extended morphological kernels for analyzing urban images
of Washington, D.C. Myint
et al
. (2011) compared pixel and
object based approaches for urban planning.
Another alternative to the assignment of a pixel to a single
class is to calculate the fractions of various land covers within
the pixel by spectral un-mixing analysis. Wu and Murray
(2003) performed spectral un-mixing analysis to calculate
VIS
fractions for a metropolitan area in Ohio for ecological stud-
ies (Wu
et al
., 2003). Yang
et al
. (2003) developed regression
model between fractions of urban classes and Landsat band
values for urban mapping of western Georgia (Yang
et al
.,
2003). Weng
et al
. (2008), one of the notable
HS
studies of an
urban area, used Principal Components for dimensionality re-
duction and then un-mixing analysis to calculate
VIS
fractions
(Weng
et al
., 2008). In their work,
HS
data provided more ac-
curate results as compared with multispectral data: especially
in case of low reflectance surfaces such as asphalt roads. They
attributed improvement in accuracy to large number of bands
in
HS
data.
Motivation and Objectives
Hyperspectral (
HS
) data provides a unique opportunity for
detecting various urban materials remotely and is consid-
ered more useful for urban mapping and ecological studies
(Gamba, 2013; Rogan
et al
., 2004; Plaza
et al
., 2009).
HS
sen-
sors collect data over narrow spectral bands contiguously and
preserve more accurate record of the energy matter interaction
(Vane
et al
. 1985; Goetz
et al
., 1985). However, the hyperspec-
tral data has not been used for urban studies as extensively
as it is used for minerals, vegetation, and environmental
monitoring studies (Xu
et al
., 2007; Weng, 2012; Shafri
et al
.,
2012). Most of the
HS
studies for urban applications are per-
formed using airborne
AVIRIS
sensor (Ridd
et al
., 1992; Ridd,
1995;Hepner
et al
., 1998; Chen
et al
., 2001; Platt
et al
., 2004),
and only a few using a space-borne sensor (Xu
et al
., 2007;
Weng
et al
., 2008). A possible reason for under-exploration
may be the effort needed for the intensive activity of spectral
library creation. The library using field spectrometer is effort
intensive (Deshpande
et al
., 2013) and creation of spectral
library using image derived spectra needs accurate reflectance
calibration methods.
The earlier studies using
HS
data for urban area do not
correct images for atmospheric effects (Ridd
et al
., 1992;
Ridd, 1995; Hepner
et al
., 1998; Chen
et al
., 2001). Though
atmospheric effects do not affect the classification accuracies
by large margins and are not mandatory in certain conditions,
the nature of hyperspectral data requires sound atmospheric
correction method for accurate results. The results by Song
et
al
. (2001) have misled some of the later studies into using
HS
data without any atmospheric correction. For example, Xu
et
al
. (2007) incorrectly use the conclusions by Song
et al
. (2001)
and do not correct space-borne hyperspectral image for atmo-
spheric effects (Xu
et al
., 2007). Another important reason for
ignoring atmospheric effects is lack of
in situ
measurements
(Yang
et al
., 2003). Additional atmospheric data such as opti-
cal depth, water vapor content, and aerosols are required for
physics-based correction methods, and it may not be available
easily for all metropolitan regions. Image-based atmospheric
correction methods are popular in multispectral data and
commonly used in hyperspectral detection of minerals. Effec-
tive and simple-to-use image-based atmospheric correction
method for hyperspectral detection of various vegetation, im-
pervious and soil surfaces in urban environment is currently
missing. There is limited research in this regard. Compara-
tively, most of the empirical methods such as
IAR
and
FAR
are
extensively studied for mineral mapping (Clark
et al
., 2002).
The primary objective of our research is to develop image-
based reflectance calibration methodology for classification
of urban Vegetation-Impervious-Soil (
VIS
) (Ridd, 1995) classes
and to assess its accuracy by comparing with results from
physics-based methods. Especially, identification of suitable
flat fields in urban area and comparative assessment of two
image-based methods.
IAR
and
FAR
, with physics-based meth-
ods are two important goals of this research. The motivation
is to come up with simple-to-use, yet effective image-based
atmospheric correction strategy for urban land use/land cover
monitoring. Our second objective is to examine whether
hyperspectral signatures of mixtures of
VIS
in different pro-
portions are effective in defining some of the urban land use
classes. Some of the specific research questions are:
• How
IAR
and
FAR
methods compare with each other
and with physics-based method such as 6SV (Vermote
et al., 2006)?
• What are the suitable flat fields in urban places?
• Which is the best flat field providing better classifica-
tion accuracy?
• Whether hyperspectral data is suitable for monitoring
sub-classes within
VIS
and at what granularity?
• Can we use average signature over a group of
VIS
pixels
in certain proportions as a surrogate definition for land
use classes such as low-economy residential, up-market
residential, industrial?
VIS
classification schema provides a unique flexibility to
apply it to either land cover or land use classification exclu-
sively. Spatial configuration of land covers in a given area as
a definition of land use class has been introduced by a few
earlier studies (Phinn
et al
., 2002; Wu
et al
., 2003; Barnsley
et
al
., 1996).
VIS
land covers at broader class granularity are good
candidates to begin this investigation. The potential of
HS
data to detect large variety of urban materials provide added
advantage for such approaches. Further,
VIS
classification
schema is more suitable for ecological studies as the classes
represent unique ecological function.
We performed various calibration experiments using
image-based and physics-based methods on recently acquired
EO-1
-Hyperion image (
USGS
, 2013a) and compared their
VIS
classification results. Further, we compared the discrimina-
tion of
VIS
classes using reflectance calculated using
IAR
,
atmospheric correction formula for clear atmosphere (
USGS
,
2011), and 6SV radiative transfer code (Vermote
et al
., 2006).
We begin further discussion with a concise overview of
physics-based methods and image-based methods in the
Reflectance Calibration Section. We then provide detailed
overall methodology and evaluation procedure followed in
this work in the Classification Section. Further, we follow
with a comprehensive analysis of the results in the Results
and Discussion Section. We conclude with research findings
and useful recommendations for
LULC
classification in urban
area using hyperspectral data.
Reflectance Calibration
Conversion of radiance values to reflectance values (re-
flectance calibration) is a fundamental step in the analysis
of hyperspectral data (Kruse 1994; Schowengerdt, 1997).
Reflectance signature can be used to study the unique spectral
response of a given material, and to identify the target mate-
rial by searching a match in a spectral library. Of the two
predominant methods, physics-based methods model absorp-
tion and scattering by various gases and aerosols to determine
atmospheric effects, whereas empirical image-based methods
use radiance information within the image to model atmo-
spheric effects. Image-based calibration methods provide
practical advantage over physics-based methods, as most of
the required information is available within the scene (Gao
et
al
., 2009; Chavez, 1996; Griffin
et al
., 2003).
366
May 2017
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
