text
stringlengths
0
820
knowledge could be sought. New insights could be gained
by using the explainable ML model and the outcomes used
to formulate hypotheses for explanations that are not yet
known to us. These hypotheses could then be tested us -
ing simulation software, for example, and confirmed or
rejected. To show the potential, previous neural network
approaches discover presumed scientific laws from obser -
vations without providing the complete underlying prior
knowledge to the learning method (e.g., in [ 80]). As illus -
trated in Figure 7 , the next promising step would be to re -
veal new hypotheses from remote sensing data. This can,
for instance, be accomplished by searching for patterns in
interpretations, which can be assigned to novel discoveries
and insights when combined with domain knowledge. This
approach may point us to previously undiscovered spatio -
temporal input–output relationships or data biases; for ex -
ample, interpretation tools and their resulting insights on
how certain land-use classes are recognized could be used
to derive improved class definitions and help with a tar -
geted acquisition of the training data.
2.29 –2.14 1.22
–0.84 0.86 –0.62Rugged Scene/
Natural/ClimbingMan-Made/
Natural Light/MetalDiving/Swimming/
Climbing
Transpor ting Things
or People/Asphalt/MatteIce/Snow/Climbing Wire/Constructing
Building/Rusty0.2
0.15
0.1
0.05
0Urban
Vegetation
Rocky
Water
(a) (b)
FIGURE 8. Explaining the factors behind landscape beauty in the United Kingdom [77] (a) The maps of landscape attributes contributing
to the estimation of beauty in a series of landscape images; these maps are learned from ground-based photos and are used to predict the
landscape attributes that are observed in single images. Such attributes are then combined end to end in a neural network that predicts
the beauty scores. The number above the single-factor maps correspond to the contribution of the factor to landscape beauty, on average.
(b) The land cover map of the United Kingdom, which is used for visual comparison of the single-factor maps learned automatically.
Authorized licensed use limited to: ASU Library. Downloaded on March 07,2024 at 22:07:36 UTC from IEEE Xplore. Restrictions apply.
99
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINELEARNING 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 spatio -
temporal 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 ma -
chine 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. Ad-
dressing 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, in -
terventions in the system are not possible because of ethi -
cal, 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 advanc -
ing 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 algo -
rithms 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 endeav -
ors, but it also has impactful practical implications. For ex -
ample, 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
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 )
ENSO4
SMreal