Environmental Engineering Reference
In-Depth Information
●
Inset 3 contains single-family dwellingunits located in sparsely
vegetated area,
●
Inset 4 contains both single-family and multifamily dwelling
units located in a sparsely vegetated area.
13.3.2
Building extraction
Automatic extraction of buildings from remote sensing data
has become more common over the years, particularly with
the availability of high spatial resolution remote sensing images
and lidar. In order to select an appropriate building detec-
tion method for population estimation, four building extraction
methods were tested in this study. These methods were based on
region-growing segmentation (Zhang, Yan and Chen, 2006), the
Hermite transform (Silvan-Cardenas and Escalante-Ram'irez,
2006), the Dempster-Shafer theory of evidence (Shafer, 1976;
Lu, Trinder and Kubik, 2006), and Definiens' eCognition seg-
mentation method. Details can be found in Silvan-Cardenas
et al
. (2009). As a result, building footprints as well as building
volumes were generated.
Datasets acquired for the study area includes lidar altime-
try measurements, demographic and geographic census data,
building footprints and land use layers, high resolution aerial
photography and a Landsat TM image. All datasets were contem-
porarily acquired around year 2000. The lidar data was acquired
in 2000 using an Optech Inc. Airborne Laser Terrain Mapper
(ALTM) 1225 instrument mounted on a single-engine craft.
The ALTM instrument delivers a cloud of three-dimensional
(3-d) points for the first and last return of a laser pulse. There
were around 40 million points in the study area, for which
the backscatter intensity was also available from each return.
Demographic and geographic data were acquired at census block
level through the US Census Bureau's American FactFinder (US
Census Bureau, 2009a) and TIGER/Line shapefiles (US Cen-
sus Bureau, 2009b) web sites, respectively. Building footprint
and land use layers, together with a high-resolution (60 cm,
2 ft) color-infrared (CIR) aerial photography, were acquired
through the City of Austin Neighborhood Planning and Zoning
Department (NPZD, City of Austin, 2009b).
13.3.3
Land use classification
Population counts have a close relationship with land use type.
Given the premise that population is only counted in residential
areas, an immediate need will be to single out residential land use
type from other land use types. Within the category of residential
land use, it is also necessary to separate different land use types
such as single family, multi-family, etc., since they are associated
with different population densities.
Conventional per-pixel classification may not be adequate
for the urban land use classification (Jensen and Cowen, 1999).
This is particularly true when high spatial resolution data, such
as the CIR photography and lidar, is employed in our analyses.
Therefore, the ''per-field'' approach appears more promising
(Pedley and Curran, 1991; Aplin, Atkinson and Curran, 1999;
Erol and Akdeniz, 2005). This method classifies land use/cover
by predetermined field boundaries with the premise that each
field pertains to a single, homogeneous class. In this study, we
adopted tax parcels as the basic mapping unit. A Meta classifier
was chosen for conducting the land use classification. The Meta
classifier transforms a multiclass problem into several binary
problems (Ichino, 1979). The Meta method was applied to clas-
sify tax parcels into nine land use classes: single family (SF),
multi-family (MF), commercial, office, industrial, civic, open
space, transportation, and undeveloped. Later, the nine classes
were subsequently aggregated to three land use classes: SF, MF,
and non-residential land use. The parcel attributes were derived
from parcel boundaries (area, perimeter and shape), neighbor-
ing parcels (distance to nearest parcel, similarity of area between
parcel and the nearest parcel, similarity of perimeter between par-
cel and the nearest parcel, and similarity of shape between parcel
and the nearest parcel, where similarity was defined as a nor-
malized absolute difference), NDVI zonal statistics (average and
standard deviation), vegetationmask (percent of vegetation cover
in parcel), land cover proportions (percent of impervious sur-
face, percent of bare ground, percent of other land cover type),
building footprint layer (number of buildings in parcel, fraction
of building cover in parcel, average, minimum, and maximum
statistics of area, perimeter, shape, volume and height of buildings
within the parcel), neighboring buildings (average,minimumand
maximumof distance to nearest neighbors) and building density.
This method is hereafter referred to as the MultiClass-tax parcel
method.
13.3
Methodology
This study involved development and application of a number of
methods, including: (1) data preprocessing, (2) building extrac-
tion, (3) land use classification, (4) population estimation and
(5) accuracy assessment. Due to space limit, methods 1-3 are
briefly introduced below while methods 4-5 are presented in
more detail.
13.3.1
Data preprocessing
For the purpose of building extraction and land use classifi-
cation, several raster layers were derived from the lidar point
cloud and the CIR photograph. The layers derived from the
CIR photograph included the following masks: vegetation, bare
ground,impervioussurface,andpervious,non-bareground.The
datasets derived from lidar consisted of the feature height, ground
mask, intensity difference, and the gray level co-occurrence
matrix (GLCM) angular second moment of a digital surface
model (DSM). The DSM was produced at a spatial resolution
of 1 m by applying a point to raster conversion tool to the
point cloud elevation values associated with the last return of
the lidar data. Then a ground mask was produced using the
multi-resolution ground filtering approach developed by Silvan-
Cardenas and Wang (2006). A bare earth digital terrain model
(DTM) was generated by utilizing the ground mask. The recov-
ered DTM was then subtracted from the DSM to produce a
feature height layer.
A reference building footprint layer was derived by editing the
NPZD's building footprint layer, which was produced by manual
digitization using aerial photos and lidar datasets collected in
2003 (City of Austin, 2009b). A reference land use layer was
obtained from the NPZD at city of Austin.
Search WWH ::
Custom Search