Urban Land Use/Land Cover Discrimination Using
Image-Based Reflectance Calibration Methods
for Hyperspectral Data
Shailesh S. Deshpande, Arun B. Inamdar, and Harrick M. Vin
Abstract
Irrespective of substantial research in land use/land cover
(
LULC
) monitoring of urban area, hyperspectral data is not
yet exploited effectively because of lack of local spectral
resources and a practical reflectance calibration method. The
objective of this research is to develop an effective methodolo-
gy for urban
LULC
classification using image-based reflectance
calibration methods: especially Vegetation-Impervious-Soil
classes (
VIS
), using hyperspectral data. We used
EO
-1 Hyper-
ion image of Pune City, India and assessed the suitability
of different land covers as reflectance calibration surfaces.
Furthermore, we performed
LULC
classification using differ-
ent reflectance calibration methods such as Internal Area
Relative Reflectance, Flat Field Relative Reflectance, and 6S
for comparative analysis. Urban
VIS
signatures extracted from
Hyperion image show distinct spectral curves at broader level.
Flat Field Relative Reflectance method provides above 90
percent average overall accuracy. An advanced physics-based
method such as 6S does not provide any added advantage
over image-based calibration methods.
Introduction
The impact of urbanization and its local and global conse-
quences are well recognized (Arnold Jr.,
et al
., 1996; Turner
et
al
., 1995; Lambin
et al
., 2001; Esch
et al
., 2009). Remote sens-
ing has been one of the important tools to monitor land use
land cover (
LULC
) in urban areas (Jensen, 1993; Jensen
et al
.,
1999; Gamba, 2013); and over the years, it has been used with
multiple perspectives from mapping, planning, to ecological
studies of urban and semi-urban places (Pickett
et al
., 2001).
The urban places are one of the most complex ecosystems
and pose many challenges for use of remote sensing. Whereas
many researchers advocate use of very high spatial resolution
imagery for
LULC
mapping (Jensen
et al
., 1999; Thomas
et al
.,
2003; Myint
et al
., 2011), high spatial resolution is not syn-
onymous to improved results (Barnsley
et al.
, 1996; Gamba,
2013). Considering the heterogeneity of
LULC
in urban area
and equally heterogeneous goals for the analysis, integrated
use of spectral and spatial multi-resolution imagery is desired
(Plaza
et al
., 2009; Shafri, 2012; Gamba, 2013).
Previous Research in Urban LULC
Though land use and land cover are related to each other
and have been used in the literature interchangeably, remote
sensing can help observe the surface properties of a given area
directly. On the other hand, apart from a few, land use classes
are to be identified using supplementary data (Anderson
et al
.,
1976; Turner
et al
., 1995; Gamba, 2013). Inferring land use pat-
tern (such as economic zones) from remote sensing is a long
standing problem (Comber
et al
., 2012). Most of the urban
studies use individual classification schema or modified
USGS
classification schema (Anderson
et al
., 1976; Cadenasso
et al
.,
2007) and consider land use land cover in integrated manner.
Maximum Likelihood (
ML
) has been one of the most com-
monly used classifier for urban
LULC
classification for both
medium and high spectral resolution remote sensing data
(Weng, 2012; Platt and Goetz, 2004; Platt
et al
., 2004) use
ML
to classify urban fringe area at Fort Collins, Colorado. They
used the modified level 2
USGS
classification schema (Ander-
son
et al
., 1976). They observed “modest but real” improve-
ment in classification accuracies by hyperspectral (
HS
) data
over multispectral data. Analysis of results indicated the im-
provement because of large number of bands rather than sig-
nal to noise ratio. Tan and Wang (2007) and Tan
et al
. (2007)
used
ASTER
and Chris-Proba data for urban mapping of Beijing
City, China. Similarly, Spectral Angle Mapper (
SAM
) has been
spectral similarly measurement and classification technique of
choice for
HS
data (Goetz
et al
., 1985; Kruse
et al
., 1993;Yuhas
et al
., 2001). Various recent urban mapping studies using
HS
data deploy
SAM
for measuring similarity between reference
spectra and target spectra and subsequent classification (Hep-
ner
et al
., 1998; Chen
et al
., 2001; Divya
et al.
, 2014). Hepner
et al
. (1998) used
SAM
for classifying
AVIRIS
image to enhance
the radar data analysis, and Chen and Hepner (2001) used
SAM
for imaging spectroscopy studies of Park City, Utah.
In addition, researchers have investigated other supervised
or semi-supervised machine learning techniques for urban
HS
data classification also. Researchers have used Artificial Neural
Network (
ANN
) (Zurada, 1992), and Support Vector Machine
(
SVM
) (Suykens
et al.
, 1999) classifiers with reasonable success
for various urban studies. Ridd
et al
. (1992) used
ANN
to clas-
sify
AVIRIS
and
SPOT
data to study urban morphology and
VIS
configuration of Salt Lake City, Utah. Herold (2002) computed
landscape metrics for studying urban land use structure of
Santa Barbara, California using Ikonos images (Herold
et al
.,
2002). Benediktsson
et al
. (2005) devised extended morpho-
logical kernels for analyzing a Pavia dataset (Benediktsson
et al
., 2005). Fauvel
et al
. (2008) further integrated
SVM
with
Shailesh S. Deshpande is with the Tata Research Development
and Design Centre (TRDDC), a division of Tata Consultancy
Services, 54-B Hadapsar Industrial Estate Pune 411013 India,
and Centre of Studies in Resource Engineering (CSRE), Indian
Institute of Technology (IIT), Bombay, Powai Mumbai 400 076.
Arun B. Inamdar is with Centre of Studies in Resource
Engineering (CSRE), Indian Institute of Technology (IIT),
Bombay.
Harrick M. Vin is with Tata Research Development and
Design Centre (TRDDC).
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 5, May 2017, pp. 365–376.
0099-1112/17/365–376
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.5.365
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2017
365
