Environmental Engineering Reference
In-Depth Information
et al
., 2003). As a further comparison, the SMA method per-
forms slightly better than the RT model. Comparing with the RT
model, the SMA method has a slightly lower estimation error
(e.g.
RMSE
pixel-based classifications, the object-based approach using the
nearest neighbor classifier requires fewer training samples since
one sample object includes several to many typical pixel samples
and their variations (Baatz
et al
., 2004).
=
=
10
.
7% and
MAE
6
.
7%) and higher model-fit
(
R
2
0
.
90). Moreover, Fig. 17.6(b) indicates the RT method
may be problematic when applied in areas with low impervious
surface factions (e.g.,
<
40%) because of the confusions between
impervious surfaces and bare soil. Finally, the integrated SMA
and RT model outperforms the individual SMA and RT models,
as it has the lowest
RMSE
(9.69%) and
MAE
(6.05%), and the
highest correlation coefficient (0.92). This may be because the
integrated model partially addresses the endmember selection
problem in the SMA model and the variable selection issue in the
RT approach.
=
17.4.2
Object-oriented classification
Three major steps were involved in the impervious surface
mapping procedure. First, object-based classification was applied
to the Quickbird imagery to classify the data into five general
land cover classes: impervious surface, forest, water, non-forested
rural, and shadow. In the next step of the classification, shadow
was further classified to impervious or non-impervious area using
class-related features. Lastly, final refined impervious maps are
generated by combining the ''shadowed'' imperious areas with
the classified impervious class.
Segmentation is the first and most important process of
the eCognition™-based object-oriented classification. It extracts
image objects at modifiable scale parameters, single layer weights,
and themixing of the homogeneity criterion concerning color and
shape. In other words, the outcome of segmentation is directly
related to several adjustable criteria - scale parameter, color, and
shape - defined by users. In particular, the scale parameter is
a measure for the maximum change in heterogeneity that may
occur when merging two image objects. A larger scale parameter
value leads to bigger objects and vice versa. Adjusting the shape
factor will also affect the overall fusion value computed based on
both spectral heterogeneity and the shape heterogeneity. There-
fore, how to choose an optimal scale parameter and shape factor
is critical to the quality of classification. To set the appropri-
ate objects for use and to evaluate how classification accuracy
changes when adjusting these two criteria, different scale param-
eters ranging from 5 to 50 and shape factors ranging from 0.1 to
0.9 were tested.
Besides the scale parameter and shape factor, the result of
object-oriented classification is also dependent on the object-
based metrics, including measures of texture, length, and shape
as well as measures of spatial relationships to other super-, sub-,
and neighboring objects, utilized in the classification process. In
this study, two extra object-based metrics - ''Ratio of Band 1'',
and''RatioofBand4''-wereselectedandtestedinaddition
to the brightness values. The ''Ratio of Band L'' is the band L
mean value of an image object divided by the sum of all spectral
band mean values (Baatz
et al
., 2004). For impervious surfaces,
the ''Ratio of the Band 1'' (blue spectral band) is comparatively
high, and the ''Ratio of Band 4'' (near infrared spectral band)
is relatively low, which may help differentiate impervious areas
from the other land cover classes. To determine if the use of the
metrics would improve the classification, step-wise classifications
by adding these metrics were also performed.
To classify shadow intodifferent land cover types, class-related
features - relative to their adjacent forest or impervious - were
added in the next level process. For example, if more than 55% of
therelativeborderoftheshadowobjectwasimpervioussurfaces
or forest, then the shadow object was assigned a new value of
impervious or forest respectively. Otherwise, the shadow was
classified as non-forest.
17.4
Object-basedmodels
for estimating
high-resolution impervious
surface
Unlike the pixel-based impervious surface estimation models
detailed in Section 17.3, object-based models use both spectral
and spatial contextual information of the image data. A bottom
up region-merging approach is usually used to segment the
image into objects consisting of multiple pixels. This case study
demonstrates an example of impervious surface mapping using
an object-oriented approach.
17.4.1
Study area and data
preparation
The study site is a 2
.
1
×
1
.
7km
2
area surrounding the Minnesota
State University, Mankato campus, with the central coordinates
of 44
.
15
◦
Nand93
.
99
◦
W. Given the varied land covers it
includes, it provides a good test site for the purposes of the study.
A Quickbird image acquired on 6 October 2003 was obtained
from the Digital Globe archive collection. The data were recorded
in 11 bits and were radiometrically and geometrically corrected.
The image includes four multispectral bands at 2.4-m resolution
and one panchromatic band at 0.6-mresolution. The sun azimuth
and elevation angle were 158.1 and 38.5 degrees and the satellite
azimuth and elevation viewangles were 27.7 and 76.3 degrees. The
primary ancillary data usedwas the National Agriculture Imagery
Program (NAIP) color digital ortho-rectified aerial photography,
acquired on 26 June 2003 by the US Department of Agriculture.
This color aerial imagery was utilized as visual reference for
selecting training and testing samples.
All data were projected to UTM (spheroid NAD83, Zone 15)
and Datum GRS 1980 system. The training and test samples
for the object-oriented classification were manually delineated in
eCognition
. The locations of samples were selected randomly
and distributed evenly across the study area. All samples were
checked against the 1-mNAIP color aerial imagery. Compared to
™
Search WWH ::
Custom Search