years.
Causal inference is believed to be the best approach to develop
Earth sciences, but this will be possible with a strong
and continuous interaction between domain knowledge
experts and computer scientists.
CONCLUSIONS
This article described the six ideas and six directions
in which geosciences, Earth observation, and AI can
achieve a lot if synergistically combined. With this article,
we have provided our appreciation for research avenues
that are new, refreshing, and exciting for scientists
willing to evolve at the interface between AI and the geosciences.
We hope that they will spark curiosity and that
the community, especially the younger generations, will
embrace them.
ACKNOWLEDGMENTS
Xiao Xiang Zhu is jointly supported by the European Research
Council (ERC) under grant ERC-2016-StG-714087,
by the Helmholtz Association through the Framework of
Helmholtz Artificial Intelligence Cooperation Unit and
Helmholtz Excellent Professorship Data Science in Earth
Observation-Big Data Fusion for Urban Research, and by
the German Federal Ministry of Education and Research
in the framework of the international future AI lab AI4EO.
Gustau Camps-Valls was partly funded by the ERC under
the ERC-SyG-2019 USMILE project (grant agreement
855187). Nathan Jacobs was partly funded by a National
Science Foundation CAREER Award (IIS-1553116). Devis
Tuia is the corresponding author.
Some of the ideas presented in this article originated
from discussions during the first workshop of the ELLIS
Program ML for Earth and Climate Science (Germany) a
few days before the COVID-19 lockdown in Europe.
AUTHOR INFORMATION
Devis Tuia (devis.tuia@epfl.ch) is with Ecole polytechnique
fédérale de Lausanne, Sion, 1950, Switzerland. He is the
corresponding author for this article. He is a Senior Member
of IEEE.
Ribana Roscher (ribana.roscher@uni-bonn.de) is with
the University of Bonn, Bonn, 53115, Germany. She is a
Member of IEEE.
101
IEEE Geoscience and Remote Sensing Magazine - June 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2021
