IEEE Geoscience and Remote Sensing Magazine - June 2016 - 52

linear global methods (e.g., PCA representation transfer)
to local methods (e.g., based on clustering, as with GM
and MA). If the first can address only rotations, translations, and to some extent scalings of the data clouds, the
others can model the per-sample transformation and allow more flexibility of the transform. The same type of
reasoning holds for semisupervised methods, which will
be able to correct for smaller shifts than methods based
on AL. When deployed in a DA setting, AL methods can
collect target labeled samples that provide evidence of
the real target class distributions, while the semisupervised method uses only unlabeled data in the target and,
therefore, cannot easily discover drastic changes in the
class distributions.
◗ A combination of methods from different families is
also possible. For example, selecting invariant features
can be a preprocessing step to kernel MA, where the labels in the target domain have been acquired by AL using the labels from the source domain.
With these simple guidelines in mind, the analyst can select
the most appropriate strategy (or combine a series of them)
according to the considered data and application.
HOW TO VALIDATE
A typical bottleneck for the employment of an adaptation
strategy is the validation of the adaptation process itself,
since it is assumed that no (or only few) labeled data are
available for the target domain. Nonetheless, one should
assess whether the adaptation was successful in the processing of the target image, even though no labeled samples are available for such
validation. To address this
crucial issue, a circular valiDa iS a RiSing FieLD oF
dation strategy is presented
and applied to remote sensinVeStigation in Remote
ing images in [64]. The stratSenSing, aS it anSweRS
egy is based on the idea that
tHe neeD FoR ReuSing
an intrinsic structure relates
aVaiLaBLe gRounD
the solutions that are consisReFeRence SampLeS to
tent with the source and the
cLaSSiFY oR FuRtHeR
target domains. A solution
pRoceSS new image
for the target domain, for
acQuiSitionS tHat
which no prior information
is available, is assumed to be
maY Be coVeRing
consistent if the solution to
DiFFeRent aReaS.
the source-domain data is associated with an acceptable
accuracy. The solution to the
source-domain data should be obtained by applying the
same DA algorithm in the reverse sense, i.e., by using the
classification labels in place of missing prior knowledge
for target-domain instances. The source-domain data is
considered as unlabeled in the reverse DA learning, and
the accuracy of the source-domain data can be evaluated
due to the available true labels for source-domain samples. This strategy can be effective for both understanding
54

if the adaptation is feasible in the considered data set and
selecting the most effective strategy.
concLuSionS
In this article, we reviewed the recent DA advances for remote sensing image analysis. DA is a rising field of investigation in remote sensing, as it answers the need for reusing
available ground reference samples to classify or further
process new image acquisitions that may be covering different areas, at different time instants, and possibly with
different sensors. The increasing satellite-data availability
trend observed in the last few years (in particular, thanks
to satellite constellations such as the Sentinels or the NASA
A-Train) and the commercialization of drone-mounted
cameras have pushed these problems to the forefront of researchers' and analysts' priorities.
We have reviewed the recent models proposed in the literature, which were grouped in four main families: 1) the
approaches based on the selection of invariant features,
2) those based on the matching of the data representation,
3) those based on the adaptation of the classifier trained
on the source domain, and 4) those based on limited but
effective sampling of labeled samples in the target domain.
With practical examples, we have provided the reader with
a thorough introduction to the field and some guidelines
for the selection of the approaches to use in real application scenarios.
We believe that DA is of the highest importance to future Earth observation since multimodality and repeated
imaging have become unavoidable [7]. The data are already
there, and new, challenging problems can now be tackled
with remote sensing. The discipline has succeeded in entering many new sectors of society, and it is now time to provide the tools to the users to perform a trustable monitoring that can be obtained in different sensor configurations
or modalities. We think that DA and machine learning in
general can contribute to providing an answer to this call.
autHoR inFoRmation
Devis tuia (devis.tuia@geo.uzh.ch) received a diploma in
geography at the University of Lausanne (UNIL) in 2004,
a master of advanced studies degree in environmental engineering at the Federal Institute of Technology of Lausanne (EPFL) in 2005, and a Ph.D. degree in environmental
sciences at UNIL in 2009. Subsequently, he worked as a
visiting postdoctoral researcher at the University of Valencia, Spain, and at the University of Colorado, Boulder. He
then worked as a senior research associate at EPFL under a
Swiss National Foundation (SNF) program. Since 2014, he
has been an SNF assistant professor in the Department of
Geography at the University of Zurich. His research interests include the development of algorithms for information extraction and data fusion of remote sensing images
using machine-learning algorithms. He serves as chair of
the Image Analysis and Data Fusion Technical Committee
of the IEEE Geoscience and Remote Sensing Society. He is
ieee Geoscience and remote sensing magazine

june 2016

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2016