See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/323587302

A Deep Learning SAR Target Classification Experiment on MSTAR Dataset

Conference Paper · March 2018

DOI: 10.23919/IRS.2018.8448048

CITATIONS READS

27 1,248

2 authors:

Rene Thaens Cristian I. Coman

NATO - North Atlantic Treaty Organisation NATO - North Atlantic Treaty Organisation
12 PUBLICATIONS   44 CITATIONS    41 PUBLICATIONS   291 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Cognitive Radar View project

Electronic Warfare and Sensor View project

All content following this page was uploaded by Rene Thaens on 29 November 2020.

The user has requested enhancement of the downloaded file.

 A Deep Learning SAR Target Classification Experiment on
MSTAR Dataset

Cristian Coman, René Thaens

NATO Communications and Information Agency

The Hague, THE NETHERLANDS
email: cristian.coman@ncia.nato.int
email: rene.thaens@ncia.nato.int

Abstract: Yet another deep learning method is proposed in this paper for the problem
of automated target recognition (ATR) in Synthetic Aperture Radar (SAR) images. Deep
convolutional neural networks (CNN) classifiers have been demonstrated on benchmark
datasets like MSTAR to outperform classical machine learning algorithms based on for-
mal feature extraction. Most of these deep learning solutions use amplitude information
only and reverse the SAR classification problem to a simple image classification task. In
this paper we analyze the potential of using additional radar information, such as phase
information, in the deep learning process.

1. Motivation
Images generated from Synthetic Aperture Radar (SAR) observations are commonly used in
the Joint Intelligence, Surveillance and Reconnaissance (JISR) domain to develop the so called
Recognized Ground Picture (RGP). Analysts are exploiting raw SAR data to respond to in-
telligence requirements, which often require identification of targets of interest or significant
activities. The exploitation of SAR images is a complex process supported by trained image
analysts, who manually search and classify targets extending just a couple of meters in large
SAR images covering tens of kilometers. The time required to complete this manual analytic
task is significant and impacts on the performance of the JISR community to respond to time
sensitive requests. Developing and maintaining experiences SAR image analysts is a lengthly
process, which can take years of training and significant resources.

Beside being a time consuming and expensive process, SAR image exploitation poses another
challenge, which is the poor quality (low resolution) of images. Physical and technical limi-
tations in radar technologies restrict to image resolutions to tens of centimeters or meters per
pixel. When observing an armored vehicle, which extends a few meters the numbers of pixels
is limited to a few tens as well. The image is more or less a collection of scatters and often even
highly train analysts have difficulties with the detection and classification of targets in SAR
images through visually guessing the shape of the target.

Researchers in the deep Convolution Neural Networks (CNN) domain have recently adopted
SAR automated target classification (ATR) as one of the benchmark problems for highlighting

The 19th International Radar Symposium IRS 2018, June 20-22, 2018, Bonn, Germany 1
978-3-7369-9545-1 2018
c DGON
the potential of these new methods in the radar domain. In [1] and [2] a five layer CNN is
reported to achieve a significant improvement in the classification accuracy when compared
with popular machine learning methods such as Support Vector Machines (SVM) or Adaptive
boosting (AdaBoost) used for SAR ATR. A regularization technique is proposed in [5] to reduce
the free parameters of CNN model and consequently reduce the training time and improve the
convergence properties. The problem of lack of SAR training data is addressed in [6] where a
multilayer auto-encoder (AE) is used to prevent the over-fitting caused by training on restricted
datasets. Most of the solutions proposed in the literature are exclusively using the magnitude
information from SAR images and disregard other radar data, such as the phase of the signal.

In this paper we analyze an simple solution for automated SAR image target recognition, which
includes phase information into the deep learning training process. The phase provides addi-
tional discriminatory information, which increases the classification accuracy and resolve some
of the challenges related to over-fitting and training on reduce datasets.

The remainder of this paper is organized as follows. Section 2 introduces the SAR data available
for the experiment and explains how the phase information is exploited in the classification pro-
cess. The architecture of the convolution neural network is presented in Section 3 where details
of the training approach is discussed as well. Experimental results are presented in Section 4
and Section 5 contains concluding remarks.

2. Synthetic Aperture Radar Data

Airborne radars used in Earth observation missions employ range and azimuth compression
techniques to generate high resolution images of observed scenes. The images constructed from
microwave radar signals are expressed in the complex domain by using amplitude and phase
information. The amplitude information gives and indication of the radar reflectivity and is
commonly presented like a black and white picture. The phase information is more difficult
to be interpreted by an analyst and often is not shown on the radar display. Still the phase
contains information about the targets and through additional feature extraction techniques this
information can be utilized for detecting and classifying target in SAR data. For example, in-
terferometric processing of phase information is a popular technique for detecting changes in
SAR images taken at different moments in time.

The Moving and Stationary Target Acquisition and Recognition (MSTAR) [7] collected by
Sandia National Laboratory in a project jointly sponsored by the Defense Advanced Research
Projects Agency and the Air Force Research Laboratory has been used in the experiment dis-
cussed in this paper. The data is available on the Internet and contains SAR images of ten
targets including tanks, armored vehicles, weapons systems and military engineer vehicles (ar-
mored personnel carrier: BMP-2, BRDM-2, BTR-60, and BTR-70; tank: T-62, T-72; weapon
system: 2S1; air defense unit: ZSU-234; truck: ZIL-131; bulldozer: D7). The data was collected
with a Sandia X-band radar operating at 9.60GHz with a bandwidth of 0.591GHz. The range
and cross range resolution are identical and equal to 0.3047m. Each target file contains a header

2
 0 6
35 35 −10 35 35
−20
−10 5
30 30 30 30
−40 −20

Range [m]
Range [m]

Range [m]
Range [m]

25 25 25 25 4
−60 −20
20 20 −30 20 20
−80 3
15 15 −30 15
15
−100 −40 2
10 10 10 10
−120 −40
−50 5 1
5 5 5
−140
0 0 0 −50 0 0
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
Cross Range [m] Cross Range [m] Cross Range [m] Cross Range [m]

(a) (b) (c) (d)

Figure 1:MSTAR chip images for target ”t72 tank”: (a) normalized amplitude [dB], (b) normal-
ized real [dB], (c) normalized imaginary [dB] and (d) phase information [rad].

where detailed sensor and sensor-target information is provided. Following the header the SAR
data is provided, first the amplitude information then the phase information. The target data is
presented as SAR chips containing between 128 and 179 rows and columns. For purpose of
this paper all the data was normalized to block of 88 rows by 88 columns by taking the central
part of the SAR chip provided in MSTAR data set. The amplitude and phase information for
the target ”t72 tank” are depicted in Fig. 1, (a) and (d) respectively. The target can be easily
observed in the amplitude image. The phase information does not present easily observable dis-
criminating features but was used to generate the real (Fig. 1, (b)) and the imaginary (Fig. 1,
(c)) parts of the original complex image. The real and imaginary parts were obtained by mul-
tiplying the normalized (0 to 1) amplitude data with the cosine and sine functions of the phase
information. The data was further normalized by translation into the positive domain of values
and re-scaling (0 to 1). This normalization process facilitates the representation of data in dB
values in Fig. 1. The use of real and imaginary parts of the signal helps with the augmentation of
the information contained in the training data set. CNNs used in visual image recognition often
employ operations such as dimension reduction and translations and rotations to increase the
training data set. For SAR classification problems the phase data can also be used as a simple
technique to extend the information contain in limited training data sets. In the data analysis
phase of the experiment reported in this paper it was possible to look into the Fourier transform
of the complex SAR image as well. Classification based on images in the frequency domain is
yet another way radar information can be employed in deep neural networks.

3. Network Model

The scope of the work reported in this paper is limited to exploration of potential benefits of
phase information in SAR classification problems. One of the idea is to compare the perfor-
mance of amplitude only and amplitude and phase SAR classification methods. For this pur-
pose, two similar networks have been tested in the experiment. The networks are composed of
two convolution layers, a flatten layer, two dense layers, and two dropout layers The parameters
of the networks are presented in Tab. 1. The only difference between the two networks is in
the shape of the input of the first convolution layer: the amplitude network receives a 2D image
with the size of 88 × 88, whereas the amplitude and phase network takes images with the shape

3
 Table 1: Parameters of the deep CNNs used in SAR classification.

layer (type) input shape kernel features output shape param

input (Conv2D) 88 × 88(88 × 88 × 3) 3×3 32 86 × 86 × 32 320(896)
conv 1 (Conv2D) 86 × 86 × 32 3×3 32 84 × 84 × 32 9248
pool 1 (MaxPool) 84 × 84 × 32 2×2 32 42 × 42 × 32 0
drop 1 (Dropout) 42 × 42 × 32 0.25 32 42 × 42 × 32 0
flat 1 (Flatten) 42 × 42 × 32 1 56448 56448 0
dense 1 (Dense) 56448 1 128 128 7225472
drop 2 (Dropout) 128 0.25 128 0
dense 2 (Dense) 128 1 10 10 1290

88 × 88 × 3. This difference is also reflected in the number of parameters to be trained in the

input layer: 320 for the amplitude only and 896 for amplitude and phase network. All layers
use rectified linear units (ReLu) activation functions, with the exception of the last dense layer
where a softmax activation function is used.

The input images for the amplitude only network simply represent the SAR absolute value
information. For the amplitude and phase network the input images are constructed with three
layers: the first layer is the amplitude data, the second layer is the real data and the third layer
is the imaginary part (see Fig. 1 (a), (b) and (c)).

4. Experimental Results
The public MSTAR data set is already partitioned on training and test subsets for all ten classes
of targets. Images for training are captured at a depression angle of 17◦ , and images for testing
are acquired at 15◦ depression angle. Test and training images are collected over the full aspect
angle range (0◦ to 360◦ ). The train/test partition based on depressions angle suggests concerns
with the classifier when images are acquired by a platform flying at different altitudes. The
range and cross-range resolution in SAR images do not depend on the target sensor geometry
(they are related to the signal bandwidth and length of the synthetic aperture) but the shadow of
objects is the main feature, which depend on the depression angle. The complete set of training
images (3671 in total for all ten classes) and test images (3203 for all ten classes) is used in
this paper. A mosaic of images (five samples per class) used for training of the amplitude only
network is depicted in Fig. 2.

The networks were trained for 40 epochs with a batch size of 32 samples. The mean value of
the correct classification accuracy on the test dataset, across all ten classes, was slightly better
for the case when the phase was used, 91% compared to 90% in case of the amplitude only
network. Such a small increase in the classification accuracy can be explained by relatively lim-
ited amount on discriminatory information contained in the phase part of the MSTAR dataset.

4
 2s1_gun bmp2_tank brdm2_truck btr60_transport btr70_transport d7_bulldozer t62_tank t72_tank zil131_truck zsu23-4_gun

Figure 2: Mosaic of SAR image samples - amplitude information.

Common SAR focusing algorithms do preserve discriminant information in the phase compo-
nent, however for the MSTAR dataset it is not evident how the focusing was performed and it
is difficult to estimate why target information is not fully reflected in the phase data.

The confusion matrix of the amplitude and phase network is depicted in Fig. 3 (a), with the
horizontal axis representing the predicted class labels and the vertical axis the true labels.

2s1_gun 235 3 1 3 9 1 14 7 0 1

bmp2_tank 500
3 561 0 2 5 0 0 16 0 0

brdm2_truck acc
btr60_transport
13

2
15

2
220

1
2

182
2

3
2

0
8

0
6

3
6

0
0

2
400 1.5 loss
val acc
Accuracy/Loss

btr70_transport 6 6 0 0 184 0 0 0 0 0
val loss
True label

300
d7_bulldozer 1 0 0 0 0 270 0 0 2 1
1.0
t62_tank 17 0 1 0 0 2 226 21 4 2 200

t72_tank 4 3 0 2 1 0 5 567 0 0

zil131_truck 11 1 0 5 0 1 0 11 242 3
100 0.5
zsu23-4_gun 2 0 0 0 0 7 0 2 1 262
0

0.0
n

n
dm nk

ck

tra rt

rt

er

nk

13 nk

ck
gu

gu
po

d7 spo

oz
tru

tru
ta

ta

ta
1_

4_
ns

ld
_

2_

2_

0 10 20 30 40
2_

1_
n
p2
2s

3-
ul
tra

t6

t7
bm

_b

u2
0_

0_

zil

zs
br

r6

r7

Epochs
bt

bt

Predicted label

(a) (b)

Figure 3: Amplitude and phase network metrics: (a) confusion matrix, (b) accuracy and loss.

The training of the networks converged fast and after about 10 epochs the validation accuracy
on the train data (acc) and on the test data (val acc) already reaches about 90% (see Fig. 3 (b)).

5
 5. Conclusions
Deep convolutional neural networks show good results when used in automatic SAR target
recognition problems. SAR phase information has the potential to improve the classification
accuracy, especially when the focusing algorithm preserves target information in the phase data.
The phase information can also be used to enhance the sample dataset when a limited number
of samples is causing over-fitting during the training phase. In this paper a simple methods was
introduced for employing phase information in the CNN process, namely the use of real and
imaginary data to generate multichannel images.

Two lines of efforts could be identified to further investigate the potential use of deep CNN in
SAR target recognition in production systems: developments in the CNN domain such as the
use of ensemble models, capsule networks, or transfer learning techniques and the enhancement
of SAR data set used in the CNN training (e.g. spectral images, polarimetry images etc.). Re-
liable CNN classification techniques are essential for the automation of the production of the
Recognized Ground Picture (RGP), in particular in the exploitation SAR, electro-optic, infrared
and full motion video data.

References
[1] H. Wang, S. Chen, F. Xu, and Ya-Qiu Jin, “Application of deep-learning algorithms to
MSTAR data”, Proc. of 2015 IEEE International Geoscience and Remote Sensing Symposium
(IGARSS),2015, pp. 3743-3745.
[2] S. Chen, H. Wang, F. Xu, and Ya-Qiu Jin, “Target Classification Using the Deep Convolutional
Networks for SAR Images”, IEEE Transactions on Geoscience and Remote Sensing,vol. 54, no. 8,
AUGUST 2016, pp. 4806-4817.
[3] Q. Zhao and J. C. Principe, “Support vector machines for SAR automatic target recognition”, IEEE
Trans. Aerosp. Electron. Syst., vol. 37, no. 2, pp. 643654, Apr. 2001.
[4] Y. J. Sun, Z. P. Liu, S. Todorovic, and J. Li, “Adaptive boosting for SAR automatic target recogni-
tion”, IEEE Trans. Aerosp. Electron. Syst., vol. 43, no. 1, pp. 112125, Jan. 2007.
[5] S. Wagner, K. Barth, and S. Brggenwirth, “A deep learning SAR ATR system using regularization
and prioritized classes”, Proc. of 2010 2017 IEEE Radar Conference (RadarConf), 2017, pp. 0772-
0777
[6] S. Deng, L. Du, C. Li, J. Ding, and H. Liu, “SAR Automatic Target Recognition Based on Euclidean
Distance Restricted Autoencoder”, IEEE Journal of Selected Topics in Applied Earth Observations
and Remote Sensing,vol. 10, no. 7, 2017, pp. 3323-3333.
[7] E. R. Keydel, S. W. Lee, and J. T. Moore, “MSTAR extended operating conditions: A tutorial”, in
Proc. of 3rd SPIE Conf. Algorithms SAR Imagerytext, 1996, vol. 2757, pp. 228242.

DISCLAIMER: Any opinions expressed herein do not necessarily reflect the views of the NCI
Agency, NATO and the NATO Nations but remain solely those of the author(s).

View publication stats