IEEE Geoscience and Remote Sensing Magazine - June 2021 - 98

neural network can be used to check whether correct decisions
have been made for the wrong reasons (the so-called
Clever Hans effect [78]). This can be seen as an additional
test for the reliability of the output due to the fact that a
high score for the network does not always mean a correct
result. In summary, these tools can increase confidence
by improving traceability, as estimates are generated and
reveal biases in the data through human-understandable
visualization.
PERSPECTIVES
Explainable ML has thus far received comparatively little
attention in remote sensing, partly because of the stillpredominant
opinion that explainability is tightly coupled
with the complexity of a model and, therefore, an increase
in explainability leads directly to a decrease in accuracy
(e.g., in [79]). In the meantime, however, several applications
have shown that this is no longer the case.
Most of the approaches considered thus far are post hoc
interpretations, but initial approaches that consider interpretability
by design are appearing. In [76], for instance, the
model is forced to predict human-interpretable concepts
before predicting the final task (Figure 8). Such approaches
have the potential to provide both reliability checks and
human-understandable explanations and could be used to
move toward physics-explicit models, similar to those discussed
in the next section.
As a further step, explanations going beyond today's
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 using
simulation software, for example, and confirmed or
rejected. To show the potential, previous neural network
approaches discover presumed scientific laws from observations
without providing the complete underlying prior
knowledge to the learning method (e.g., in [80]). As illustrated
in Figure 7, the next promising step would be to reveal
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 spatiotemporal
input-output relationships or data biases; for example,
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 targeted
acquisition of the training data.
2.29
-2.14
1.22
Urban
Rugged Scene/
Natural/Climbing
-0.84
Man-Made/
Natural Light/Metal
0.86
Diving/Swimming/
Climbing
-0.62
0.2
0.15
0.1
0.05
Vegetation
Rocky
Water
Transporting Things
or People/Asphalt/Matte
Ice/Snow/Climbing
(a)
Wire/Constructing
Building/Rusty
(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.
98
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2021

IEEE Geoscience and Remote Sensing Magazine - June 2021

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