The spatial resolution for UAVSAR and TerraSAR
imagery are 5.5m and 1m respectively. This research focuses
on analyzing the ability of each polarization channel in
identifying the landslides with different frequency bands of
synthetic aperture radar imagery. The UAVSAR multi
polarized, multi-look radar image acquired on 25 January,
2010 and the TerraSAR-X dual polarized high resolution
spotlight imagery acquired on 15 September, 2010 were used
in the analysis. In this study, the HH and VV backscattering
behavior of X-band and L-band radar backscatter from the
landslide has been investigated.
A. Study Area
The study area is a stretch of 230 km of levees along the
lower Mississippi River along the western boundary of the
state of Mississippi. At the time of image acquisition there was
one active landslide (latitude 32.5685, longitude -91.0393)
north of Vicksburg, Mississippi. A subset of 0.7 km length
containing this active landslide was chosen as the area of
analysis. The geo-referenced layers used in the analyses have
been masked by a 40 meter buffer from the crown of the levee
on the river side (where the landslides primarily occur). The
ground truth data was collected by the US Army Corp. of
Engineers (USACE) and the landslide polygons were drawn
using a GPS instrument during field data collection trips. This
data contains the location and timing of landslide appearance,
dimensions of the landslides, and their repair status.
III. DISCRETE WAVELET TRANSFORM (DWT) FEATURES
In SAR images, texture and intensity are the two important
parameters for the classification tasks. The roughness and
related textural characteristics of the soil affect the amount and
pattern of radar backscatter. Statistical texture analysis is very
important in this study since it allows better representation and
segmentation of various objects on the levee.
In this study, textural features derived from the SAR
imagery using discrete the wavelet transform (DWT) have
been used in the classification tasks. In DWT, the procedure
starts with passing the original SAR backscatter coefficients
through a set of high-pass and low-pass filters in a filter bank
followed by down sampling by a factor of two as shown in Fig.
1. The outputs from the low-pass branch are called wavelet
approximation coefficients and the outputs from the high-pass
branch are called wavelet detail coefficients. The
approximation features provide coarse textural information
from the image whereas the detail coefficients provide the
detail information.
IV. ANOMALY DETECTION
Detecting anomalies in the radar imagery involves
locating pixels with spectral signatures that are significantly
different from the background. RX Anomaly detector, a
training–free unsupervised classification scheme, detects
signatures that are distinct from the surroundings with no prior
knowledge. These unsupervised techniques are very fast and
do not depend on ground truth information. The algorithm
uses the covariance matrix to calculate the Mahalanobis
distance between the test pixels and the mean of the
background pixels. Suppose D is the number of spectral bands
and r is a Dx1 column pixel vector of the image. Then the RX
detector (RXD) implements a filter specified by
where μ is the global sample mean of the image subset and
KDxD is the sample covariance matrix of the image.