TARGET DETECTION AND CLASSIFICATION USING SEISMIC SIGNAL PROCESSING IN UNATTENDED GROUND SENSOR SYSTEMS Yuxin Tian Texas

A&M University Department of Electrical Engineering College Station, TX 77843 Email: ytian@ee.tamu.edu
ABSTRACT The most challenging problem in target detection and classication is the extraction of a robust feature vector which can effectively represent a specic type of target. The use of the seismic signals in unattended ground sensor systems brings new challenges to the problem because of the complexity of the seismic waves and their highly dependency on the underlying geology. This paper proposes a new feature extraction algorithm spectral statistics and wavelet coefcients characterization (SSWCC). SSWCC extracts a feature vector from both the frequency and the time-frequency domain analysis of the seismic signals, including the spectrum, the power spectral density (PSD) and the wavelet coefcients. The SSWCC algorithm is designed for real-time applications, and has shown its robustness and effectiveness through a series of experiments. Extensive performance evaluation is conducted to derive the optimal conguration of the different parameters. The overall classication accuracy can reach as high as 90%. 1. INTRODUCTION An Unattended Ground Sensor (UGS) is a device placed on the ground which automatically gathers and interprets sensor data from remote targets, then transfers results back to some higher level processing center [1]. Until now, UGS systems have been used in a variety of applications ranging from industrial monitoring to military information gathering. Typically used sensors in UGS systems include acoustic and seismic sensors. Most of the target detection and classication methods have been proposed based on acoustic signal processing because seismic waves are more complicate to analyze. Seismic waves propagate in different forms, different directions, different speeds, and are highly dependent on the underlying geology. However, seismic sensor also has an advantage over acoustic systems. Acoustic vehicle recognition will be affected by Doppler effects, by noises introduced from various moving parts of vehicles, and by atmospheric and terrain variations, while seismic waves are less sensitive to these factors. This paper discusses the use of seismic signals for target detection and classication. In general, seismic waves can be classied into two categories: body waves and surface waves. Body waves travel at a higher speed through the interior of the Earth and propagate in three dimensions, while surface waves travel near the
The author performed the work while at the University of Tennessee. This research was supported in part by DARPA under grant N66001-0018946

Hairong Qi The University of Tennessee Department of Electrical and Computer Engineering Knoxville, TN 37996 Email: hqi@utk.edu
surface of the Earth and propagate in two dimensions. Research of seismic waves shows that if the disturbance, such as an earthquake, explosion or human interference occurs near the surface, a signicant amount of seismic energy is dispersed near the surface and is transmitted as surface waves (70% of the energy of the impact is distributed in the Rayleigh waves, and the remaining 30% of the energy is transmitted into the earth via body waves [2]). Therefore, in UGS systems, when we study the seismic waves generated by the ground vehicles, we pay more attention to the surface waves, especially on Rayleigh waves. Based on different processing domains, current feature extraction methods for seismic signals can be classied into three categories: time domain methods, frequency domain methods, and time-frequency domain methods. Generally speaking, time-domain analysis is always not very accurate because of the interfering noise, the complicated waveforms, and the variation of the terrains. Most target classication research use either frequency-domain or time-frequency domain methods. High-order spectral analysis is popularly used in frequency-domain analysis [3]. Time-spectral combination analysis was proposed in [4] for target localization from a three-axial seismometer. This paper develops a new feature extraction algorithm, spectral statistics and wavelet coefcients characterization (SSWCC), which extracts a series of feature elements from the spectrum, the power spectral density (PSD) and the wavelet coefcients of the seismic signals collected from the UGS systems. Compared with other algorithms, SSWCC synthesizes more information from different domain-transformations of the original signal and generates a robust feature vector. In order to reduce the correlation existed among feature elements, SSWCC also applies principal component analysis (PCA) which reduces the dimension of the feature space without losing important information. Both k-nearestneighbor (kNN) and minimum distance classiers are used to classify the feature vectors. We show that simple classiers can achieve high accuracy as long as the feature vector is robust and representative, which is important in real-time applications. 2. SSWCC FEATURE ANALYSIS Feature extraction plays an important role in target classication since the performance of the classier largely depends on the quality of the feature vector. SSWCC extracts features from both frequency and time-frequency domains. A block diagram of the algorithm is shown in Fig. 1. The rst step of feature extraction is signal preprocessing, where

Seismic Signal

Signal Preprocessing

Removing DC Component

Wavelet Denoising

Spectral Analysis
Amplitude Statistics
1012.5 1012.5

PSD Analysis
Peak Analysis Peak Analysis Shape Statistics

Wavelet Analysis
Coefficient Analysis Energy Analysis

Feature Analysis

Feature Normalization

are needed. The result of using Blackman window is shown in Fig. 2. It is obvious that the noise level in Fig. 2 (D) is much less than that in Fig. 2 (B), and the dominant frequency peaks are more distinctive than those in Fig. 2 (B) as well. From the spectrum of the signal we extract three groups of features: the amplitude statistics (Eq. 1), the shape statistics (Eq. 2) and the location of the largest three peaks from the spectrum. The amplitude statistics provides a statistical measurement of local spectral energy content over the signal band. The shape statistics provides a higher order statistics in terms of mean, standard deviation, skewness and kurtosis [3]. The peak locations in the spectrum reveal the dominant frequencies of the vibration of vehicles. amp = amp = amp = amp =
1 N 1 N 1 N 1 N N i=1 C(i) N 2 i=1 (C(i) amp ) C(i)amp 3 N ) i=1 ( amp C(i)amp 4 N ) 3 i=1 ( amp N 1 i=1 iC(i) S N 2 i=1 (i shape ) C(i) ishape 3 N i=1 ( shape ) C(i) ishape 4 N i=1 ( shape ) C(i) 3

Shape Statistics

Principal Component Analysis Classification

Feature Vectors

Classifier

(1)

Fig. 1. Block diagram of SSWCC feature extraction. shape =

shape = the DC component of the original signal is rst eliminated, and wavelet denoising is applied to remove the noise in the original signal. The next step is feature extraction. The spectrum, the PSD and the wavelet transforms of the preprocessed signals are derived and potential features are extracted. To eliminate the scaling effects among different features, all the potential features are normalized. Principal component analysis (PCA) is then applied to help optimize the feature vector and to reduce the dimension of the feature space by selecting the most signicant feature components. The output of PCA will be used as the input to the classier. This paper focuses on feature analysis. Readers are referred to [5] for more detailed discussion. 2.1. Spectral Analysis Seismic signals from ground vehicles are generated from the vibration of vehicle engines and their strike on the ground. Thus, frequency responses of the signals from a certain engine should be identical. Spectral analysis is able to reveal the frequency information of a vehicle, and serve as one of the major features.
Original Time Series Data x(t) 2
1.5 1 0.5 0 0.5 1 1.5 1012 1013 1013.5 1014 (t) 0.2 0 0.6 0.4

shape = shape =
1 S

1 S 1 S

(2)

where S = N C(i), C(i) is the spectral magnitude for the ith i=1 frequency bin, and N is the number of the frequency bins. 2.2. Power Spectral Density Analysis Power spectral density (PSD) describes the energy distribution of a signal in the frequency domain. Mathematically, PSD is dened as the Fourier transform of the autocorrelation of the time series signal. In practice, Welchs averaged, modied periodogram method is used. At a certain noise level, the peaks in the PSD shows the most powerful frequency components in the signal. Therefore, the shape of the PSD also represents the characteristics of the signal. Figure 3 shows the PSD shapes of four different targets. Difference can be observed from the peak locations and the energy distribution pattern.
Magnitude(dB) Magnitude(dB)
0 10 20 30 40 50 10 0 10 20 30 40 0 50 100 50 0 50 100

Spectrum of Signal x(t) W(f)

1

0.8

50

100

60

f(Hz)

f(Hz)

x(t)2
1.5 1 0.5 0 0.5 1 1.5 1012

Time Series Data After Blackman Window

Spectrum of Signal x(t) W(f)

1

(A) PSD of Vehicle "DW" Magnitude(dB) Magnitude(dB)

20 30 40 50 35 40 45 50 55 60 60 0 50 100 65

(B) PSD of Vehicle "AAV"

0.8 0.6 0.4 0.2 0

1013

1013.5

1014 (t)

50

100

70

f(Hz)

50

100

f(Hz)

(C) PSD of Vehicle "LAV"

(D) PSD of "No Vehicle"

Fig. 2. Effect of the Blackman window. Fig. 3. PSDs of different vehicles. The Fourier spectrum is derived from a 2-second segment of the seismic signal. Although a rectangular window generates the maximal sharpness, it brings the frequency leakage to the spectrum simultaneously. Therefore, other forms of window function
Student Version of MATLAB

Conventional methods of analyzing the PSD is to calculate the averaged PSD in a series of time blocks to represent the signature of the signal. However, in our problem, signals are processed in
Student Version of MATLAB

real time, which means information from other time blocks is not available. Therefore, we select the locations of the largest three peaks from each PSD as the features of the signal. To describe the shape and the distribution of the signal, we use the shape statistics dened in Eq. 2. Thus we obtained 7 features from the PSD analysis. 2.3. Wavelet Analysis Wavelet transform is designed to analyze the non-stationary signals. Many algorithms have been proposed to develop robust feature vectors from the wavelet transform. These algorithms, however, require more complex computation and are not suitable for real-time applications. SSWCC derives a set of features from the wavelet coefcients generated from a three-band discrete wavelet transform. Figure 4 shows a segment of seismic signal and its wavelet transform. We use a 2-second time series segment. The data are sampled at 256 Hz. Thus the wavelet coefcients are composed of 512 samples. The rst 64 points correspond to C4, which belong to the third scale residue of the wavelet transform. The second 64 points correspond to C3 that belong to the third scale high pass band. The next 128 points correspond to C2 that belong to the second scale high pass band. The last 256 points are C1 that belong to the rst scale high pass band. In SSWCC, we choose functions from the Daubechies wavelet family as the low-pass and the high-pass lter.
0.5
Amplitude

applying the transformation matrix to the original feature vectors, we can then derive a set of most dominant features. 4. PERFORMANCE EVALUATION We use two supervised classiers to test the performance of the feature vector. A training set and a test set are generated for the evaluation. The raw data come from one-axial seismic sensors. We select two types of vehicles, DW (Dragon Wagon) and AAV to test the feature extraction algorithm we developed. The classier needs to distinguish three classes: DW, AAV, and No Vehicle. The signal is sampled at 256Hz. For a 2-second segment, 512 samples are selected from the raw data. The training set and the test set are constructed from different RUNs. We have built a training set including 320 samples and a test set including 500 samples. We use Performance Level (PL) to measure the quality of the classication, PL = Number of Correct Classication Number of Total Observation (3)

4.1. Effect of Parameter k in kNN and Energy Loss The energy loss, , selected in PCA determines the number of feature vectors used in the classier. The number of the nearest neighbors included in the classication, k, also affects the classier performance. We design two experiments to illustrate the effects of and k on the performance level: 1) x , and change k to evaluate the classier; 2) x k at the maximum performance level and change to evaluate the classier. Figure 5 shows the PL with a changing k but a xed . We can see that in most cases, the performance is above 80% when k varies from 1 to 180. The highest PL can reach 87% when k is within the interval [130, 170].
PL
0.9 0.85

0.8
Amplitude

0.6 0.4 0.2 0 0 0.2

C4 C3 C2 C1

0.4 0.5 0 100 200 300 400 500

n

0.6

100

200

300

400

500
n

= 0.01 PL 0.85

0.9

= 0.05

Fig. 4. Different levels of the wavelet coefcients Existing methods use the original wavelet coefcients to serve as a feature set which greatly increases the dimensionality of the feature space. In SSWCC, we choose to calculate only three features from each level of the transformation: the average, the vari2 ance, and the energy (||p||2 = n= pn where pn is the wavelet coefcient). For a three-band wavelet transform, we can obtain 4 3 = 12 features. 3. PRINCIPAL COMPONENT ANALYSIS The elements of the feature vectors can be strongly correlated with each other. Therefore, it is necessary to reduce the feature set to a minimum but sufcient one. The principal component analysis (PCA) technique is a popular algorithm inStudent Version of MATLAB reducdimensionality tion while keeping the most important information untouched in the feature sets by a linear transformation matrix. The transformation matrix is made up of the eigenvectors of the feature vector covariance matrix. Let i represent the eigenvalues of the covariance matrix, vi the eigenvectors, then by choosing the K largest i s and the corresponding eigenvectors, we can construct the transformation matrix. The choice of K is determined by M K i=1 i 1 where is the loss of energy. By i=1 i /

0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35

0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35

20

40

60

80

100

120

140

160

180

200

220

240

(A) PL
0.9 0.85 0.8 0.75 0.7 Student Version of MATLAB 0.65 0.6 0.55 0.5 0.45 0.4 0.35

k PL
0.9

20

40

60

80

100

120

140

160

180

200

220

240

(B) = 0.2

= 0.1

0.85 0.8 0.75 0.7

Student Version of MATLAB 0.65 0.6 0.55 0.5 0.45 0.4 0.35

20

40

60

80

100

120

140

160

180

200

220

240

(C)

20

40

60

80

100

120

140

160

180

200

220

240

(D)

Fig. 5. PL of kNN at different and k.

Student Version of MATLAB Student Version of MATLAB

Next, we x k and change from 0 to 0.5, the performance is shown in Fig. 6. The results show the maximum peak always appears when the energy loss is around 0.2 0.3, and the maximum PL can reach 87%. From the above experiments, we observe that when k changes from 120 to 180, the performance level is stable, in most cases, around 85%. On the other hand, when the energy loss is set between 0.2 and 0.3, the classier has the best performance. The PL can reach 87%.

PL 0.87
0.86 0.85 0.84 0.83 0.82 0.81 0.8 0.79 0.78 0.77 0.76 0.75 0.74 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35

0.88

k = 121

PL 0.87
0.86 0.85 0.84 0.83 0.82 0.81 0.8 0.79 0.78 0.77 0.76 0.75

0.88

k = 141

0.4

0.45

(A) PL0.88
0.87 0.86 0.85 0.84 0.83 0.82 0.81 0.8 0.79 0.78 0.77 0.76 0.75 0.74 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45

0.5

0.74

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

(B)
0.88 PL 0.87 0.86 0.85 0.84 0.83

0.5

k = 151

k = 155

generate the training set and the test set. The result is shown in Fig. 8. Compare Fig. 7 and Fig. 8, we can see that the PL of the minimum-distance classier can reach around 82%. The minimumdistance classier always performs worse than the kNN, which is due to the Gaussian assumption the minimum-distance classier made about the probability density function of the sample variables.
1 1

Student Version of MATLAB

0.82 0.81 0.8 0.79 0.78 0.77 0.76 0.75

Student Version of MATLAB

PL
0.95 0.9

= 0.01

PL
0.95

= 0.1
0.9

0.85

0.85

0.8

0.8

(C)

0.5

0.74

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

(D)

0.5

0.75

0.75

0.7

0.7

0.65

0.65

Fig. 6. PL of kNN with different k, 0 0.5.

Student Version of MATLAB Student Version of MATLAB

0.6

0.6

0.55

0.55

0.5

20

40

60

80

100

120

140

160

180

200

220

240

260

280

300

0.5

20

40

60

80

100

120

140

160

180

200

220

240

260

280

300

Test Time (A)

1

(B)
1

Test Time

PL

= 0.2
0.95 0.9

PL
0.95

4.2. Selection of Training Set and Test Set No matter how good the feature extraction algorithm is or how well the classication algorithm performs, if the training set is not representatively chosen, the performance level can still be very low. In this experiment, we change the content of the training set and the test set in order to test the algorithms on a more reasonable basis. Our approach is to combine the training set and the test set to form a new set which contains about 800 samples. Then we separate this set into two parts randomly and treat one set as the training set and the other one as the test set. We apply the kNN classier to the newly formed training set and the test set. The PL is shown in Fig. 7. For each energy loss , we randomly divide the data set into the training set and the test set for 450 times. For the rst 50 times, we choose k = 5, for the next 50 times, k is increased by 20. The PL prole in
PL
1 1

= 0.3
0.9

0.85 Student Version of MATLAB 0.8

0.85 Student Version of MATLAB 0.8

0.75

0.75

0.7

0.7

0.65

0.65

0.6

0.6

0.55

0.55

0.5

20

40

60

80

100

120

140

160

180

200

220

240

260

280

300

0.5

20

40

60

80

100

120

140

160

180

200

220

240

260

280

300

(C)

Test Time

(D)

Test Time

Fig. 8. Performance of the minimum-distance classier.

Student Version of MATLAB Student Version of MATLAB

5. CONCLUSIONS A new feature extraction algorithm, spectral statistics and wavelet coefcients characterization (SSWCC), is proposed to do target detection and classication using seismic signals. SSWCC obtains statistic features from the Fourier spectrum, the PSD, and the wavelet coefcients. The minimum-distance classier and the kNN classier are used to do the classication. The performance level of the classication could reach as high as 90%. 6. REFERENCES [1] G. E. Sleefe, S. Peglow, and R. Hamrick, The application of unattended ground sensors to stationary targets, in Peace and Wartime Applications and Technical Issues for Unattended Ground Sensors. SPIE, July 1997, vol. 3081, pp. 2129.

PL
0.95 0.9 0.9

0.95

0.85

0.85

0.8

0.8

0.75

0.75

0.7

0.7

0.65

0.65

0.6

0.6

0.55

K=5 K=25
0 50

K=45

0.5

K=65

K=85

K=125 K=165 K=145 K=105

0.55

K=5 K=25
0 50

K=45

100

150

200

250

(A) PL
1

=0.01

300

350

400

450

0.5

K=65

K=85

K=125 K=165 K=145 K=105

100

150

200

250

Test Time
1

(B) PL
0.95 0.9

=0.1

300

350

400

450

Test Time

0.95

0.9

0.85 Student Version of MATLAB 0.8

0.85 Student Version of MATLAB 0.8

0.75

0.75

0.7

0.7

0.65

0.65

0.6

0.6

0.55

K=5 K=25
0 50

K=45

0.5

K=65

K=85

K=125 K=165 K=145 K=105

300 350 400 450

0.55

K=5 K=25
0 50

K=45
150

100

150

200

250

0.5

K=65

K=85

K=125 K=165 K=105 K=145

300 350 400 450

100

200

250

(C)

=0.15

Test Time

(D)

=0.20

Test Time

Fig. 7. Performance of the kNN classier when the training set and the test set are randomly chosen.
Student Version of MATLAB Student Version of MATLAB

[2] G. Succi, G. Prado, R. Gampert, T. Pedersen, and H. Dhaliwal, Problems in seismic detection and tracking, in Unattended Ground Sensor Technologies and Applications II. SPIE, July 2000, vol. 4040, pp. 165173. [3] D. H. Kil and F. B. Shin, Pattern Recognition and Prediction with Application to Signal Characterization, American Institute of Physics, 1996. [4] R. A. Gramann, M. B. Bennett, and T. D. OBrien, Vehicle and personnel detection using seismic sensors, in Sensors, C3I, Information, and Training Technologies for Law Enforcement. SPIE, January 1999, vol. 3577, pp. 7485. [5] Y. Tian, Target detection and classication using seismic signal processing in unattended ground sensor systems, M.S. thesis, University of Tennessee, Knoxville, TN, July 2001.

Fig. 7 is obviously better than the PL we obtained from the last two experiments, which shows the importance of the generation of the training set. We can achieve a better performance in this experiment. The highest performance level can reach 97%. When is from 0.15 to 0.3, the PL is stable and almost insensitive to the selection of k. 4.3. Performance of Different Classiers In this experiment, the minimum-distance classier is used to classify the feature vector. We use the same method as Sec. 4.3 to