Geoscience Reference
In-Depth Information
11.4.1
Data Preparation
Preprocessing of the CASI data included the conversion of at-sensor spectral
radiance to surface reflectance values using ATCOR-4 (Richter and Schläpfer
2013
), spectral smoothing using a Savitzky-Golay filter (Savitzky and Golay
1964
),
and image clipping to exclude areas affected by clouds and cloud shadows from
further analysis. Spectral smoothing is a commonly applied pre-classification step
to reduce the noise contained in spectral signatures collected by hyperspectral
sensors. It enables an improved identification of spectral key characteristics for
urban surface material mapping. In principle, Savitzky-Golay filters approximate the
“true” signature of noise-contaminated image spectra by a higher-order polynomial
operating within a predefined moving window. As opposed to conventional mean
filters, they mostly preserve the position and width of absorption features as well
as their absolute minima and maxima. Ideally, the filtering result should reflect a
compromise between spectral preservation and smoothing. Using a symmetrical
kernel size and a third order polynomial, the selection of filter parameters applied in
this study is based on Vaiphasa (
2006
).
Preprocessing of the LiDAR data involved two further steps. First, a normalized
digital surface model (nDSM) was calculated from the LiDAR data. Since Houston
is located at the seaside on a flat terrain, a constant elevation value of 13 m was
subtracted from the DSM (The City of Houston
2013
). The resulting nDSM contains
the height of urban objects relative to the ground. Second, the nDSM was smoothed
to enable the creation of derivative LiDAR products (e.g., nDSM slope). Even
though the calculation of the nDSM is potentially too simplistic at times, one has to
consider that the LiDAR DSM was made available in GeoTIFF format only (Image
Analysis and Data Fusion Technical Committee
2013
). Without having the original
LiDAR point cloud, there is hardly any possibility to apply more sophisticated
processing techniques to the input data. This is particularly true for existing nDSM
generation approaches from the literature since these usually rely on LiDAR raw
data. Moreover, the elevation threshold used to generate the nDSM was not only
taken from official numbers (The City of Houston
2013
) but was also compared
against other values and was found to be most suitable to make a discrimination
between urban objects and the ground.
With respect to the subsequent surface material classification, additional features
were derived from the input data. Among those features are the average reflectance
of all CASI bands (i.e., image brightness), the normalized difference vegetation
index (NDVI) (Tucker
1979
), and the slope of the nDSM (in percent) (Zevenbergen
and Thorne
1987
). The latter is useful for identifying transitions between flat areas
and elevated objects (e.g., trees and buildings Priestnall et al.
2000
).
11.4.2
Material Mapping
An overall number of 11 surface material classes were extracted from the data basis.
The surface materials were primarily chosen to meet the requirements for urban
Search WWH ::
Custom Search