IEEE Geoscience and Remote Sensing Magazine - June 2021 - 99

LEARNING CAUSE-AND-EFFECT
RELATIONSHIPS FROM DATA
Earth is a highly complex, dynamic, and networked system,
where very different physical, chemical, and biological
processes interact in several spheres and at diverse spatiotemporal
scales. Despite the great predictive capabilities
of current machine and deep learning methods, there
is still
little actual
learning-understanding is harder
than predicting. ML algorithms excel at fitting arbitrary,
functional data relationships but do not have a clear
notion of the underlying causal relationships. ML is far
from problem understanding and even moreso from machine
intelligence (see the critical perspectives in blogs
by Gary Marcus and Michael Jordan and the articles in
[1] and [22]).
Earth system data analysis aims to extract information
from multivariate, nongridded data sets, where missing
data, nonlinearities, and nonstationarities are present in
the wild. Variables and physical processes are coupled in
space and time, and (tele)connections can be large range,
discontinuous, and variant in strength and intensity. Addressing
this problem will allow us to identify the right set
of predictors, develop robust models, and avoid getting the
right answer for the wrong reasons. The links with physics
(direction 4 in Table 1) and interpretability (direction 5 in
Table 1) are very strong.
CAUSALITY AS THE WAY FORWARD
Causal inference aims at discovering and explaining the
causal structure of a system [20], [81], [82]. Very often, interventions
in the system are not possible because of ethical,
practical, or economic reasons. Then, observational
causal inference comes into play to extract cause-and-effect
relationships from multivariate data sets, going beyond the
commonly adopted correlation approach, which merely
captures associations between variables.
Today, the science of causal inference [19], [83] is advancing
fast and, under reasonable assumptions, can unravel
causal relationships among two or more coupled variables
even in the presence of nonlinearities and nonstationarities
and even when time is not involved. Several rigorous algorithms
have been developed in the last decade that enable
us to make inferences across multiple variables to discover
plausible causal relationships from observations. Causal
inference is, of course, very relevant for scientific endeavors,
but it also has impactful practical implications. For example,
learning causal structures enables us to build more
parsimonious and robust models: that means faster, more
fault-tolerant, and interpretable models.
A TAXONOMY OF CAUSAL-DISCOVERY METHODS
Causal-discovery methods can be divided into four main
families. First, Granger causality (GC) [84] is the most
0.4
0.8
ENSO4
0.08
-0.8
-0.4
SMreal
SMimaginary
VODreal
VODimaginary
2011 2012 2013 2014 2015 2016 2017
(a)
0.02
0.04
0.06
ENSO4 --> SM
(b)
FIGURE 9. Examples of causal-inference approaches in remote sensing and the geosciences. (a) Nonlinear GC, showing a time series of
ENSO4 (ENSO in region 4) that captures sea-surface temperature anomalies in the central equatorial Pacific, SM, and vegetation optical
depth (VOD) to explore their causal relationships extracted using the nonlinear principal component analysis in [91]. (b) The five-day lagged
correlation (top) and causal (bottom) maps of ENSO4 and SM interannual components using the kernel GC method in [86]. Results show
that many of the correlations, even the highest ones (~ .),08t
are not causal, thus suggesting mere spurious associations. (continued )
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
99
Corr (ENSO4, SM)

IEEE Geoscience and Remote Sensing Magazine - June 2021

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