GAS PLUME DETECTION IN HYPERSPECTRAL VIDEO SEQUENCE
USING TENSOR NUCLEAR NORM
Wenting Shang
1
, Zebin Wu
1
, Jie Wei
1
, Yang Xu
1
, Ling Qian
1
, Zhihui Wei
1
,
Jocelyn Chanussot
2
, Andrea L.Bertozzi
3
1
School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing
210094, China
2
GIPSA-Lab, Grenoble Institute of Technology, France
3
Department of Mathematics, University of California Los Angeles (UCLA), USA
ABSTRACT
Gas plume detection (GPD) of Hyperspectral video
sequences (HVSs) has become a hot topic in the field of
remote sensing. The traditional HVS processing methods
reshape the extracted video to a 2-D matrix, which is at
expense of destroying spatial or spectral structure. In this
paper, we propose a novel method of Multi-feature Tensor
Decomposition (MTD), where the 3-dimensional (3-D)
structure of the extracted video can be seen as a 3-order
tensor, thus the spatial and temporal structures in HVS are
preserved. We employ the tensor nuclear norm to model the
low-rank property of the background, and apply tensor
sparse norm to constrain the sparsity of the gas plume.
Moreover, taking into consideration the continuity in both
spatial and temporal domain of the gas plume, we add a 3-D
total variation regularization (3DTV) in the proposed
detection model, and assume the support of the gas plume in
different features are the same. The final objective function
of gas plume detection is efficiently solved by augmented
Lagrangian multiplier algorithm (ADMM). Experimental
results demonstrate the effectiveness and high detection
accuracy of the proposed method.
Index Terms— Hyperspectral video sequences
(HVSs), Multi-feature Tensor Decomposition(MTD), low-
rank, sparsity, 3-D total variation(3DTV).
1. INTRODUCTION
Standoff detection of chemical clouds is necessary to many
military and civilian applications. In fact, most gases are in
This work was supported in part by the National Natural Science
Foundation of China under Grant No. 61772274, 61471199, 61701238,
91538108, 61671243, 11431015, the Fundamental Research Funds for the
Central Universities under Grant No.30917015104, the Jiangsu Provincial
Natural Science Foundation of China under Grant BK20170858, the
Jiangsu Province Six Top Talents Project of China WLW-011.
*Corresponding author. Email: zebin.wu@gmail.com
restrained portion of the long-wave infrared (LWIR) domain
[1]. It is necessary of a fine sampling of the electromagnetic
spectrum. Real-time Hyperspectral video sequences HVS
therefore appears as the most suited type of imagery for
such tracking applications. However, the commonly used
processing method would destroy the original spatial or
spectral of HVS.
As each frame of the HVSs is a HSI, every frame is a
high dimensional dataset. Due to the strong coherent of
different HSI bands, the redundancy features need to be
removed. Principal Component Analysis (PCA) is a
common preprocessing to reduce the dimensions and
preserve the useful components.
Many methods have been proposed for gas plume
detection of HVSs, such as [2] that separate the plume from
the background through performing clustering of spectral
data, which exhibits more spatial continuity. Tochon et al. [3]
use the temporal redundancy between two consecutive
frames with the position of the plume is first estimated. The
temporal continuity of HVS is discovered. However, the
spectral characteristic is not fully considered. In order to
retain both of the spatial and spectral structures, we propose
a new method Multi-feature Tensor Decomposition(MTD),
which employs the tensor to take the advantage of its multi-
dimensional structure. Here, the single channel video is
considered as a 3-order tensor, and the single channel video
is composed of the background and gas plume. The data of
the background part have a low-dimensional character, so
we can use tensor low rankness to define the structure. Since
only a few fraction of the pixels belongs to the gas plume,
the left part characterized by the tensor sparse term. In
addition, 3-D total variation (3DTV) regularization for video
can explicitly describe the gas plume continuity in both the
spatial and temporal directions, so we add it to constrain the
gas plume. Although, different features have different
ability in gas plume detection, the position of the gas plume
in the HVS is the same for all features[4]. Thus, the detected
gas plume should have the same support in the result. A
joint sparsity constraint is utilized which can effectively fuse
all the extracted features’ detection result.