2018 Fifth International Workshop on Earth Observation and Remote Sensing Applications
978-1-5386-6642-5/18/$31.00 ©2018 IEEE
Change Detection Combining Spatial-spectral
Features and Sparse Representation Classifier
Qiong Ran
*
College of Information Science
and Technology
Beijing University of Chemical
Technology
Bejing, China
ranqiong@mail.buct.edu.cn
Shizhi Zhao
College of Information Science
and Technology
Beijing University of Chemical
Technology
Beijing, China
Wei Li
College of Information Science
and Technology
Beijing University of Chemical
Technology
Beijing, China
Abstract—In this paper, we propose a spatial-spectral one-
class sparse representation classifier (OCSRC) method to solve
the multi-temporal change detection problem for identifying
disaster-affected areas. The OCSRC method is adapted from the
classical multi-class sparse representation classifier (SRC) from
an earlier work. Based on the spectral based OCSRC, the
spectral-spatial OCSRC is brought up by applying the spatial-
spectral features to the one class sparse representation process
instead of the original spectral bands. The spectral-spatial
features discussed in this paper includes Gabor filter, adaptive
weighted filter (AWF) and collaborative representation filter
(CRF). These features are calculated from the original image
with a convolution process to combine the information from the
neighboring pixels. Performances of OCSRC with these three
features and original spectral feature are tested and compared
with multi-temporal multispectral HJ-1A images acquired in
Heilongjiang province before and after the flood in 2013, with
detailed discussion with two sub-images and massive application
with the entire image. Receiver-operating-characteristics (ROC)
curve, which is widely used to evaluate accuracy for two class
problems such as target detection, is employed to evaluate the
results. It shows that OCSRC combined with spatial and
temporal characteristics outperform the cases with only spectral
feature by a lower false positive rate (FPR) at defined true
positive rate (TPR), namely less detection errors, and lead to
better change detection result.
Keywords—Change detection, one class sparse representation
classifier, sparse representation, spatial-spectral features, disaster
monitoring
I. INTRODUCTION
Detection for changes of the land surface has become a
necessity since landscape is always changing due to human
activities and natural forces. Detection of changes can be done
with two or more registered remote sensing images of the same
area at different times. Change detection have many important
applications, such as locating disaster affected areas, detecting
oil spill, etc.
The change detection task can be converted to a
classification problem by defining the change as a class. Thus
many classification techniques can be used in change
detection[1], including the classic k-means clustering,
supervised and unsupervised multiple support vector machines
classifiers. The combined use of several classifiers is also
conducted, by adopting post-class change detection means
under an iterative Bayesian-Markovian framework, or
exploiting the merits of different classifiers for an improved
performance[2].
Sparse representation (SR) [3] has been widely used in
remote sensing image classification, in this case a testing pixel
is sparsely represented as a linear combination of all labeled
samples via ℓ0- or ℓ1-norm regularization, then the
representation residual with the samples from each class is
calculated. The pixel is then classified to the class with the
smallest residual. Considering the changed areas as one and the
only concerned class, the one-class sparse representation-based
classifier (OCSRC) has previously been proposed for change
detection [4]. The above paper uses only the spectral features,
however, the spatial information can’t be ignored in image
classification. many spatial-spectral feature classification
methods have been proposed [5].
In this paper, we discuss Gabor filter, adaptive weighted
filter(AWF) and collaborative representation filter (CRF). The
spatial features first extracted, then applied to one class
classification.
II.
METHODS
A. OCSRC
With the one class classification concept, the changed area
is the only concerned class. By correctly classifying the