Soft Computing
https://doi.org/10.1007/s00500-020-04697-y
METHODOLOGIES AND APPLICATION
A rough-GA based optimal feature selection in attribute profiles
for classification of hyperspectral imagery
Arundhati Das
1
· Swarnajyoti Patra
1
© Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract
Morphological attribute profiles are robust in capturing the spectral–spatial information of hyperspectral imagery. To incor-
porate maximum spatial information, generation of a profile using multiple attributes with large number of threshold values
is a well-known approach. Although the profile contains very rich spatial information, at the same time its dimensionality
increases. This raises two critical problems for hyperspectral image classification: (i) curse of dimensionality and (ii) com-
putational complexity. To mitigate such problems, the only supervised feature selection technique that exists in the literature
is computationally demanding. In this article, a fast supervised feature selection technique by exploiting rough set theory
and genetic algorithms is proposed. Our technique computes the relevance and significance of each feature in the profile
using rough set theory. Then, based on the relevance and significance values a novel fitness function of genetic algorithms
is designed to select an optimal subset of features from the constructed profile. To show the effectiveness of the proposed
technique, it is compared with the existing state-of-the-art technique by considering three real hyperspectral data sets.
Keywords Hyperspectral images · Mathematical morphology · Attribute profiles · Feature selection · Genetic algorithms ·
Rough sets · Support vector machines
1 Introduction
Hyperspectral imagery (HSI) is capable of capturing spec-
tral information in hundreds of narrow spectral channels
(Chang 2007). The ability to represent a target object by
such a huge spectral variation makes HSI better for char-
acterization of different land-cover objects as compared to
the multispectral images. Over the last decade, researchers
devoted great effort to the classification of HSI for numerous
applications like urban area management, forest monitoring,
soil moisture estimation, crop monitoring, wetland mapping,
etc. (Khan et al. 2018). However, the classification of HSI is
still a challenging task because of two main reasons. First,
as the adjacent bands of HSI are highly correlated it retains
huge redundant information that not only increases computa-
Communicated by V. Loia.
B Swarnajyoti Patra
swpatra@tezu.ernet.in
Arundhati Das
arundha@tezu.ernet.in
1
Computer Science and Engineering Department, Tezpur
University, Assam 784028, India
tional complexity but also introduces curse of dimensionality
problem (Hughes 1968). Second, since the scales, shapes,
and geometries of the different objects are heterogeneous in
nature, proper exploitation of spatial information plays an
important role in the HSI classification. In the literature, the
first issue is addressed by adapting feature selection or fea-
ture extraction techniques that reduces the dimensionality
of hyperspectral imagery by removing redundant informa-
tion (Fan et al. 2019; Fang et al. 2019; Feng et al. 2016;
Shang et al. 2018; Wang et al. 2018; Yuan et al. 2017).
The second issue is addressed by fusing spectral and spa-
tial information (Bhardwaj et al. 2019; Fauvel et al. 2013;
Ghamisi et al. 2017). A large number of algorithms have
been developed to fuse spectral and spatial information from
HSI data, including Markov random field (MRF) modelling
(Tarabalka et al. 2010), segmentation (Ghamisi et al. 2014b),
edge preserving filtering (Kang et al. 2014), Gabor feature
extraction (Zhu et al. 2015), deep learning (Chen et al. 2014),
dictionary learning (Soltani-Farani and Rabiee 2015), object-
based image analysis (Blaschke 2010), etc. An alternative
widely used approach to fuse spectral and spatial informa-
tion is based on mathematical morphology (Benediktsson
123