Proceedings of the 37th Chinese Control Conference

July 25-27, 2018, Wuhan, China

Target Detection of Hyperspectral Image Based on Convolutional

Neural Networks
Xuefeng Liu1, Congcong Wang1, Qiaoqiao Sun1, Min Fu2
1. Qingdao University of Science & Technology, Qingdao 266061, China
E-mail: [email protected]
2. Ocean University of China, Qingdao 266100, China
E-mail: [email protected]

Abstract: Convolutional neural networks (CNN) has been applied in image classification and target detection
successfully, however, it is rarely introduced to the field of hyperspectral image (HSI) target detection. Therefore, in
this paper, a hyperspectral image (HSI) target detection method based on CNN is proposed. Firstly, the raw HSI data
is preprocessed and the spectral information could be obtained. Secondly, to extract the feature information, a CNN
is trained and the parameters of the network are adjusted according to a HSI. Finally, the targets will be calibrated
according to the extracted features. To estimate the target detection performance of the proposed method, deep belief
network (DBN) and SVM methods are compared in the experiment of the real world AVIRIS HSI experiment.
Numerical results show that the proposed method has promising prospect in the field of HSI target detection.

Key Words: Deep learning, CNN, Target recognition, DBN, Remote sensing image


(2) Target detection: spectral features of the objects are

1 Introduction known. The purpose is to detect the pixel matching with the
In the 1980s, hyperspectral remote sensing technology spectrum characteristics; traditional detection methods with
was developed on the basis of multispectral remote sensing spectral angle mapping (Spectral Angle Mapping, AM) [13],
imaging technology, and has been highly valued by remote matched subspace (Matched Subspace Detector, SD),
sensing scientists and scholars in the world [1]. orthogonal subspace projection (Orthogonal Subspace
Hyperspectral remote sensing is a multi-dimensional Projection. SP) [14].
information acquisition technology combined with imaging (3) A class of classifications: to find a class and reject the
technology and spectroscopy technology, which has the other class [6].
characteristics of simultaneous detection of two-dimensional The domestic Zhao et al. applied the St OMP algorithm
geometric spatial information and one-dimensional (1-D) and the sparse model algorithm to the hyperspectral target
spectral information [2]. detection and achieved good effect [11]. Liao et al. proposed
Target detection is one of the most important fields in the a Two Steps Reconstruction (SR) method [1]. Song et al.
application of hyperspectral remote sensing technology. presented a combining sparse representation of target
Hyperspectral image contains spatial information and detection method [6]. Wu et al. presented support vector
spectral information in the scene [3]. The spectrum machine (SVM) detection method based on adaptive
information is characterized by the inherent characteristics parameter [4] which achieved good detection results. And
of a substance with different chemical components [4]. target detection algorithms based on spectral uncertainty,
Target detection of HSIs can make use of the spectral-spatial matched filtering correlation (MFC) and manifold
information of the image to greatly improve the ability of compression perception are proposed respectively [13-18].
quantitative analysis of the target and background [5]. The An automatic algorithm for local background subspace
target detection of HSIs can reveal the camouflage, estimation is proposed by S. Matteoli et al. In 2012, P.
concealment and deception of the enemy in the military [6], Bajorski proposed target detection under misspecified
and it can be used to monitor public safety, food safety, models in Hyperspectral Images [19, 20]. Taghipour et al.
quality and safety, etc [7-9]. At the same time, it has great using attribute profiles achieved anomaly detection [21].
potential application value in many fields, such as pollution In recent years, deep learning has achieved good
monitoring [10], interstellar exploration, hydrology, performance in hyperspectral image classification [22] and it
atmospheric monitoring [11]. achieves high detection accuracy in general image target
According to the known specimens, target detection can detection. In this paper, the target detection for specific small
be divided into three forms: targets is realized based on classification and achieve better
(1) Abnormal target detection: pixels are obviously detection results.
different from the background. The most famous ones are
RX and Kernel RX algorithm [12]. 2 Overview of CNN:

The convolution neural network (CNN) [21] is a

×The
feedforward neural network. The artificial neurons can
work was supported by the National Natural Science Foundation
of China (Grant No. 61401244, 61773227) and the Science and Technolo respond to a portion of the surrounding units and have a good
gy Project of Shandong Provincial University (Grant No. J15LN39).

9255
Authorized licensed use limited to: National Institute of Technology. Downloaded on April 11,2025 at 11:25:50 UTC from IEEE Xplore. Restrictions apply.
performance for large image processing. It includes the The corresponding formulas are as follows:
convolution layers and the pooling layers.
1 x1  Wn 2 x2  Wn 3 x3  bn )
2
an f (Wn(1) (1) (1) (1)
(2)
Nowadays, CNN has become one of the research focuses
in many fields, especially in pattern recognition [21]. h ( x) a
w ,b
(3)
1
f (W a  W a  W a  b )
( 2)
11 1
( 2) ( 2)
21 2
( 2) ( 2)
13 3
( 2)
1
( 2)
(3)
Because it can avoid the complex preprocessing of images
and it can directly input original images. Generally speaking, More similar, can be extended to 2,3,4,5, ..., a hidden
the basic structure of CNN consists of two layers, one is the layer. The weights are updated by back propagation
feature extraction layer. The input of each neuron is
connected to the local receptive domain of the previous layer algorithms in the training process.
and the local features are extracted [22]. Once the local
features are extracted, the position relationship between 3 Target detection based on CNN classification
other characteristics can be determined. The other is the
feature mapping network layer, each computing layer is The framework is made up of Blob, Solver, Net, Layer,
composed of multiple feature maps and each feature map is a Proto in this paper and its structure diagram is shown in Fig.
flat plane where all neuron weights are equal. 3.
The feature mapping structure uses the sigmoid function Solver signal Input
Blob
blob
with small kernel function as the activation function, which
Blob
makes the feature maps have the invariance of the Indefinite
dimension matrix˖
displacement. In addition, the number of parameters of the Param N×C×H×W
Train Net Test Net Layer
network is reduced by shared weights. db Blob N:Num
C˖Channel
Each coiling layer in CNN is closely following a io H:Height
computing layer for local average and two times extraction. Blob
W:Width

This unique feature extraction structure reduces the feature Output

Solver Proto
resolution. Blob

The CNN has unique advantages in speech recognition

and image processing due to its special structure with shared Fig. 3: The framework
weights and local connection.
The multi-dimensional image can be directly input into the Solver is responsible for the training of deep network,
network which avoids the complexity of data reconstruction each Solver contains a training network object and a test
in feature extraction and classification. Each unit of the network object. Each network is made up of several Layer.
neural network is as follows: The input and output of each Layer Feature map is expressed
as Input Blob and Output Blob.
x1
A specific network framework consists of an input layer, a
x2 h ( x)
w ,b convolution layer, a convergence layer and a complete
connection layer, as shown in Fig. 3.
x3 Different feature maps can be obtained by convolution.
+1 The data is placed with a different convolution kernel and
each feature map represents a feature.
Fig. 1: Unit of the neural network
Soft-max
The corresponding formulas are as follows˖ layer
Convolutional
h ( x)
w ,b
T
f (W x) f ( ¦ W x  b)
3
i 1 I i
(1) layer1
Convolutional
layer2
Convolutional
layer3
Convolutional
Layer...
Among them, the unit can also be called the Logistic 31
28 6
160
14 12 4 1
...
...
regression model. A neural network model is formed when 4 2 2
3 2 2
2
3 2
multiple units are combined to have a layered structure. The 4 2 80
2 8 3
2 3 2
following figure shows a neural network with an implicit 39 36 18 3 16 40 6 ..
layer. 20
40
60 .
40 60 Max-pooling
20 20 Max-pooling
x 1
α1
(2)
Input
layer
Max-pooling layer 2
layer 3
n
layer 1 Deep hidden identify
features(DeepID)
Fig. 4: Overall CNN framework
x 2 α (2)
2 h w ,b
( x)
α (2) In the pooling layer, the value of the neuron (x, y) in the
3
Layer
x 3 L3
first layer is m feature map.
N1 1 N 2 1
lm
vxy f (¦ ¦ ¦w n1n2
lmp v((lx1)n1p)( y  n2 )  blm ) (4)
p n1 0 n 2 0
+1 +1
Layer Layer
where p indexes the feature map in the ( l − 1) th layer
L1 L2 connected to the current mth feature map is the weights of
Fig. 1: The neural network position (n1, n2) connected to the pth feature map, N1 and
N2 are the height and the width of the convolution kernel,

9256
Authorized licensed use limited to: National Institute of Technology. Downloaded on April 11,2025 at 11:25:50 UTC from IEEE Xplore. Restrictions apply.
and blm is the bias of the mth feature map in the lth layer.
f is the activation function, the commonly used activation
function such as RELU.

7 2 6 0 8 1 2 3 0 1
2 1 2 0 6 2 2 2 0 1
 
7 6 8 2 3 1 (e)  (f)
1 3 1 4 3 2 3 1 2 0
3 4 6 . . . Fig. 6: Original target image
0 3 1 1 2 0 3 1 1 2
3 4 7 . . .
1 0 3 1 7 1 0 3 1 7 HSIs also reflect the spectral information of the target on
Original image Max-pooling Original image Mean-pooling the general image information. Each pixel corresponds to a
Fig. 5: Max-pooling and mean-pooling (with 3×3 kernels and spectral curve, and the spectral curves of the pixels in the
stride 3) same category are similar. The spectral information of each
pixel point is represented by one of the one dimensional
The softmax-loss layer, and the sotfmax-loss calculation formula vectors.
is: Figure 7 is the spectral curve of the pixels corresponding
f yi
e to the four different objects in the original data which is the
L = -¦ y j log p yj Pyi (5)
j ¦ je j plane, the roadside, the roof and the grass. The target
classification detection can be achieved by using the spectral
The overall network of target detection based on CNN is characteristics of the target.
as shown Fig. 4. Firstly, the raw HSI data is preprocessed to But relying solely on the spectral information of the target
extract the spectral information and convert it into the to classify the target is far from reaching the precision we
corresponding data format. Secondly, the preprocessed data want. HSIs can provide spatial information while providing
is input to the network to extract the feature information. target spectral information. So we need to take the spatial
Finally, the target is calibrated according to the extracted information into account.
feature information. First, in order to obtain the spatial information of each
pixel in the image, we extract the target pixels from each
4. EXPERIMENTS AND ANALYSIS band of the original data. Then, the block images of (13×13)
around the target pixels in each dimension are taken as
In order to verify the effect of the experiment, the data
training and testing samples.
acquired from advanced infrared imaging spectrometer
The spectral information of each pixel corresponding to a
AVIRIS (airborne visible infrared imaging spectrometer) is
spectral curve, spectral curve pixels in the same category is
used to do related simulation experiments in this paper [2].
similar. The spectral information of each pixel is represented
AVIRIS is an imaging spectrometer that uses push broom
by a 1-D vector, the vector can be converted into 2 -D
imaging. It can get 224 spatial information of wavelength at a
matrix (12h12) as training and testing samples.
wavelength range of 0.4 to 2.45 nm. The image used in this
0.7
0.6
experiment is part of the Santiago Airport, which covers the
spectrum range from visible to near infrared. The 0.58
0.65
experimental data can be used to simulate the 158 pure bands
reflectance
reflectance

which removing the absorption band of water vapor and the 0.6 0.56

noise band.
0.55 0.54

0.5 0.52
0 50 100 150 0 50 100 150
wavelength wavelength
(a)airplane (b) road
0.6
0.8

0.58
0.6
reflectance

0.56
reflectance

 0.4
(a) (b) 0.54

0.2 0.52

0 0.5
0 50 100 150 0 50 100 150
wavelength wavelength
(c) roof (d) grass
Fig. 7: Spectral characteristic curve
  A total of 36000 small samples are selected and the
(c)  (d)
images containing the pixels of the airplane are marked as
the target and the images that do not contain the pixels of the
airplane are marked as the background.

9257
Authorized licensed use limited to: National Institute of Technology. Downloaded on April 11,2025 at 11:25:50 UTC from IEEE Xplore. Restrictions apply.
 25% samples are randomly selected as test sample, the
remaining 75% are training samples. The network input data
is the lmdb format data which is converted by the jpg data.
And tag marks are 0 and 1, 0 is the target, and 1 is the
background. After the picture subtracted the mean, training
and testing will increase the speed and precision, so we add
the mean file to improve the calculation speed and accuracy.
In the network the number of convolution cores is 5 with
relu as activation function, test iter is 100, momentum is 0.9,
weight decay is 0.004, and max iter is 6000.
Figure 8 (a) shows the relationship between test accuracy
and accuracy-iters and (b) implies the loss-iters line diagram
for spectral information. The accuracy is 92.1%, loss is
0.252 when iteration is 6000 times. 
Figure 9 (a) and(b) are the experimental data curves (a) Accuray-iters
obtained by using spatial-spectral information. According to
graph analysis, the accuracy is 93.7% and is obviously
higher than the accuracy of the spectral information or
spatial information.
Figure 10 (a) is the spectral information train-loss-iters
line diagram and (b) is full information line diagram.
Figure 11 (a-i) are SVM, DBN and CNN three image
visualization methods in the corresponding accuracy.


(b) Loss-iters

Fig. 9: Accuray-iters and loss-iters line diagram


(a) Accuray-iters


(a) Spectral Train- loss-iters


(b) Loss-iters
,
Fig. 8: Accuray-iters and loss-iters line diagram

9258
Authorized licensed use limited to: National Institute of Technology. Downloaded on April 11,2025 at 11:25:50 UTC from IEEE Xplore. Restrictions apply.
 Experimental data show that when the number of
iterations is the same, the measurement accuracy of spectral
information is much higher than that of space information.
According to table 1˖This article uses the group data to do
the same test with another neural network framework DBN.
The test accuracy of DBN network is 91%, which is
obviously superior to the precision of 85% of the traditional
SVM. The experiment proves that the theory of deep
learning has great potential in this field. But the results of the
entire image input are not very well, and some parts are still
not properly calibrated. Through the adjustment of the later
parameters and network optimization, the higher test
accuracy can be achieved.
Conclusion
 In this paper, CNN is applied to the target detection of
(b) Full-information- loss-iters
HSIs on basis of the target classification. A hyperspectral
Fig. 10: Train- loss-iters line diagram image target detection method based on CNN is proposed.
The paper use CNN, DBN two deep learning models and the
Accuracy is precision, that is: traditional SVM method to simulate the real ARIVIS
tp hyperspectral data, and get effective results. The results of
p CNN and DBN are better than that of SVM; And the method
tp  fp of CNN has better results than DBN. But, because the
p is precision of Accuray , tp is the correct test sample, sample selection, parameter setting and the network structure
fp is the false test sample. is not optimal, the results of the experiment are not very well.
Future work will optimize the network parameters to achieve
Table 1: Precision comparison table better experimental results, and try various network
frameworks to achieve hyperspectral target detection.
Method Accuray
SVM 85.1% References
DBN 91.2%
[1] Chen Y, Nasrabadi N M, Tran T D. Sparse Representation for
CNN 93.7% Target Detection in Hyperspectral Imagery[J]. IEEE Journal
of Selected Topics in Signal Processing, 2011, 5(3):629-640.
[2] Borhani M, Ghassemian H. Kernel Multivariate
Spectral–Spatial Analysis of Hyperspectral Data[J]. IEEE
Journal of Selected Topics in Applied Earth Observations &
Remote Sensing, 2015, 8(6): 2418-2426.
[3] Khazai S, Homayouni S, Safari A, et al. Anomaly Detection in
Hyperspectral Images Based on an Adaptive Support Vector
Method[J]. IEEE Geoscience & Remote Sensing Letters, 2011,
8(4):646-650.
   [4] Wu Yiquan, Zhou Yang, Long Yunlin. Small Target
(a) (b)(c) Detection in Hyperspectral Remote Sensing Image Based on
Adaptive Parameter SVM[J]. Guangxue Xuebao/acta Optica
Sinica, 2015, 35(9):0928001.
[5] Zhao CH, Li XH, Wang YL. Research advance on anomaly
detection for hyperspectral imagery[J]. Journal of Electronic
Measurement and Instrumentation, 2014, 28(08): 803-811.
[6] Zhao C, Jing X, Wei L I. Hyperspectral image target detection
   algorithm based on St OMP sparse representation[J]. Journal
(d)  (e)  (f) of Harbin Engineering University, 2015, 36(7):992-996.
[7] Song Y G, Wu Z B, Sun L, et al. A novel spectral-spatial
sparse method for hyperspectral target detection[J]. Binggong
Xuebao/acta Armamentarii, 2014, 35(6):834-841.
[8] Wang H L. Research on Hyperspectral Imaging Algorithm for
Targets Detection[J]. Electro-optics & Passive
Countermeasures, 2004.
[9] Cheng B Z, Zhao C H. A particle swarm optimization
  clustering-based approach for hyperspectral image anomaly
(g)  (h) (i)
targets detection[J]. Journal of Optoelectronics Laser, 2013,
Fig. 11: Visual result diagram 24(10):2047-2054.
[10] Zhao C, Li X, Zhu H. Hyperspectral imaging target detection
algorithm based on spatial 4 neighborhoods for sparse

9259
Authorized licensed use limited to: National Institute of Technology. Downloaded on April 11,2025 at 11:25:50 UTC from IEEE Xplore. Restrictions apply.
 representation[J]. Journal of Harbin Engineering University, Hyperspectral Images[J]. IEEE Geoscience & Remote
2013, 34(9):1171-1178. Sensing Letters, 2016, 13(8):1129-1133.
[11] Wu Chao, Wu Yiquan. Target Detection in Hyperspectral [17] Cheng B. Anomaly targets detection of hyperspectral imagery
Image Using Projection Pursuit Based on Chaotic Particle based on sparse representation[C]. IEEE International
Swarm Optimization[J]. Acta Optica Sinica, 2011, Conference on Computer Communication and the Internet.
31(12):1211003. IEEE, 2016:388-391.
[12] Matteoli S, Acito N, Diani M, Corsini G. An Automatic [18] Tan Jing. Research on the target detection of hyperspectral
Approach to Adaptive Local Background Estimation and imaging information [D]. Huazhong University of Science
Suppression in Hyperspectral Target Detection [J]. IEEE and Technology, 2011.
Transactions on Geoscience and Remote Sensing, 2011, [19] Bajorski P. Target Detection Under Misspecified Models in
49(2): 790-800. Hyperspectral Images[J]. IEEE Journal of Selected Topics in
[13] Jin Tianming. Research on Hyperspectral target detection Applied Earth Observations and Remote Sensing, 2012,.5(2):
based on kernel sparse and spatial constraint [D]. Harbin 470-477.
Institute of Technology, 2016. [20] Xue Q. Optical system design of multi-model hyperspectral
[14] Cheng B, Zhao C, Zhang L, et al. Joint Spatial Preprocessing and imager for spaced-based atmospheric remote sensing[J]. Acta
Spectral Clustering Based Collaborative Sparsity Anomaly Detection Optica Sinica, 2014, 34(8): 0822005-690.
for Hyperspectral Images[J]. Acta Optica Sinica, 2017. [21] Taghipour A, Ghassemian H. Hyperspectral Anomaly
[15] Zhang L, Zhang L, Tao D, et al. Sparse Transfer Manifold Detection Using Attribute Profiles[J]. IEEE Geoscience &
Embedding for Hyperspectral Target Detection[J]. IEEE Remote Sensing Letters, 2017, (99): 1-5.
Transactions on Geoscience & Remote Sensing, 2013, [22] Alam F I, Zhou J, Liew W C, et al. CRF learning with CNN
52(2):1030-1043. features for hyperspectral image segmentation [J], Geoscience
[16] Xu Y, Wu Z, Xiao F, et al. A Target Detection Method Based and Remote Sensing Symposium. IEEE, 2016: 6890-6893
on Low-Rank Regularized Least Squares Model for

9260
Authorized licensed use limited to: National Institute of Technology. Downloaded on April 11,2025 at 11:25:50 UTC from IEEE Xplore. Restrictions apply.