Environmental Engineering Reference
In-Depth Information
data sets. The data base is currently available with the regional
atmospheric models used by the French and German weather
services, MESO-NH and COSMO respectively.
Lemonsu
et al
. (2006) present a general methodology of
urban classification for typical North-American cities that is
currently applied to update LULC information in the regional
forecast model of the Meteorological Service of Canada. The
methodology is based on the joint processing of medium-
resolution Advanced Spaceborne Thermal Emission and Reflec-
tion Radiometer (ASTER) and Landsat-7 Enhanced Thematic
Mapper Plus (ETM
+
) imagery, digital elevation models (DEMs)
derived from the Shuttle Radar Topography Mission (SRTM-
DEM) and national elevation DEMs for both the USA and
Canada. A decision tree model is applied to identify urban
LULC classes. First a 15 m resolution unsupervised classifica-
tion of preprocessed remotely sensed data (ASTER or ETM
from the Landsat sensors or MODIS (Woodcock and Strahler,
1987; Small, 2007). Current atmospheric and weather prediction
models operate at grid scales two to three orders of magnitude
larger than the scale of land cover classifications produced from
very high to ultra high resolution data, but as computational
power continues to increase it may soon become possible to
utilize such fine-grained land cover and biophysical information
(such as vegetation indices) directly. At present, however, fine
spatial resolution land cover information can be aggregated to
larger grid scales as model inputs.
The majority of urban and suburban LULC classification
techniques using remotely sensed data have been developed for
use with medium to high resolution (50-4 m/pixel; Ehlers, 2004)
multispectral and hyperspectral data collected by satellite-based
and airborne sensors. Classification approaches include unsuper-
vised/supervised clustering algorithms and fuzzy membership
classifiers (Jensen, 1996; Zhang and Foody, 1998); techniques
that integrate both remotely sensed and ancillary georeferenced
data such as expert systems (Gong and Howarth, 1990; Stefanov,
Ramsey and Christensen, 2001; Wentz
et al
., 2008) and neu-
ral networks (Civco, 1993; Foody, McCulloch and Yates, 1995;
Weng and Hu, 2008); decision trees (Mahesh and Mather, 2003);
image spectral analysis (Lu and Weng, 2006; Rashed
et al
., 2003
Ridd, 1995, Small, 2005; Wu and Murray, 2003), and image
spatial analysis (Bian, 2003; Myint, Lam and Tyler, 2004). Many
of the above-listed techniques, particularly spectral analysis, rely
on multiple bands of information spanning the visible through
midinfrared (or thermal infrared) portions of the electromag-
netic spectrum; this condition is satisfied by most of the current
generation of high to low resolution orbital and airborne sensors.
Direct application of these spectrally-based techniques to
very high and ultra high resolution orbital and airborne data has
been problematic due to the relatively low spectral information
content of these systems, which typically consists of visible red,
green, blue (and/or a panchromatic band in these wavelengths),
and a near infrared band. Some of these spectral limitations
can be overcome by adding other information such as light
detection and ranging (lidar) data (e.g., Tooke
et al
., 2009), or
extracting spatial information from the fine resolution data such
as texture (e.g., Su
et al
., 2008) to provide additional infor-
mation for LULC classification. Newer satellite-based systems
such as DigitalGlobe's WorldView-2 (launched in 2009) and the
RapidEye constellation of satellites (launched in 2008) exceed
the multispectral capabilities of previous very high to ultra high
resolution sensors by adding addition bands in the visible and
near-infrared wavelengths (Ehlers, 2009). The site revisit times
of these new sensor systems are also reduced to 1-4 days by
off-nadir pointing capability (WorldView 2) and multiple satel-
lite platforms (RapidEye). An extensive review of existing and
planned Earth observing satellite systems is presented in Petrie
and Stoney (2009).
A relatively recent classification approach known as object-
based image analysis (OBIA) takes advantage of the high spatial
information content of very high to ultra high resolution data,
and is particularly well-suited for land use and land cover clas-
sification of urban and suburban areas (Banzhaf and Netzband,
2004; Blaschke andHay, 2001; Blaschke, Burnett and Pekkarinen,
2006; Hofmann
et al
., 2008; also see Chapter 7, this volume).
OBIA uses image segmentation to identify homogeneous image
objects at several different scales, rather than the classical pixel-
based (and single-scale) classification approaches. A particular
pixel included in an image segment might have simultaneous
)
is performed to obtain 11 generalized urban/non-urban classes.
Building heights at 15 m spatial resolution are determined based
on the difference of SRTM-DEM and the national DEMs, and
used as additional criteria to aggregate the 15 m classifications to
a 60 m resolution fraction of natural surface and urban cover.
The decision-tree model is then applied to derive twelve urban
LULC classes: high buildings, mid-high buildings, low buildings,
very low buildings, sparse buildings, industrial areas, roads and
parking, roadmix, dense residential, medium-density residential,
low-density residential, mix of natural and built. This methodol-
ogy was refined by Leroux
et al
. (2009) to create a fully automated
geospatial database processing approach for generation of urban
LULC. Designed for use with Canadian cities the LULC output
can be used directly in mesoscale atmospheric models. The auto-
mated system incorporates vector-based land use and land cover
data, SRTM-DEMs, Canadian Digital Elevation Data DEMs, and
census data to derive urban LULC classifications. Such an auto-
mated systemmay be applicable in other countries where similar
data are available.
For the American community Weather Research and Fore-
casting (WRF) model (Skamarock
et al
., 2005) Moderate Reso-
lution Imaging Spectroradiometer, or MODIS, 1-km global data
were made available recently by Boston University (Friedl
et al
.,
2002). Those data are classified according to a 20-class LULC
classification developed by the International Global Biosphere
Program (Lambin and Geist, 2006). Since the classification is
based on 2001 images the representation of urban areas in terms
of extent has been improved. However, only one urban land use
class is considered (''urban built-up''). For WRF applications to
urban areas users are currently still encouraged to provide their
own updated urban LULC dataset.
The current generation of very high (1-4 m/pixel) to ultra
high (
<
1 m/pixel) spatial resolution sensors on board Earth-
observing satellites (e.g., IKONOS, Quickbird, Orbview) has
enabled the development of sophisticated approaches to classifi-
cation of urban LULC and measurement of biophysical variables
(also see chapters in Part III of this volume). Very high to ultra
high resolution multispectral data is now available for numerous
urban centers around the world. These data provide spatial res-
olution similar to or better than traditional aerial photography,
but with increased spectral information and geometric accuracy.
Urban/suburban land cover classes such as residential lawns,
driveways and parking lots, and even individual tree canopies
can be fully resolved in these data. This allows avoidance of sub-
pixel spectral mixing of built and natural surface materials that
is a problem with coarser resolution data such as that obtained
+
Search WWH ::
Custom Search