IEEE Geoscience and Remote Sensing Magazine - December 2017 - 15

be further explored include nonlinear spectral unmixing,
hyperspectral image enhancement, and hyperspectral timeseries analysis.
INTERPRETATION OF SYNTHETIC
APERTURE RADAR IMAGES
Over the past several years, many studies related to deep learning for SAR image analysis have been published. Among these,
deep learning techniques have been used most in typical applications, including automatic target recognition (ATR), terrain surface classification, and parameter inversion. This section
reviews some of the relevant studies in this area.
AUTOMATIC TARGET RECOGNITION
SAR ATR is an important application, in particular, for military surveillance [39]. A standard architecture for efficient
ATR consists of three stages: detection, discrimination, and
classification. Each stage tends to perform a more complicated and refined processing than its predecessor and selects
the candidate objects for the next-stage processing. However, all three stages can be treated as a classification problem
and, for this reason, deep learning has made its mark.
Chen and Wang [40] introduced CNNs into SAR ATR and
tested them on the standard ATR data set MSTAR [41]. They
found the major issue to be the lack of sufficient training
samples as compared to optical images. This might cause severe overfitting and, therefore, greatly limit the capability of
generalizing the model, so data augmentation is employed
to counteract overfitting. Chen et al. [42] propose to further
remove all fully connected layers from conventional CNNs,
which are accountable for most trainable parameters. The
final performance is demonstrated as superior compared to
conventional CNNs on the MSTAR data set (i.e., a state-ofthe-art accuracy of 99.1% in standard operating condition).
Extensive experiments have been conducted to test the

generalization capability of the so-called AConvNets, and
they have proved to be quite robust in several extended operating conditions. The removal of the fully connected layers, originally designed to be trainable classifiers, might be
justifiable in this case because the limited number of target
types can be seen as the feature templates that the AConvNets are extracting.
Many authors have applied CNNs to SAR ATR and tested
the results on the MSTAR data
set, e.g., [43]-[46]. Among these
studies, the one common findSAR ATR IS AN IMPORTANT
ing is that data augmentation is
necessary and the most critical
APPLICATION, IN
step for SAR ATR using CNNs.
PARTICULAR, FOR
Various augmentation strateMILITARY SURVEILLANCE.
gies have been offered, including translation, rotation, and
interpolation. Cui et al. [47]
introduce DBN to SAR ATR, where stacked RBMs are used to
extract features that are then fed to a trainable classifier.
Wagner [48] suggests using a CNN to first extract feature
vectors and then feed them to an SVM for classification. The
CNN is trained with a fully connected layer, but only the
previous activations are used. A systematic data augmentation approach is employed, which includes elastic distortions and affine transformations. It is intended to mimic
typical imaging errors, such as a changing range (which is
scale dependent on the depression angle) or an incorrectly
estimated aspect angle.
Additional studies applying CNNs to the ATR problem
are also of interest. Bentes et al. [49] applied a CNN to ship-
iceberg discrimination, tested on TerraSAR-X StripMap images. Schwegmann et al. [50] applied a specific type of deep
NNs, highway networks, to the discrimination of ships in
SAR imagery and achieved promising results. Ødegaard et al.

Classification Map
One Pixel

Hyperspectral Image

x k-1
xk
x k+1

Network
p

GRU

GRU
Input Layer

Recurrent Layer + Batch Normalization + PRetanh

Fully Connected Layer

Softmax Layer

FIGURE 7. The RNN proposed for the hyperspectral image classification task in [36]. GRU: gated recurrent unit; PRetanh: parametric recti-

fied tanh.
DECEMBER 2017

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2017