text
stringlengths
0
820
on the problem considered, an analyst can favor one or
another. However, there are some guidelines that should be taken into account that will depend mainly on the data
available (e.g., the sensors to be used) and on the effort
already provided by the analyst (e.g., whether a classifier is already available or if labels in the target domain are
available or can be acquired easily). The guidelines are
as follows: ◗If the data to be used are acquired by different sensors, they are associated with different feature spaces. In this
case, only heterogeneous DA (i.e., methods that allow
for aligning spaces of different dimensionality) should be considered. Accordingly, multiview feature-representation
transfer methods such as CCA and KCCA or MA (see the
“Adapting Data Distributions” section) are the recom -
mended choices.
◗If a classifier trained on the source is already available and the effort of training is considerable, methods of
the third and fourth families (i.e., the adaptation of
classifiers and the adaptation by selective sampling, respectively) should be preferred. These methods
build on the model that is already defined on the
source domain, while those of the two other families imply the definition of a new classifier that is success -
ful in all domains.
◗Whenever it is possible to acquire new labeled samples in the target, it should be done. There is no better way
to correct for a data set shift than by having examples of
the class-conditional distribution in the target. The AL and MA methods are to be preferred in that case.
◗The level of data set shift the methods can cope with goes along with their level of flexibility. Representation
methods relying on labeled samples from the target can
cope with strong nonlinear deformations because they allow for a kind of feature registration between the do-
mains, while those that do not use target samples (e.g.,
PCA, TCA, and GM) are successful only if the data distri-butions are already prealigned and have not undergone
drastic shifts, such as the cases where the signature of a
target class becomes identical with one of the others in the source. Among the unlabeled methods, the differenc-
es in their flexibility should be considered, going from 500 550 600 650 700 750 800 8506065707580859095
Number of Pixels in Training SetOA (%)Source Labeled
Target Labeled
Random Sampling
TrAdaBoost
Traditional AL
TrAdaBoost + AL
0 5 10 15 20 25 30 350102030405060708090100
AL Iteration
(a) (b)Alphas (%)% Source Norm. Alphas > 0.02
% Target Norm. Alphas > 0.2
Figu Re 10. The AL results over the KSC data of Figure 9 (adapted from [67]). (a) The learning rates (OA) for different methods. (b) The
percentage of source (dashed red line) and target (solid blue line) a weights larger than 0.02 (source) and 0.2 (target) along the iterative AL
process in the TrAdaBoost + AL experiment, which is shown by the black line in (a).
Authorized licensed use limited to: ASU Library. Downloaded on March 08,2024 at 03:13:37 UTC from IEEE Xplore. Restrictions apply.
ieee Geoscience and remote sensin G ma Gazine june 201654
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, transla-
tions, and to some extent scalings of the data clouds, the others can model the per-sample transformation and al-
low 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 semisuper-
vised 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 la -
bels in the target domain have been acquired by AL us -
ing 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 pro -
cessing of the target image, even though no labeled sam -
ples are available for such validation. To address this
crucial issue, a circular vali -
dation strategy is presented
and applied to remote sens -
ing images in [ 64]. The strat -
egy is based on the idea that an intrinsic structure relates
the solutions that are consis -
tent with the source and the
target domains. A solution
for the target domain, for which no prior information
is available, is assumed to be
consistent if the solution to the source-domain data is as -
sociated 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 sam -
ples. This strategy can be effective for both understanding if the adaptation is feasible in the considered data set and
selecting the most effective strategy.
conc LuSionS
In this article, we reviewed the recent DA advances for re -
mote sensing image analysis. DA is a rising field of investi -
gation 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 dif -
ferent 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 re -
searchers’ and analysts’ priorities.
We have reviewed the recent models proposed in the lit -
erature, 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,