combined with a spatial distance weight, is adopted to determine the
purity of each candidate endmember in the neighborhood window. In
the SSPP method (Martín and Plaza, 2012), the pixel's spectral purity is
characterized by using the Root Mean Square Error (RMSE) (Keshava
and Mustard, 2002) between two images: the original image and a
filtered version obtained by using Multi-Scale Gaussian Filtering
(Young and Van Vliet, 1995). The lower the RMSE, the higher the si-
milarity between a pixel and its neighbors. In Xu et al. (2015),a
Homogeneous Index (HI) is assigned to each candidate. Such HI is
calculated by using the SID criterion between itself and its neighbors,
and the maximum SID value is adopted. In the SSPM method (Kowkabi
et al., 2016b), the values in cluster labels between the candidate and its
neighbors are adopted as the homogeneity criterion, where the cluster
label for each pixel is obtained by the K-means clustering algorithm.
The basic idea of all the above mentioned methods is to perform
spectral similarity measurements within a local neighborhood window
by using a spectral distance metric, such as the Euclidean Distance (ED),
Spectral Correlation Measure (SCM), SAD, and SID. However, there are
some pending issues in those methods. For instance, in the AMEE
method (Plaza et al., 2002), the properties of the spatial kernel (shape
and size) strongly influence the MEI result. Although a progressively
increased kernel size can be adopted, this will heavily increase the
computational burden. The problem of determining an optimal neigh-
borhood size also exists in other methods (Li and Zhang, 2011; Martín
and Plaza, 2012; Xu et al., 2015; Kowkabi et al., 2016b). In addition,
the setting of parameters is a difficult issue for all methods.
For better examining the spatially homogeneous regions in hyper-
spectral images, unsupervised clustering methods have been adopted in
existing spatial-spectral unmixing models. Specifically, RBSPP (Martín
and Plaza, 2011), SSPP (Martín and Plaza, 2012) and SSPM (Kowkabi
et al., 2016b) apply unsupervised clustering algorithms [such as
ISODATA (Ball and Hall, 1965), K-means, or the Hierarchical Seg-
mentation (HSEG) algorithm (Tilton, 2003)] to segment the original or
transformed image [e.g., by using the Principal Component Transform
(PCT)] into a set of spectral clusters, each made up of one or more
spatially connected regions. Compared with the SSEE (Rogge et al.,
2007a), SPEE (Mei et al., 2010), HEEA (Li and Zhang, 2011), LLC
(Canham et al., 2011) and the method in Xu et al. (2015), which all
segment the original image into non-overlapping sub-blocks with a
fixed size and shape, the partitions obtained by clustering algorithms
generally exhibit better smoothness and coherence. However, in most
of the aforementioned clustering algorithms, only spectral information
is considered during the clustering process, and the size of segmented
partitions is subject to diversity and inconsistencies. The SSPP method
(Martín and Plaza, 2012), which includes the spatial information,
however fuses the spatial and spectral information in a separate way.
Therefore, how to naturally combine the spatial information during the
clustering stage is a main difficulty for SSPP. In addition, other relevant
issues such as the computational complexity of clustering algorithms,
the determination of the number of clusters in advance, and the com-
putational e
fficiency
need further discussion.
In recent years, many superpixel-based segmentation methods, such
as Graph-based Algorithms (Felzenszwalb and Huttenlocher, 2004),
Turbopixel method (Levinshtein et al., 2009), and SLIC (Achanta et al.,
2012), have been proposed. In essence, any image segmentation
method can generate superpixels, and the aforementioned methods
show faster speed and more accurate segmentation results than some
traditional methods such as Mean-Shift-based (Vedaldi and Soatto,
2008) and Watershed-based approaches (Vincent and Soille, 1991).
Superpixel methods have also been widely used in the hyperspectral
imaging community; relevant work can be found in Thompson et al.
(2010) and Saranathan and Parente (2016). In hyperspectral images, a
superpixel represents a homogeneous region which contains a number
of pixels that exhibit spatial continuity and spectral similarity. When
superpixels are used to represent spatial correlation information in
hyperspectral images, compared with the fixed size and shape of
window-based methods, they can represent adaptive spatial neighbor-
hoods as expected in natural scenes, which generally exhibit arbitrary
morphologies or sharp boundaries, thus reducing the sensitivity to
noise and outliers and significantly improving the computational effi-
ciency for subsequent processing tasks, such as endmember extraction
and spectral unmixing. In Thompson et al. (2010) and Saranathan and
Parente (2016),anefficient graph-based image segmentation algorithm
(Felzenszwalb and Huttenlocher, 2004), which performs an agglom-
erative clustering of pixels as nodes on a graph such that each super-
pixel represents the minimum spanning tree of the constituent pixels, is
adopted. This method adjusts well to natural image boundaries, but
produces superpixels with very irregular sizes and shapes (Achanta
et al., 2012). In Massoudifar et al. (2014), the well known Ultrametric
Contour Map (UCM) algorithm is extended to hyperspectral images,
conducting comparative experiments on the original hyperspectral
image, the transformed image by PCT, and a monochromatic image,
respectively. The results indicate that superpixel estimation on the
transformed image is generally more efficient than other methods. The
most important aspect of superpixel segmentation methods is how to
determine the number of superpixels. In order to obtain a good super-
pixel representation, over-segmentation is generally adopted. In
Saranathan and Parente (2016), a segment uniformity criterion is pro-
posed to control the segmentation scale, which adopts a threshold to
limit maximum variability inside one segment. The threshold is com-
puted by a statistical analysis of the within-class and between-class
spectral divergences of several endmember classes.
In Saranathan and Parente (2016), Thompson et al. (2010), and
Massoudifar et al. (2014), superpixel segmentation methods have been
shown to succeed when applied to hyperspectral imagery, and pre-
sented competitive results and significant improvements in the sub-
sequent processing steps (particularly from a computational stand-
point). However, the processing time of these superpixel segmentation
methods increases as the size of the images become larger. As a result,
finer segmentations are likely to result in significant computation times
and sensitivity to noise. Based on this observation, in this work we have
developed a new spatial preprocessing approach for hyperspectral un-
mixing. Compared with the aforementioned superpixel segmentation
methods, the proposed approach is faster and requires less memory
space. In addition, its computational complexity is not affected by the
image size. A detailed description of our proposed method is given in
the following section.
3. Regional clustering-based spatial preprocessing
In this section we describe a new method for spatial preprocessing
which consists of two main steps. The
first one is a clustering procedure
which divides the original image into a set of homogeneous partitions.
As opposed to the RBSPP (Martín and Plaza, 2011) and the SSPP
(Martín and Plaza, 2012), we introduce a new efficient clustering
strategy which naturally integrates the spatial and the spectral in-
formation contained in the data. In a second step, we extract candidate
pixels from each partition. After these two steps, the candidate pixels
are gathered together and fed to a spectral-based method to obtain the
final endmembers and their corresponding abundances. An illustrative
flowchart of the method is shown in Fig. 1. A detailed description of
each step is given below.
3.1. Regional clustering
The proposed clustering strategy is inspired by the SLIC algorithm.
The core idea of SLIC is to constrain the search scope from the whole
image to a local region around the centroid. Compared with the SLIC
X. Xu et al.
Remote Sensing of Environment 204 (2018) 333–346
335