IEEE Geoscience and Remote Sensing Magazine - June 2016 - 45

0.1
90

0.098

0.094

0.096

0.092
0.09
−0.65

65
−0.645

−0.64
−∆
(a)

−0.635

−0.63

0.098

0.096

0.094

0.092

0.1

0.09
−0.65

−0.645

−0.64
−∆
(b)

−0.635

−0.63

FiguRe 5. The Pareto front estimated using a multiobjective genetic algorithm for the selection of six features. Each dot corresponds to a
feature set minimizing (1). The color indicates the OA on (a) the source test set T s and (b) the target test set T t , according to the reported
color scale bar (adapted from [30]).

In the following, we report the experimental results obtained on a hyperspectral image acquired by the Hyperion
sensor of the Earth Observation 1 satellite in an area of the
Okavango Delta, Botswana [32] (see Figure 4). For more information about the experimental setting and the obtained
results, see [30]. The labeled reference samples were collected on two spatially disjoint areas with slightly different
characteristics, thus representing two different domains.
The samples taken on the first area, which was considered
as the source domain, were partitioned into a training set
T s and a test set T s by random sampling. Samples taken on
the second area (i.e., the target domain) were used to derive
a training set T t and test set T t according to the same procedure. The estimated Pareto front for the selection of six
features is reported in Figure 5.
Each point on the two graphs corresponds to a different
selected feature subset F, i.e, a feature set minimizing (1).
In Figure 5(a), the color of the points indicates the overall
accuracy (OA) obtained on the source-domain test set T s
using an SVM classifier trained using T s (according to the
reported color scale bar). In Figure 5(b), the color indicates
the OA obtained by the same SVM classifier on the targetdomain test set T t .
The results show that the solutions with higher relevance
D result in better classification accuracies on the source domain. However, relevance alone is not sufficient for selecting features that are stable for the classification on the target
domain. We observe that the most accurate solutions on the
target domain T t are those that exhibit a good tradeoff between the relevance and invariance terms, which confirms
the importance of the invariance term and shows that the
P measure is able to capture the information of feature
stability. To select the subset of features that leads to good
generalization capabilities on different domains, tradeoff
solutions between the two competing objectives should be
identified. The selected subset of features results in an OA
of 91% on the source domain and 80.7% on the target. The
set of features selected according to the optimization of D
results in an OA of 92.7% on the source but only 64.4% on
june 2016

ieee Geoscience and remote sensing magazine

the target, showing that the features selected by accounting
for the data set shift between the domains can significantly
improve the generalization capability on the target domain.
ADAPTING DATA DISTRIBUTIONS
The second family reviewed considers DA methods that
aim to adapt the representation of the original data, regardless of the model that will process them afterward. A review
of the methods proposed in computer vision and machine
learning can be found in [16]. Here, we will focus on the approaches proposed in the remote sensing literature. This type
of adaptation is often done by relative normalization methods,
i.e., methods that do not provide physical units as an output
but that instead provide similarly distributed digital numbers.
Their aim is to make the data
distributions more similar
across the domains to train a
single model that can simultHe aim oF Da metHoDS
taneously classify the source
iS to aDapt a cLaSSiFieR
and target domains.
tRaineD on tHe SouRce
In general, a data represenDomain to maKe
tation transformation with the
pReDictionS on tHe
aim of making data sources
taRget Domain.
more similar should have the
following desirable properties:
◗ The method should be able
to align unpaired data (see
the "Unpaired" column in Table 1), which allows the
alignment of noncoregistered data (not even imaging the
same location) or data with different spatial resolutions.
◗ The method should be able to align data of different dimensionality (see the "D Dimensionality" column in Table 1) to
allow multisource classification.
◗ The method should be able to align several domains at
the same time (see the "Multisource" column in Table 1)
to enhance multitemporal adaptation instead of pairwise adaptation.
◗ The method should be able to align in a nonlinear
way (see the "Nonlinear" column in Table 1), since the
47

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